diff --git a/.gitignore b/.gitignore index 9e3a0b499f..b92bb9cc12 100644 --- a/.gitignore +++ b/.gitignore @@ -5,6 +5,7 @@ python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/ python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/ python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/ *.DS_Store +*.vs build/ build_doc/ *.user @@ -15,6 +16,7 @@ build_doc/ .cproject .pydevproject .settings/ +CMakeSettings.json Makefile .test_env/ third_party/ diff --git a/CMakeLists.txt b/CMakeLists.txt index 920c20d6f8..48e52961a9 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -204,22 +204,24 @@ include(external/snappy) # download snappy include(external/snappystream) include(external/threadpool) -set(WITH_ANAKIN OFF CACHE STRING "Disable Anakin first, will add it later." FORCE) +include(flags) # set paddle compile flags +include(cudnn) # set cudnn libraries, must before configure +include(cupti) +include(configure) # add paddle env configuration + if(WITH_GPU) include(cuda) include(tensorrt) include(external/anakin) +elseif() + set(WITH_ANAKIN OFF CACHE STRING "Anakin is used in GPU only now." FORCE) endif() -include(cudnn) # set cudnn libraries, must before configure -include(cupti) -include(configure) # add paddle env configuration include(generic) # simplify cmake module include(package) # set paddle packages include(ccache) # set ccache for compilation include(util) # set unittest and link libs include(rdma) # set rdma libraries -include(flags) # set paddle compile flags include(version) # set PADDLE_VERSION include(coveralls) # set code coverage include(inference_lib) # add paddle fluid inference libraries diff --git a/cmake/configure.cmake b/cmake/configure.cmake index c35096e09b..e03e15bfc0 100644 --- a/cmake/configure.cmake +++ b/cmake/configure.cmake @@ -50,7 +50,11 @@ if(NOT WITH_PROFILER) endif(NOT WITH_PROFILER) if(NOT CMAKE_CROSSCOMPILING) - if(WITH_AVX AND AVX_FOUND) + if(WITH_AVX AND AVX512F_FOUND) + set(SIMD_FLAG ${AVX512F_FLAG}) + elseif(WITH_AVX AND AVX2_FOUND) + set(SIMD_FLAG ${AVX2_FLAG}) + elseif(WITH_AVX AND AVX_FOUND) set(SIMD_FLAG ${AVX_FLAG}) elseif(SSE3_FOUND) set(SIMD_FLAG ${SSE3_FLAG}) @@ -99,12 +103,21 @@ if(WITH_GPU) endif() if(WITH_ANAKIN) if(${CUDA_VERSION_MAJOR} VERSION_LESS 8) - message(FATAL_ERROR "Anakin needs CUDA >= 8.0 to compile") + message(WARNING "Anakin needs CUDA >= 8.0 to compile. Force WITH_ANAKIN=OFF") + set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when CUDA >= 8.0." FORCE) endif() if(${CUDNN_MAJOR_VERSION} VERSION_LESS 7) - message(FATAL_ERROR "Anakin needs CUDNN >= 7.0 to compile") + message(WARNING "Anakin needs CUDNN >= 7.0 to compile. Force WITH_ANAKIN=OFF") + set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when CUDNN >= 7.0." FORCE) endif() endif() + if(WITH_ANAKIN) + # NOTICE(minqiyang): the end slash is important because $CUDNN_INCLUDE_DIR + # is a softlink to real cudnn.h directory + set(ENV{CUDNN_INCLUDE_DIR} "${CUDNN_INCLUDE_DIR}/") + get_filename_component(CUDNN_LIBRARY_DIR ${CUDNN_LIBRARY} DIRECTORY) + set(ENV{CUDNN_LIBRARY} ${CUDNN_LIBRARY_DIR}) + endif() elseif(WITH_AMD_GPU) add_definitions(-DPADDLE_WITH_HIP) set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -D__HIP_PLATFORM_HCC__") diff --git a/cmake/cudnn.cmake b/cmake/cudnn.cmake index 9eebea816c..cd51533926 100644 --- a/cmake/cudnn.cmake +++ b/cmake/cudnn.cmake @@ -25,8 +25,25 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS $ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/lib64 $ENV{CUDNN_ROOT}/lib - /usr/lib) -find_library(CUDNN_LIBRARY NAMES libcudnn.so libcudnn.dylib # libcudnn_static.a + /usr/lib + ${CUDA_TOOLKIT_ROOT_DIR} + ${CUDA_TOOLKIT_ROOT_DIR}/lib/x64 + ) +set(CUDNN_LIB_NAME "") +if (LINUX) +set(CUDNN_LIB_NAME "libcudnn.so") +endif(LINUX) + +if(WIN32) +# only support cudnn7 +set(CUDNN_LIB_NAME "cudnn.lib" "cudnn64_7.dll") +endif(WIN32) + +if(Apple) +set(CUDNN_LIB_NAME "libcudnn.dylib" "libcudnn.so") +endif(Apple) + +find_library(CUDNN_LIBRARY NAMES ${CUDNN_LIB_NAME} # libcudnn_static.a PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist} NO_DEFAULT_PATH DOC "Path to cuDNN library.") diff --git a/cmake/external/anakin.cmake b/cmake/external/anakin.cmake index 403873a510..78be074909 100644 --- a/cmake/external/anakin.cmake +++ b/cmake/external/anakin.cmake @@ -2,6 +2,11 @@ if (NOT WITH_ANAKIN) return() endif() +option(ANAKIN_ENABLE_OP_TIMER "Get more detailed information with Anakin op time" OFF) +if(ANAKIN_ENABLE_OP_TIMER) + add_definitions(-DPADDLE_ANAKIN_ENABLE_OP_TIMER) +endif() + INCLUDE(ExternalProject) set(ANAKIN_SOURCE_DIR ${THIRD_PARTY_PATH}/anakin) # the anakin install dir is only default one now @@ -11,33 +16,45 @@ set(ANAKIN_LIBRARY ${ANAKIN_INSTALL_DIR}) set(ANAKIN_SHARED_LIB ${ANAKIN_LIBRARY}/libanakin.so) set(ANAKIN_SABER_LIB ${ANAKIN_LIBRARY}/libanakin_saber_common.so) -# TODO(luotao): ANAKIN_MODLE_URL will move to demo ci later. -set(ANAKIN_MODLE_URL "http://paddle-inference-dist.bj.bcebos.com/mobilenet_v2.anakin.bin") +# TODO(luotao): ANAKIN_MODLE_URL etc will move to demo ci later. +set(INFERENCE_URL "http://paddle-inference-dist.bj.bcebos.com") +set(ANAKIN_MODLE_URL "${INFERENCE_URL}/mobilenet_v2.anakin.bin") +set(ANAKIN_RNN_MODLE_URL "${INFERENCE_URL}/anakin_test%2Fditu_rnn.anakin2.model.bin") +set(ANAKIN_RNN_DATA_URL "${INFERENCE_URL}/anakin_test%2Fditu_rnn_data.txt") execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_SOURCE_DIR}") -execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL}") +execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL} -N") +execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_RNN_MODLE_URL} -N") +execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_RNN_DATA_URL} -N") include_directories(${ANAKIN_INCLUDE}) include_directories(${ANAKIN_INCLUDE}/saber/) +include_directories(${ANAKIN_INCLUDE}/saber/core/) +include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/x86/) +include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/cuda/base/cuda_c/) -set(ANAKIN_COMPILE_EXTRA_FLAGS +set(ANAKIN_COMPILE_EXTRA_FLAGS -Wno-error=unused-but-set-variable -Wno-unused-but-set-variable - -Wno-error=unused-variable -Wno-unused-variable + -Wno-error=unused-variable -Wno-unused-variable -Wno-error=format-extra-args -Wno-format-extra-args -Wno-error=comment -Wno-comment -Wno-error=format -Wno-format + -Wno-error=maybe-uninitialized -Wno-maybe-uninitialized -Wno-error=switch -Wno-switch - -Wno-error=return-type -Wno-return-type + -Wno-error=return-type -Wno-return-type -Wno-error=non-virtual-dtor -Wno-non-virtual-dtor + -Wno-error=ignored-qualifiers + -Wno-ignored-qualifiers -Wno-sign-compare - -Wno-reorder + -Wno-reorder -Wno-error=cpp) ExternalProject_Add( extern_anakin ${EXTERNAL_PROJECT_LOG_ARGS} - # TODO(luotao): use PaddlePaddle/Anakin later + DEPENDS ${MKLML_PROJECT} + # Anakin codes error on Intel(R) Xeon(R) Gold 5117 CPU, temporary do not compile avx512 related code. GIT_REPOSITORY "https://github.com/luotao1/Anakin" - GIT_TAG "3957ae9263eaa0b1986758dac60a88852afb09be" + GIT_TAG "211d1fc5d813d70c0c14072f9083cf25f40940ea" PREFIX ${ANAKIN_SOURCE_DIR} UPDATE_COMMAND "" CMAKE_ARGS -DUSE_GPU_PLACE=YES @@ -46,6 +63,8 @@ ExternalProject_Add( -DPROTOBUF_ROOT=${THIRD_PARTY_PATH}/install/protobuf -DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml -DCUDNN_ROOT=${CUDNN_ROOT} + -DCUDNN_INCLUDE_DIR=${CUDNN_INCLUDE_DIR} + -DENABLE_OP_TIMER=${ANAKIN_ENABLE_OP_TIMER} ${EXTERNAL_OPTIONAL_ARGS} CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR} ) diff --git a/cmake/flags.cmake b/cmake/flags.cmake index 1120677a37..e0556a0bab 100644 --- a/cmake/flags.cmake +++ b/cmake/flags.cmake @@ -102,7 +102,6 @@ set(COMMON_FLAGS -fno-omit-frame-pointer -Wall -Wextra - -Werror -Wnon-virtual-dtor -Wdelete-non-virtual-dtor -Wno-unused-parameter @@ -115,6 +114,11 @@ set(COMMON_FLAGS -Wno-error=terminate # Warning in PADDLE_ENFORCE ) +# https://github.com/PaddlePaddle/Paddle/issues/12773 +if (NOT WIN32) +list(APPEND COMMON_FLAGS -Werror) +endif() + set(GPU_COMMON_FLAGS -fPIC -fno-omit-frame-pointer @@ -142,6 +146,11 @@ else() ${GPU_COMMON_FLAGS}) endif() +if(UNIX AND NOT APPLE) + # except apple from nix*Os family + set(LINUX TRUE) +endif(UNIX AND NOT APPLE) + foreach(flag ${COMMON_FLAGS}) safe_set_cflag(CMAKE_C_FLAGS ${flag}) diff --git a/cmake/simd.cmake b/cmake/simd.cmake index 53c2de332e..3eacf4d86a 100644 --- a/cmake/simd.cmake +++ b/cmake/simd.cmake @@ -10,6 +10,7 @@ if(CMAKE_COMPILER_IS_GNUCC OR CMAKE_COMPILER_IS_GNUCXX OR CMAKE_CXX_COMPILER_ID set(SSE3_FLAG "-msse3") set(AVX_FLAG "-mavx") set(AVX2_FLAG "-mavx2") + set(AVX512F_FLAG "-mavx512f") elseif(MSVC) set(MMX_FLAG "/arch:MMX") set(SSE2_FLAG "/arch:SSE2") @@ -81,5 +82,16 @@ int main() return 0; }" AVX2_FOUND) +# Check AVX512F +set(CMAKE_REQUIRED_FLAGS ${AVX512F_FLAG}) +set(AVX512F_FOUND_EXITCODE 1 CACHE STRING "Result from TRY_RUN" FORCE) +CHECK_CXX_SOURCE_RUNS(" +#include +int main() +{ + __m512i a = _mm512_undefined_epi32(); + return 0; +}" AVX512F_FOUND) + set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_RETAINED}) -mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND) +mark_as_advanced(MMX_FOUND SSE2_FOUND SSE3_FOUND AVX_FOUND AVX2_FOUND AVX512F_FOUND) diff --git a/doc/fluid/design/others/graph_survey.md b/doc/fluid/design/others/graph_survey.md index 6c6db08f46..97f395133b 100644 --- a/doc/fluid/design/others/graph_survey.md +++ b/doc/fluid/design/others/graph_survey.md @@ -28,7 +28,7 @@ def get_symbol(num_classes=10, **kwargs): -Varible here is actually a Symbol. Every basic Symbol will correspond to one Node, and every Node has its own NodeAttr. There is a op field in NodeAttr class, when a Symbol represents Variable(often input data), the op field is null. +Varible here is actually a Symbol. Every basic Symbol will correspond to one Node, and every Node has its own AnyAttr. There is a op field in AnyAttr class, when a Symbol represents Variable(often input data), the op field is null. Symbol contains a data member, std::vector outputs, and NodeEntry cantains a poniter to Node. We can follow the Node pointer to get all the Graph. diff --git a/doc/fluid/dev/new_op_cn.md b/doc/fluid/dev/new_op_cn.md index 587d819f79..c00f73be95 100644 --- a/doc/fluid/dev/new_op_cn.md +++ b/doc/fluid/dev/new_op_cn.md @@ -119,10 +119,29 @@ $$Out = scale*X$$ 这个例子有`AddAttr("scale", "...").SetDefault(1.0);` : 增加`scale`系数，作为参数属性，并且设置默认值为1.0。 +### 定义GradProtoMaker类 +每个Op的必须有一个对应的GraProtoMaker，若未定制对应前向Op的GradProtoMaker，fluid提供了DefaultGradProtoMaker，默认注册会使用全部输入输出，包括Input, Output, Output@Grad等，使用不需要的变量的会造成显存浪费。 +下面示例定义了ScaleOp的GradProtoMaker。 + +```cpp +class ScaleGradMaker : public framework::SingleGradOpDescMaker { + public: + using framework::SingleGradOpDescMaker::SingleGradOpDescMaker; + + std::unique_ptr Apply() const override { + auto *grad_op = new framework::OpDesc(); + grad_op->SetType("scale"); + grad_op->SetInput("X", OutputGrad("Out")); + grad_op->SetOutput("Out", InputGrad("X")); + grad_op->SetAttr("scale", GetAttr("scale")); + return std::unique_ptr(grad_op); + } +}; +``` ### 定义Operator类 -下面的点实现了MulOp的定义： +下面实现了MulOp的定义： ```cpp class MulOp : public framework::OperatorWithKernel { @@ -334,3 +353,83 @@ ctest -R test_mul_op - 注册Op时的类型名，需要和该Op的名字一样。即不允许在`A_op.cc`里面，注册`REGISTER_OPERATOR(B, ...)`等，这将会导致单元测试出错。 - 如果Op没有实现CUDA Kernel，请不要创建空的`*_op.cu`，这将会导致单元测试出错。 - 如果多个Op依赖一些共用的函数，可以创建非`*_op.*`格式的文件来存放，如`gather.h`文件。 + +### PADDLE_ENFORCE使用注意 + +实现Op时检查数据的合法性需要使用PADDLE_ENFORCE以及PADDLE_ENFORCE_EQ等宏定义，基本格式如下： + +``` +PADDLE_ENFORCE(表达式, 错误提示信息) +PADDLE_ENFORCE_EQ(比较对象A, 比较对象B, 错误提示信息) +``` + +如果表达式为真，或者比较对象A=B，则检查通过，否则会终止程序运行，向用户反馈相应的错误提示信息。 +为了确保提示友好易懂，开发者需要注意其使用方法。 + +#### 总体原则 + +任何使用了PADDLE_ENFORCE与PADDLE_ENFORCE_**检查的地方，必须有详略得当的备注解释！**错误提示信息**不能为空！ + +#### 提示信息书写标准 + +1. [required] 哪里错了？为什么错了？ + - 例如：`ValueError: Mismatched label shape` +2. [optional] 期望的输入是什么样的？实际的输入是怎样的？ + - 例如：`Expected labels dimension=1. Received 4.` +3. [optional] 能否给出修改意见？ + - 例如：`Suggested Fix:If your classifier expects one-hot encoding label,check your n_classes argument to the estimatorand/or the shape of your label.Otherwise, check the shape of your label.` + +如果并非必要或者简洁的描述即可表达清楚以上要点，根据情况书写亦可。 + +##### FAQ 典型问题 + +1. 无报错信息或报错信息过于简单，不能给用户提供有效的提示！ + +问题示例1 ：未写提示信息 +``` +PADDLE_ENFORCE(ctx->HasInput("X"), ""); +``` +问题示例2 ：提示信息过于简单 +``` +PADDLE_ENFORCE(i != nullptr, "I must be set"); // I是什么？ +``` + +2. 在报错信息中使用开发人员定义的变量缩写，不易理解！ + +问题示例： +``` +PADDLE_ENFORCE(forward_pd != nullptr, + "Fail to find eltwise_fwd_pd in device context"); //eltwise_fwd_pd用户可能看不懂 +``` + +3. OP内部调用非法接口：Op内部如果出现Output = ShareDataWith(Input) +问题示例： +```cpp +auto *out = ctx.Output("Out"); +auto *in = ctx.Input("X"); +out->ShareDataWith(*in); +``` +Op内部如果出现Output = ShareDataWith(Input)，相当于operator图的中有一条隐藏边，连接了Input和Output，这条边无法在图分析中表达，引发基于图优化的错误。 + +4. OP实现的性能实践 +调用了eigen的broadcast, chop等操作，性能会比手写cuda kernel差几倍以上。此时cpu的实现可以复用eigen，gpu实现可以实现cuda kernel. + + +#### OP InferShape检查提示信息特别说明 + +- 检查输入输出变量，请统一遵循以下格式 +`Input(变量名) of OP名 operator should not be null.` + +正确示例： +``` +PADDLE_ENFORCE(ctx->HasInput("Input"), + "Input(Input) of LSTMP operator should not be null."); +``` + +- 反向Op的输入输出检查，要写明反向Op的名字 + +正确示例： +``` +PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of LoDResetGrad opreator should not be null."); +``` diff --git a/doc/fluid/dev/use_eigen_cn.md b/doc/fluid/dev/use_eigen_cn.md index 75922e7d85..56203d6fad 100644 --- a/doc/fluid/dev/use_eigen_cn.md +++ b/doc/fluid/dev/use_eigen_cn.md @@ -7,7 +7,7 @@ Eigen Tensor模块对element-wise计算提供了强大的支持，并且书写一份代码，可以同时在CPU、GPU执行。但Eigen Tensor是一个正在开发中的模块，因此可能测试不够完备，文档较少。 -关于Eigen Tensor模块的详细介绍请参考[文档1](https://github.com/RLovelett/eigen/blob/master/unsupported/Eigen/CXX11/src/Tensor/README.md) 和[文档2](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md) +关于Eigen Tensor模块的详细介绍请参考[Eigen文档](https://bitbucket.org/eigen/eigen/src/default/unsupported/Eigen/CXX11/src/Tensor/README.md) ## paddle::framework::Tensor diff --git a/paddle/fluid/API.spec b/paddle/fluid/API.spec index e963902a50..9250cde1b2 100644 --- a/paddle/fluid/API.spec +++ b/paddle/fluid/API.spec @@ -78,7 +78,7 @@ paddle.fluid.io.load_vars ArgSpec(args=['executor', 'dirname', 'main_program', ' paddle.fluid.io.load_params ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.io.load_persistables ArgSpec(args=['executor', 'dirname', 'main_program', 'filename'], varargs=None, keywords=None, defaults=(None, None)) paddle.fluid.io.save_inference_model ArgSpec(args=['dirname', 'feeded_var_names', 'target_vars', 'executor', 'main_program', 'model_filename', 'params_filename', 'export_for_deployment'], varargs=None, keywords=None, defaults=(None, None, None, True)) -paddle.fluid.io.load_inference_model ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename'], varargs=None, keywords=None, defaults=(None, None)) +paddle.fluid.io.load_inference_model ArgSpec(args=['dirname', 'executor', 'model_filename', 'params_filename', 'pserver_endpoints'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.io.get_inference_program ArgSpec(args=['target_vars', 'main_program'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.initializer.ConstantInitializer.__init__ ArgSpec(args=['self', 'value', 'force_cpu'], varargs=None, keywords=None, defaults=(0.0, False)) paddle.fluid.initializer.UniformInitializer.__init__ ArgSpec(args=['self', 'low', 'high', 'seed'], varargs=None, keywords=None, defaults=(-1.0, 1.0, 0)) @@ -153,6 +153,7 @@ paddle.fluid.layers.image_resize ArgSpec(args=['input', 'out_shape', 'scale', 'n paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', 'resample'], varargs=None, keywords=None, defaults=('BILINEAR',)) paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name'], varargs=None, keywords=None, defaults=(None, None, None)) paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None) +paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.random_crop ArgSpec(args=['x', 'shape', 'seed'], varargs=None, keywords=None, defaults=(None,)) paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], varargs=None, keywords=None, defaults=None) paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)) @@ -250,7 +251,6 @@ paddle.fluid.layers.logical_not ArgSpec(args=[], varargs='args', keywords='kwarg paddle.fluid.layers.uniform_random_batch_size_like ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) paddle.fluid.layers.gaussian_random ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) paddle.fluid.layers.gaussian_random_batch_size_like ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) -paddle.fluid.layers.scatter ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) paddle.fluid.layers.sum ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) paddle.fluid.layers.slice ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) paddle.fluid.layers.shape ArgSpec(args=[], varargs='args', keywords='kwargs', defaults=None) diff --git a/paddle/fluid/framework/CMakeLists.txt b/paddle/fluid/framework/CMakeLists.txt index 1d62792b80..2ec422cc17 100644 --- a/paddle/fluid/framework/CMakeLists.txt +++ b/paddle/fluid/framework/CMakeLists.txt @@ -99,8 +99,13 @@ else() cc_library(executor SRCS executor.cc DEPS op_registry device_context scope framework_proto glog lod_rank_table feed_fetch_method) endif() - -cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph graph_viz_pass multi_devices_graph_pass multi_devices_graph_print_pass multi_devices_graph_check_pass) +if (NOT WIN32) +cc_library(parallel_executor SRCS parallel_executor.cc DEPS + threaded_ssa_graph_executor scope_buffered_ssa_graph_executor + graph graph_viz_pass multi_devices_graph_pass + multi_devices_graph_print_pass multi_devices_graph_check_pass + fast_threaded_ssa_graph_executor) +endif() # NOT WIN32 cc_library(prune SRCS prune.cc DEPS framework_proto) cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context) @@ -115,6 +120,8 @@ cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc) # cc_test(channel_test SRCS channel_test.cc) cc_test(tuple_test SRCS tuple_test.cc ) +cc_test(rw_lock_test SRCS rw_lock_test.cc) + # disable test temporarily. # TODO https://github.com/PaddlePaddle/Paddle/issues/11971 # cc_test(concurrency_test SRCS concurrency_test.cc DEPS go_op channel_close_op channel_create_op diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h index 8428bf8e33..14ca3e9620 100644 --- a/paddle/fluid/framework/attribute.h +++ b/paddle/fluid/framework/attribute.h @@ -128,7 +128,8 @@ struct ExtractAttribute { attr_value = &boost::get(attr); } catch (boost::bad_get& bad_get) { PADDLE_THROW("Cannot get attribute %s by type %s, its type is %s", - attr_name_, typeid(T).name(), attr.type().name()); + attr_name_, paddle::platform::demangle(typeid(T).name()), + paddle::platform::demangle(attr.type().name())); } return attr_value; } @@ -160,7 +161,7 @@ struct ExtractAttribute { attr_value = &boost::get(attr); } catch (boost::bad_get& bad_get) { PADDLE_THROW("Cannot get attribute %s by type bool, its type is %s", - attr_name_, attr.type().name()); + attr_name_, paddle::platform::demangle(attr.type().name())); } return attr_value; } @@ -186,7 +187,7 @@ struct ExtractAttribute { attr_value = &boost::get(attr); } catch (boost::bad_get& bad_get) { PADDLE_THROW("Cannot get attribute %s by type int64_t, its type is %s", - attr_name_, attr.type().name()); + attr_name_, paddle::platform::demangle(attr.type().name())); } return attr_value; } diff --git a/paddle/fluid/framework/details/CMakeLists.txt b/paddle/fluid/framework/details/CMakeLists.txt index 8f6c4163d6..abd5459f6d 100644 --- a/paddle/fluid/framework/details/CMakeLists.txt +++ b/paddle/fluid/framework/details/CMakeLists.txt @@ -42,3 +42,5 @@ cc_test(gather_op_test SRCS gather_op_handle_test.cc DEPS var_handle op_handle_b cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_executor.cc DEPS ssa_graph_executor) #cc_test(reduce_op_handle_test SRCS reduce_op_handle_test.cc DEPS var_handle op_handle_base scope ddim memory # device_context reduce_op_handle ) +cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc + DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context) diff --git a/paddle/fluid/framework/details/execution_strategy.h b/paddle/fluid/framework/details/execution_strategy.h index 716d674fa2..5183be878e 100644 --- a/paddle/fluid/framework/details/execution_strategy.h +++ b/paddle/fluid/framework/details/execution_strategy.h @@ -19,10 +19,13 @@ namespace framework { namespace details { struct ExecutionStrategy { + enum ExecutorType { kDefault = 0, kExperimental = 1 }; + size_t num_threads_{0}; bool use_cuda_{true}; bool allow_op_delay_{false}; size_t num_iteration_per_drop_scope_{100}; + ExecutorType type_{kDefault}; }; } // namespace details diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc new file mode 100644 index 0000000000..7606f2bc06 --- /dev/null +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc @@ -0,0 +1,175 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. +#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h" +#include +#include +#include "paddle/fluid/framework/details/fetch_op_handle.h" +#include "paddle/fluid/framework/details/multi_devices_helper.h" + +namespace paddle { +namespace framework { +namespace details { + +FastThreadedSSAGraphExecutor::FastThreadedSSAGraphExecutor( + const ExecutionStrategy &strategy, const std::vector &local_scopes, + const std::vector &places, + std::unique_ptr &&graph) + : strategy_(strategy), + local_scopes_(local_scopes), + places_(places), + graph_(std::move(graph)), + pool_(strategy.num_threads_ + + 1), // add one more thread for generate op_deps + fetch_ctxs_(places) { + auto &ops = graph_->Get("ops"); + + for (auto &op : ops) { + int dep = static_cast(op->NotReadyInputSize()); + op_deps_.emplace(op.get(), dep); + if (dep == 0) { + bootstrap_ops_.emplace_back(op.get()); + } + } + + PrepareAtomicOpDeps(); +} + +FeedFetchList FastThreadedSSAGraphExecutor::Run( + const std::vector &fetch_tensors) { + std::unique_ptr>> + op_deps = atomic_op_deps_.get(); + PrepareAtomicOpDeps(); + + paddle::framework::FeedFetchList fetches; + fetches.resize(fetch_tensors.size()); + std::unordered_map> fetched_vars; + std::vector> fetch_nodes; + std::vector> fetch_ops; + + for (auto &fetch_var_name : fetch_tensors) { + for (auto &var_map : graph_->Get("vars")) { + auto it = var_map.find(fetch_var_name); + if (it != var_map.end()) { + fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get()); + } + } + } + + for (size_t i = 0; i < fetch_tensors.size(); ++i) { + auto &var_name = fetch_tensors[i]; + auto fetched_var_it = fetched_vars.find(var_name); + PADDLE_ENFORCE(fetched_var_it != fetched_vars.end(), + "Cannot find fetched variable.(Perhaps the main_program " + "is not set to ParallelExecutor)"); + + auto &vars = fetched_var_it->second; + + fetch_nodes.emplace_back(new ir::Node("fetch", ir::Node::Type::kOperation)); + auto *op = new FetchOpHandle(fetch_nodes.back().get(), &fetches, i, + &local_scopes_); + fetch_ops.emplace_back(op); + + for (auto &p : places_) { + op->SetDeviceContext(p, fetch_ctxs_.Get(p)); + } + + for (auto *var : vars) { + op->AddInput(var); + } + + (*op_deps)[op] = static_cast(op->NotReadyInputSize()); + } + + size_t num_complete = 0; + remaining_ = 0; + BlockingQueue complete_q; + for (auto op : bootstrap_ops_) { + RunOpAsync(op_deps.get(), op, &complete_q); + } + + while (num_complete != op_deps->size()) { + size_t num_comp = complete_q.Pop(); + if (num_comp == -1UL) { + int remaining = 0; + while (true) { + remaining = remaining_; + if (remaining == 0) { + break; + } + for (int i = 0; i < remaining; ++i) { + complete_q.Pop(); + } + } + exception_.ReThrow(); + } + num_complete += num_comp; + } + // Wait FetchOps. + if (!fetch_ops.empty()) { + fetch_ops.clear(); + } + return fetches; +} +void FastThreadedSSAGraphExecutor::RunOpAsync( + std::unordered_map> *op_deps, + OpHandleBase *op, BlockingQueue *complete_q) { + ++remaining_; + this->pool_.enqueue([=] { + OpHandleBase *op_to_run = op; + size_t complete = 0; + while (op_to_run != nullptr) { + try { + op_to_run->Run(strategy_.use_cuda_); + ++complete; + } catch (...) { + exception_.Catch(std::current_exception()); + --remaining_; + complete_q->Push(-1UL); + return; + } + auto &outputs = op_to_run->Outputs(); + op_to_run = nullptr; + for (auto &output : outputs) { + for (auto &pending_op : output->PendingOps()) { + std::atomic &deps = op_deps->at(pending_op); + if (deps.fetch_sub(1) == 1) { // pending_op ready + if (op_to_run == nullptr) { + op_to_run = pending_op; + } else { + this->RunOpAsync(op_deps, pending_op, complete_q); + } + } + } + } + } + --remaining_; + complete_q->Push(complete); + }); +} +void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() { + atomic_op_deps_ = pool_.enqueue([&] { + std::unordered_map> *op_deps = + new std::unordered_map>; + for (auto &pair : op_deps_) { + (*op_deps)[pair.first] = pair.second; + } + return std::unique_ptr< + std::unordered_map>>(op_deps); + }); +} + +const ir::Graph &FastThreadedSSAGraphExecutor::Graph() const { return *graph_; } +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h new file mode 100644 index 0000000000..dad3a231cb --- /dev/null +++ b/paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h @@ -0,0 +1,64 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once +#include +#include +#include "ThreadPool.h" +#include "paddle/fluid/framework/blocking_queue.h" +#include "paddle/fluid/framework/details/exception_holder.h" +#include "paddle/fluid/framework/details/execution_strategy.h" +#include "paddle/fluid/framework/details/ssa_graph_executor.h" + +namespace paddle { +namespace framework { +class Scope; +namespace details { + +class OpHandleBase; +class FastThreadedSSAGraphExecutor : public SSAGraphExecutor { + public: + FastThreadedSSAGraphExecutor(const ExecutionStrategy &strategy, + const std::vector &local_scopes, + const std::vector &places, + std::unique_ptr &&graph); + FeedFetchList Run(const std::vector &fetch_tensors) override; + const ir::Graph &Graph() const override; + + private: + ExecutionStrategy strategy_; + std::vector local_scopes_; + std::vector places_; + std::unique_ptr graph_; + + std::unordered_map op_deps_; + std::vector bootstrap_ops_; + + ::ThreadPool pool_; + platform::DeviceContextPool fetch_ctxs_; + std::atomic remaining_; + + void RunOpAsync(std::unordered_map> *op_deps, + OpHandleBase *op, BlockingQueue *complete_q); + + void PrepareAtomicOpDeps(); + + std::future< + std::unique_ptr>>> + atomic_op_deps_; + ExceptionHolder exception_; +}; +} // namespace details +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc index ee9f9184da..3812f0abf1 100644 --- a/paddle/fluid/framework/details/op_handle_base.cc +++ b/paddle/fluid/framework/details/op_handle_base.cc @@ -158,6 +158,16 @@ void OpHandleBase::RunAndRecordEvent(platform::Place p, #endif } +size_t OpHandleBase::NotReadyInputSize() const { + std::unordered_set res; + for (auto *var : inputs_) { + if (var->GeneratedOp() != nullptr) { + res.emplace(var); + } + } + return res.size(); +} + } // namespace details } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/details/op_handle_base.h b/paddle/fluid/framework/details/op_handle_base.h index 2d7f189428..9fbefabc84 100644 --- a/paddle/fluid/framework/details/op_handle_base.h +++ b/paddle/fluid/framework/details/op_handle_base.h @@ -81,6 +81,8 @@ class OpHandleBase { return res.size(); } + size_t NotReadyInputSize() const; + const std::vector &Outputs() const { return outputs_; } size_t NoDummyInputSize() const; diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index 923a7083d4..da0955a9a0 100644 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -5,8 +5,12 @@ cc_library(pass SRCS pass.cc DEPS graph node graph_helper) cc_library(graph_viz_pass SRCS graph_viz_pass.cc DEPS graph pass graph_helper) cc_library(graph_traits SRCS graph_traits.cc DEPS graph) cc_library(graph_pattern_detecter SRCS graph_pattern_detecter.cc DEPS graph graph_helper graph_traits) +cc_library(fc_fuse_pass SRCS fc_fuse_pass.cc DEPS graph graph_pattern_detecter) +cc_library(infer_clean_graph_pass SRCS infer_clean_graph_pass.cc DEPS graph pass) + cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper) cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry) cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry) cc_test(test_graph_pattern_detecter SRCS graph_pattern_detecter_tester.cc DEPS graph_pattern_detecter) +cc_test(test_fc_fuse_pass SRCS fc_fuse_pass_tester.cc DEPS fc_fuse_pass graph_pattern_detecter graph pass graph_traits framework_proto) diff --git a/paddle/fluid/framework/ir/fc_fuse_pass.cc b/paddle/fluid/framework/ir/fc_fuse_pass.cc new file mode 100644 index 0000000000..f4327742ea --- /dev/null +++ b/paddle/fluid/framework/ir/fc_fuse_pass.cc @@ -0,0 +1,192 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/fc_fuse_pass.h" +#include +#include +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { +namespace ir { + +bool VarOutLinksToOp(Node* node, const std::string& op_type) { + for (auto* out : node->outputs) { + if (out->IsOp() && out->Op()->Type() == op_type) { + return true; + } + } + return false; +} + +void BuildFCPattern(PDPattern* pattern) { + // make sure the selected MUL op has one input argument is a parameter. + auto* mul_parameter_var = pattern->NewNode( + [](Node* node) { + return node->IsVar() && node->outputs.size() == 1UL && + node->outputs.front()->Op()->Type() == "mul" && node->Var() && + node->Var()->Persistable(); // check is a parameter + }, + "mul_weight" /*name*/); + + auto* mul_tmp_input_var = pattern->NewNode( + [](Node* node) { + bool result = + node->IsVar() && node->outputs.size() >= 1UL && node->Var() && + !node->Var()->Persistable(); // this input is not an parameter. + if (!result) return false; + // check whether one output is MUL op. + for (auto* op : node->outputs) { + if (op->IsOp() && op->Op()->Type() == "mul") return true; + } + return false; + }, + "mul_tmp_var" /*name*/); + + // select a MUL op + auto* mul_op = pattern->NewNode( + [](Node* node) { + return node->IsOp() && // start from an Op + node->Op()->Type() == "mul"; // type is mul + // the output should be consumed only by one element_add, that check + // leaves in a Var PDNode. + }, + "mul" /*name*/); + + // make sure the MUL op's output has only one consumer and links to an + // ELEMENTWISE_ADD op. + auto* mul_out_var = pattern->NewNode( + [](Node* node) { + return node->IsVar() && // starts from a Var + node->outputs.size() == 1UL && // only has one consumer + node->outputs.front()->IsOp() && // check basic logic + node->Var() && // not a ControlDepVar + node->outputs.front()->Op()->Type() == + "elementwise_add"; // a very strong validation + }, + "mul_out"); + // this check is not essential, just to make the corresponding variable Node + // retrival easier. + auto* elementwise_add_tmp_var = pattern->NewNode( + [](Node* node) { + return node->IsVar() && node->outputs.size() >= 1UL && node->Var() && + VarOutLinksToOp(node, "elementwise_add"); + }, + "elementwise_add_tmpvar"); + + // select an ELEMENTWISE_ADD op + auto* elementwise_add_op = pattern->NewNode( + [](Node* node) { + return node->IsOp() && node->Op()->Type() == "elementwise_add"; + }, + "elementwise_add" /*name*/); + + // get the ELEMENTWISE_ADD op's output + auto* elementwise_add_out_var = pattern->NewNode( + [](Node* node) { + return node->IsVar() && node->inputs.size() == 1UL && node->Var() && + node->inputs.front()->Op()->Type() == "elementwise_add"; + }, + "elementwise_add_out"); + + pattern->AddEdge(mul_parameter_var, mul_op); + pattern->AddEdge(mul_tmp_input_var, mul_op); + pattern->AddEdge(mul_op, mul_out_var); + pattern->AddEdge(mul_out_var, elementwise_add_op); + pattern->AddEdge(elementwise_add_tmp_var, elementwise_add_op); + pattern->AddEdge(elementwise_add_op, elementwise_add_out_var); +} + +// Replace the node `from` in the links to `to` +bool LinksReplace(std::vector* links, Node* from, Node* to) { + for (auto*& n : *links) { + if (n == from) { + n = to; + return true; + } + } + return false; +} + +std::unique_ptr FCFusePass::ApplyImpl( + std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + + std::unordered_set nodes2delete; + + GraphPatternDetecter gpd; + BuildFCPattern(gpd.mutable_pattern()); + +#define GET_NODE(id) \ + PADDLE_ENFORCE(subgraph.count(gpd.pattern().RetriveNode(#id)), \ + "pattern has no Node called %s", #id); \ + auto* id = subgraph.at(gpd.pattern().RetriveNode(#id)); \ + PADDLE_ENFORCE_NOT_NULL(id, "subgraph has no node %s", #id); + + auto handler = [&](const GraphPatternDetecter::subgraph_t& subgraph, + Graph* g) { + VLOG(4) << "handle FC fuse"; + // Currently, there is no FC op available, so I will just simulate the + // scenerio. + // FC's fusion is simple, just op fuse, no need to process the + // parameters. + GET_NODE(mul_tmp_var); // x + GET_NODE(mul_weight); // Y + GET_NODE(elementwise_add_tmpvar); // bias + GET_NODE(elementwise_add_out); // Out + GET_NODE(mul); // MUL op + GET_NODE(elementwise_add); // ELEMENT_ADD op + GET_NODE(mul_out); // tmp +#undef GET_NODE + + // Create an FC Node. + OpDesc desc; + std::string fc_x_in = mul_tmp_var->Name(); + std::string fc_Y_in = mul_weight->Name(); + std::string fc_bias_in = elementwise_add_tmpvar->Name(); + std::string fc_out = elementwise_add_out->Name(); + desc.SetInput("Input", std::vector({fc_x_in})); + desc.SetInput("W", std::vector({fc_Y_in})); + desc.SetInput("Bias", std::vector({fc_bias_in})); + desc.SetOutput("Out", std::vector({fc_out})); + desc.SetType("fc"); + auto fc_node = g->CreateOpNode(&desc); // OpDesc will be copied. + fc_node->inputs = + std::vector({mul_tmp_var, mul_weight, elementwise_add_tmpvar}); + fc_node->outputs.push_back(elementwise_add_out); + + // Update link relatons + PADDLE_ENFORCE(LinksReplace(&mul_tmp_var->outputs, mul, fc_node)); + PADDLE_ENFORCE(LinksReplace(&mul_weight->outputs, mul, fc_node)); + PADDLE_ENFORCE(LinksReplace(&elementwise_add_tmpvar->outputs, + elementwise_add, fc_node)); + PADDLE_ENFORCE( + LinksReplace(&elementwise_add_out->inputs, elementwise_add, fc_node)); + + // Drop old nodes + graph->RemoveNode(mul); + graph->RemoveNode(elementwise_add); + graph->RemoveNode(mul_out); // tmp variable + }; + + gpd(graph.get(), handler); + + return graph; +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(fc_fuse_pass, paddle::framework::ir::FCFusePass); diff --git a/paddle/fluid/operators/fill_constant_op.cu.cc b/paddle/fluid/framework/ir/fc_fuse_pass.h similarity index 50% rename from paddle/fluid/operators/fill_constant_op.cu.cc rename to paddle/fluid/framework/ir/fc_fuse_pass.h index 51ccaefa43..eb43dd4486 100644 --- a/paddle/fluid/operators/fill_constant_op.cu.cc +++ b/paddle/fluid/framework/ir/fc_fuse_pass.h @@ -12,15 +12,25 @@ // See the License for the specific language governing permissions and // limitations under the License. -#include "paddle/fluid/operators/fill_constant_op.h" -#include "paddle/fluid/platform/float16.h" +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/graph_pattern_detecter.h" +#include "paddle/fluid/framework/ir/pass.h" -namespace ops = paddle::operators; -REGISTER_OP_CUDA_KERNEL( - fill_constant, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel) +namespace paddle { +namespace framework { +namespace ir { + +/* + * Fuse the MUL and ELEMENTWISE_ADD to a FCOp. + */ +class FCFusePass : public Pass { + public: + virtual ~FCFusePass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const; +}; + +} // namespace ir +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc b/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc new file mode 100644 index 0000000000..87ba417b1a --- /dev/null +++ b/paddle/fluid/framework/ir/fc_fuse_pass_tester.cc @@ -0,0 +1,90 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/framework/ir/fc_fuse_pass.h" + +#include + +namespace paddle { +namespace framework { +namespace ir { + +void SetOp(ProgramDesc* prog, const std::string& type, + const std::vector& inputs, + const std::vector& outputs) { + auto* op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + op->SetInput("Xs", inputs); + op->SetOutput("Ys", outputs); +} + +// a->OP0->b +// a->OP1->c +// (b, c)->mul->d +// (d, e)->elementwise_add->f +ProgramDesc BuildProgramDesc() { + ProgramDesc prog; + for (auto& v : std::vector({"a", "b", "c", "d", "e", "f"})) { + auto* var = prog.MutableBlock(0)->Var(v); + var->SetType(proto::VarType::SELECTED_ROWS); + if (v == "c") { + var->SetPersistable(true); + } + } + + SetOp(&prog, "OP0", std::vector({"a"}), + std::vector({"b"})); + SetOp(&prog, "OP1", std::vector({"a"}), + std::vector({"c"})); + SetOp(&prog, "mul", std::vector({"b", "c"}), + std::vector({"d"})); + SetOp(&prog, "elementwise_add", std::vector({"d", "e"}), + std::vector({"f"})); + + return prog; +} + +TEST(FCFusePass, basic) { + auto prog = BuildProgramDesc(); + + std::unique_ptr graph(new ir::Graph(prog)); + + auto pass = PassRegistry::Instance().Get("fc_fuse_pass"); + + int pre_nodes = graph->Nodes().size(); + + graph = pass->Apply(std::move(graph)); + + int after_nodes = graph->Nodes().size(); + + // Remove 3 Nodes: MUL,ELEMENTWISE_ADD, mul_out + // Add 1 Node: FC + EXPECT_EQ(pre_nodes - 2, after_nodes); + + // Assert fc op in newly generated graph + int fc_count = 0; + + for (auto* node : graph->Nodes()) { + if (node->IsOp() && node->Op()->Type() == "fc") { + ++fc_count; + } + } + EXPECT_EQ(fc_count, 1); +} + +} // namespace ir +} // namespace framework +} // namespace paddle + +USE_PASS(fc_fuse_pass); diff --git a/paddle/fluid/framework/ir/graph.cc b/paddle/fluid/framework/ir/graph.cc index f87d5212c0..2a6bf4ac23 100644 --- a/paddle/fluid/framework/ir/graph.cc +++ b/paddle/fluid/framework/ir/graph.cc @@ -117,7 +117,15 @@ Graph::Graph(const ProgramDesc &program) : program_(program) { } // For output args, always create a new var. for (auto &each_var_name : op->OutputArgumentNames()) { - ir::Node *var = CreateVarNode(all_vars.at(each_var_name)); + ir::Node *var = nullptr; + if (all_vars.count(each_var_name) != 0) { + var = CreateVarNode(all_vars.at(each_var_name)); + } else { + // Operation output vars can be @EMPTY@. For example, while_grad + // can have multi @EMPTY@ outputs with no VarDesc. + // TODO(panyx0718): Add a test. + var = CreateEmptyNode(each_var_name, ir::Node::Type::kVariable); + } var_nodes[each_var_name].push_back(var); node->outputs.push_back(var); var->inputs.push_back(node); @@ -208,7 +216,8 @@ Graph::Graph(const ProgramDesc &program) : program_(program) { // Add write after write dependence ir::Node *upstream_op = (*it_old)->inputs.empty() ? nullptr : (*it_old)->inputs[0]; - if (upstream_op) { + // TODO(zcd): Add a test. + if (upstream_op && upstream_op != write_op) { ir::Node *dep_var = CreateControlDepVar(); write_op->inputs.push_back(dep_var); upstream_op->outputs.push_back(dep_var); diff --git a/paddle/fluid/framework/ir/graph.h b/paddle/fluid/framework/ir/graph.h index 5736a5c4e2..25e33861c0 100644 --- a/paddle/fluid/framework/ir/graph.h +++ b/paddle/fluid/framework/ir/graph.h @@ -98,11 +98,13 @@ class Graph { // Create a normal variable with non-null VarDesc. ir::Node *CreateVarNode(VarDesc *var_desc) { + PADDLE_ENFORCE(var_desc); return AddNode(new ir::Node(var_desc)); } // Create a normal runnable operator with OpDesc. ir::Node *CreateOpNode(OpDesc *op_desc) { + PADDLE_ENFORCE(op_desc); return AddNode(new ir::Node(op_desc)); } @@ -134,6 +136,14 @@ class Graph { return ret; } + void RemoveNode(ir::Node *node) { + PADDLE_ENFORCE(node_set_.find(node) != node_set_.end()); + node_set_.erase(node); + nodes_.erase(node); + } + + const ProgramDesc &program() const { return program_; } + private: // This method takes ownership of `node`. ir::Node *AddNode(ir::Node *node) { @@ -143,12 +153,6 @@ class Graph { return node; } - void RemoveNode(ir::Node *node) { - PADDLE_ENFORCE(node_set_.find(node) != node_set_.end()); - node_set_.erase(node); - nodes_.erase(node); - } - // NOTE: program_ shouldn't be exposed to user. const ProgramDesc &program_; std::map attrs_; diff --git a/paddle/fluid/framework/ir/graph_helper.cc b/paddle/fluid/framework/ir/graph_helper.cc index b1c19e6535..dc81a2cac5 100644 --- a/paddle/fluid/framework/ir/graph_helper.cc +++ b/paddle/fluid/framework/ir/graph_helper.cc @@ -104,7 +104,7 @@ std::map> BuildOperationAdjList( for (auto &adj_n : var->inputs) { PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation); adj_list[n].insert(adj_n); - VLOG(3) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) + VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast(adj_n) << " -> " << n->Name() << reinterpret_cast(n) << " via " << var->Name() << reinterpret_cast(var); } diff --git a/paddle/fluid/framework/ir/graph_pattern_detecter.cc b/paddle/fluid/framework/ir/graph_pattern_detecter.cc index f27d9b0509..e197861251 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detecter.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detecter.cc @@ -12,6 +12,7 @@ // See the License for the specific language governing permissions and // limitations under the License. +#include #include #include @@ -24,12 +25,30 @@ namespace paddle { namespace framework { namespace ir { +size_t PDPattern::id_ = 0UL; + PDNode* PDPattern::NewNode(PDNode::teller_t&& teller, const std::string& name) { + if (!name.empty()) { + PADDLE_ENFORCE_EQ(node_map_.count(name), 0, + "PDNode's name should be unique, get duplicate [%s]", + name); + } + nodes_.emplace_back(new PDNode(std::move(teller), name)); auto* cur = nodes_.back().get(); + node_map_[name] = cur; return cur; } +PDNode* PDPattern::RetriveNode(const std::string& id) const { + auto it = node_map_.find(id); + if (it == node_map_.end()) { + return nullptr; + } + + return it->second; +} + void PDPattern::AddEdge(PDNode* a, PDNode* b) { PADDLE_ENFORCE(a); PADDLE_ENFORCE(b); @@ -50,15 +69,18 @@ void GraphPatternDetecter::operator()(Graph* graph, } bool GraphPatternDetecter::MarkPDNodesInGraph(const ir::Graph& graph) { + VLOG(4) << "mark pdnodes in graph"; if (graph.Nodes().empty()) return false; for (auto& node : GraphTraits::DFS(graph)) { for (const auto& pdnode : pattern_.nodes()) { if (pdnode->Tell(&node)) { + VLOG(4) << "pdnode " << pdnode->name() << " marked"; pdnodes2nodes_[pdnode.get()].insert(&node); } } } + VLOG(3) << pdnodes2nodes_.size() << " nodes marked"; return !pdnodes2nodes_.empty(); } @@ -66,10 +88,20 @@ struct HitGroup { std::unordered_map roles; bool Match(Node* node, PDNode* pat) { + if (nodes_.count(node)) { + if (!roles.count(pat)) return false; + return roles[pat] == node; + } return !roles.count(pat) || roles.at(pat) == node; } - void Register(Node* node, PDNode* pat) { roles[pat] = node; } + void Register(Node* node, PDNode* pat) { + roles[pat] = node; + nodes_.insert(node); + } + + private: + std::unordered_set nodes_; }; // Tell whether Node a links to b. @@ -103,6 +135,7 @@ GraphPatternDetecter::DetectPatterns() { // Extend a PDNode to subgraphs by deducing the connection relations defined // in edges of PDNodes. for (const auto& edge : pattern_.edges()) { + VLOG(4) << "check " << edge.first->name() << " -> " << edge.second->name(); // Each role has two PDNodes, which indicates two roles. // Detect two Nodes that can match these two roles and they are connected. auto& pre_groups = bi_records[step % 2]; @@ -126,6 +159,7 @@ GraphPatternDetecter::DetectPatterns() { } } } + VLOG(3) << "step " << step << " get records: " << cur_groups.size(); } for (auto& group : bi_records[step % 2]) { diff --git a/paddle/fluid/framework/ir/graph_pattern_detecter.h b/paddle/fluid/framework/ir/graph_pattern_detecter.h index 1778bf0000..68c39902b5 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detecter.h +++ b/paddle/fluid/framework/ir/graph_pattern_detecter.h @@ -96,7 +96,8 @@ class PDPattern { void AddEdge(PDNode* a, PDNode* b); - PDNode* NewNode(PDNode::teller_t&& teller, const std::string& name = ""); + PDNode* NewNode(PDNode::teller_t&& teller, const std::string& name = NewID()); + PDNode* RetriveNode(const std::string& id) const; const std::vector>& nodes() const { return nodes_; } const std::vector& edges() const { return edges_; } @@ -107,8 +108,12 @@ class PDPattern { FRIEND_TEST(PDPattern, NewNode); #endif + static std::string NewID() { return "pdnode-" + std::to_string(id_++); } + std::vector> nodes_; std::vector edges_; + std::unordered_map node_map_; + static size_t id_; }; /* diff --git a/paddle/fluid/framework/ir/graph_test.cc b/paddle/fluid/framework/ir/graph_test.cc index b1b8d1c586..cadda49c39 100644 --- a/paddle/fluid/framework/ir/graph_test.cc +++ b/paddle/fluid/framework/ir/graph_test.cc @@ -200,9 +200,11 @@ TEST(GraphTest, WriteAfterWrite) { ASSERT_TRUE(ir::IsControlDepVar(*n->inputs[1])); control_dep2 = n->inputs[1]; ASSERT_EQ(n->inputs.size(), 2); - ASSERT_EQ(control_dep1, control_dep2); } } + ASSERT_NE(control_dep1, nullptr); + ASSERT_NE(control_dep2, nullptr); + ASSERT_EQ(control_dep1, control_dep2); } } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/ir/graph_traits.h b/paddle/fluid/framework/ir/graph_traits.h index edbe45acb9..f42bab20ed 100644 --- a/paddle/fluid/framework/ir/graph_traits.h +++ b/paddle/fluid/framework/ir/graph_traits.h @@ -12,7 +12,11 @@ // See the License for the specific language governing permissions and // limitations under the License. +#pragma once + #include +#include + #include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/framework/ir/node.h" diff --git a/paddle/fluid/framework/ir/graph_viz_pass.cc b/paddle/fluid/framework/ir/graph_viz_pass.cc index 8cb812d138..e7ff0c1dac 100644 --- a/paddle/fluid/framework/ir/graph_viz_pass.cc +++ b/paddle/fluid/framework/ir/graph_viz_pass.cc @@ -25,6 +25,7 @@ static const char kGraphVizPath[] = "graph_viz_path"; std::unique_ptr GraphVizPass::ApplyImpl( std::unique_ptr graph) const { const std::string graph_viz_path = Get(kGraphVizPath); + VLOG(3) << "draw IR graph viz to " << graph_viz_path; std::unique_ptr fout(new std::ofstream(graph_viz_path)); PADDLE_ENFORCE(fout->good()); std::ostream& sout = *fout; diff --git a/paddle/fluid/framework/ir/infer_clean_graph_pass.cc b/paddle/fluid/framework/ir/infer_clean_graph_pass.cc new file mode 100644 index 0000000000..f885567da1 --- /dev/null +++ b/paddle/fluid/framework/ir/infer_clean_graph_pass.cc @@ -0,0 +1,69 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" + +namespace paddle { +namespace framework { +namespace ir { + +class InferCleanGraphPass : public Pass { + public: + virtual ~InferCleanGraphPass() {} + + protected: + std::unique_ptr ApplyImpl(std::unique_ptr graph) const { + PADDLE_ENFORCE(graph.get()); + + auto is_valid_node = [](Node* x) { + return x && IsControlDepVar(*x) && x->IsVar() && !x->Var(); + }; + + std::unordered_set invalid_nodes; + for (auto* node : graph->Nodes()) { + if (is_valid_node(node)) { + invalid_nodes.insert(node); + } + } + + // remove nodes from the graph. + for (auto* node : invalid_nodes) { + graph->RemoveNode(node); + } + + // clean edges. + for (auto* node : graph->Nodes()) { + CleanEdges(&node->inputs, invalid_nodes); + CleanEdges(&node->outputs, invalid_nodes); + } + + return graph; + } + + void CleanEdges(std::vector* nodes, + const std::unordered_set& to_remove) const { + auto it = std::remove_if(nodes->begin(), nodes->end(), + [&](Node* x) { return to_remove.count(x); }); + nodes->erase(it, nodes->end()); + } +}; + +} // namespace ir +} // namespace framework +} // namespace paddle + +REGISTER_PASS(infer_clean_graph_pass, + paddle::framework::ir::InferCleanGraphPass); diff --git a/paddle/fluid/framework/ir/node.h b/paddle/fluid/framework/ir/node.h index 9c0765ab8c..063c70fb7b 100644 --- a/paddle/fluid/framework/ir/node.h +++ b/paddle/fluid/framework/ir/node.h @@ -34,14 +34,15 @@ class Node { explicit Node(VarDesc* var_desc) : name_(var_desc->Name()), - var_desc_(var_desc), + var_desc_(new VarDesc(*var_desc)), op_desc_(nullptr), type_(Type::kVariable) {} explicit Node(OpDesc* op_desc) : name_(op_desc->Type()), var_desc_(nullptr), - op_desc_(op_desc), + op_desc_(new OpDesc(*op_desc)), // TODO(panyx0718) the pointer in the + // original OpDesc might go out. type_(Type::kOperation) {} Type NodeType() const { return type_; } @@ -50,12 +51,12 @@ class Node { VarDesc* Var() { PADDLE_ENFORCE(type_ == Type::kVariable); - return var_desc_; + return var_desc_.get(); } OpDesc* Op() { - PADDLE_ENFORCE(type_ == Type::kOperation); - return op_desc_; + PADDLE_ENFORCE(IsOp()); + return op_desc_.get(); } bool IsOp() const { return type_ == Type::kOperation; } @@ -66,8 +67,8 @@ class Node { protected: const std::string name_; - VarDesc* var_desc_; - OpDesc* op_desc_; + std::unique_ptr var_desc_; + std::unique_ptr op_desc_; Type type_; private: diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 03f7e71c03..122dc161b4 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -202,6 +202,52 @@ std::vector OpDesc::AttrNames() const { } void OpDesc::SetAttr(const std::string &name, const Attribute &v) { + // NOTICE(minqiyang): pybind11 will take the empty list in python as + // the std::vector type in C++; so we have to change the attr's type + // here if we meet this issue + proto::AttrType attr_type = static_cast(v.which() - 1); + if (attr_type == proto::AttrType::INTS && + boost::get>(v).size() == 0u) { + // Find current attr via attr name and set the correct attribute value + const proto::OpProto::Attr &attr = GetProtoAttr(name); + switch (attr.type()) { + case proto::AttrType::BOOLEANS: { + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to BOOLEANS"; + this->attrs_[name] = std::vector(); + break; + } + case proto::AttrType::INTS: { + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to INTS"; + this->attrs_[name] = std::vector(); + break; + } + case proto::AttrType::FLOATS: { + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to FLOATS"; + this->attrs_[name] = std::vector(); + break; + } + case proto::AttrType::STRINGS: { + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to STRINGS"; + this->attrs_[name] = std::vector(); + break; + } + case proto::AttrType::BLOCKS: { + VLOG(11) << "SetAttr: " << Type() << ", " << name + << " from INTS to BLOCKS"; + this->SetBlocksAttr(name, std::vector()); + return; + } + default: + PADDLE_THROW("Wrong attr type %d", attr.type()); + } + need_update_ = true; + return; + } + this->attrs_[name] = v; need_update_ = true; } @@ -229,6 +275,19 @@ Attribute OpDesc::GetAttr(const std::string &name) const { return it->second; } +const proto::OpProto::Attr &OpDesc::GetProtoAttr( + const std::string &name) const { + const proto::OpProto &proto = OpInfoMap::Instance().Get(Type()).Proto(); + for (int i = 0; i != proto.attrs_size(); ++i) { + const proto::OpProto::Attr &attr = proto.attrs(i); + if (attr.name() == name) { + return attr; + } + } + + PADDLE_THROW("Attribute %s is not found in proto %s", name, proto.type()); +} + Attribute OpDesc::GetNullableAttr(const std::string &name) const { auto it = attrs_.find(name); if (it != attrs_.end()) { diff --git a/paddle/fluid/framework/op_desc.h b/paddle/fluid/framework/op_desc.h index b77d84125a..2422392e24 100644 --- a/paddle/fluid/framework/op_desc.h +++ b/paddle/fluid/framework/op_desc.h @@ -81,6 +81,8 @@ class OpDesc { Attribute GetAttr(const std::string &name) const; + const proto::OpProto::Attr &GetProtoAttr(const std::string &name) const; + Attribute GetNullableAttr(const std::string &name) const; int GetBlockAttrId(const std::string &name) const; diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc index 275cb8c592..81cb24bdda 100644 --- a/paddle/fluid/framework/parallel_executor.cc +++ b/paddle/fluid/framework/parallel_executor.cc @@ -25,6 +25,7 @@ limitations under the License. */ #include "paddle/fluid/platform/nccl_helper.h" #endif +#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h" #include "paddle/fluid/framework/details/multi_devices_graph_check_pass.h" #include "paddle/fluid/framework/details/multi_devices_graph_print_pass.h" #include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h" @@ -193,8 +194,14 @@ ParallelExecutor::ParallelExecutor( member_->local_scopes_, member_->use_cuda_, build_strategy); #endif - member_->executor_.reset(new details::ThreadedSSAGraphExecutor( - exec_strategy, member_->local_scopes_, places, std::move(graph))); + if (exec_strategy.type_ == ExecutionStrategy::kDefault) { + member_->executor_.reset(new details::ThreadedSSAGraphExecutor( + exec_strategy, member_->local_scopes_, places, std::move(graph))); + } else { + member_->executor_.reset(new details::FastThreadedSSAGraphExecutor( + exec_strategy, member_->local_scopes_, places, std::move(graph))); + } + member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor( exec_strategy, member_->local_scopes_, std::move(var_infos), member_->places_, std::move(member_->executor_))); diff --git a/paddle/fluid/framework/program_desc.cc b/paddle/fluid/framework/program_desc.cc index 20bdc7830f..344c001a69 100644 --- a/paddle/fluid/framework/program_desc.cc +++ b/paddle/fluid/framework/program_desc.cc @@ -55,11 +55,20 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) { auto all_ops = blocks_[block_id]->AllOps(); for (size_t op_id = 0; op_id < all_ops.size(); ++op_id) { auto &op = all_ops[op_id]; + for (const std::string &attr_name : op->AttrNames()) { if (op->GetAttrType(attr_name) == proto::AttrType::BLOCK) { int sub_block_id = o.Block(block_id).Op(op_id)->GetBlockAttrId(attr_name); op->SetBlockAttr(attr_name, MutableBlock(sub_block_id)); + } else if (op->GetAttrType(attr_name) == proto::AttrType::BLOCKS) { + std::vector sub_block_ids = + o.Block(block_id).Op(op_id)->GetBlocksAttrIds(attr_name); + std::vector block_descs; + for (int block_id : sub_block_ids) { + block_descs.push_back(MutableBlock(block_id)); + } + op->SetBlocksAttr(attr_name, block_descs); } } } @@ -68,24 +77,16 @@ ProgramDesc::ProgramDesc(const ProgramDesc &o) { ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) { desc_ = desc; - for (auto &block_desc : *desc_.mutable_blocks()) { - blocks_.emplace_back(new BlockDesc(this, &block_desc)); - } - for (auto &block : blocks_) { - for (auto *op : block->AllOps()) { - for (const auto &attr : op->Proto()->attrs()) { - if (attr.type() == proto::AttrType::BLOCK) { - size_t blk_idx = attr.block_idx(); - op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); - } - } - } - } + InitFromProto(); } ProgramDesc::ProgramDesc(const std::string &binary_str) { PADDLE_ENFORCE(desc_.ParseFromString(binary_str), "Fail to parse program_desc from binary string."); + InitFromProto(); +} + +void ProgramDesc::InitFromProto() { for (auto &block_desc : *desc_.mutable_blocks()) { blocks_.emplace_back(new BlockDesc(this, &block_desc)); } @@ -95,6 +96,13 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) { if (attr.type() == proto::AttrType::BLOCK) { size_t blk_idx = attr.block_idx(); op->SetBlockAttr(attr.name(), this->MutableBlock(blk_idx)); + } else if (attr.type() == proto::AttrType::BLOCKS) { + auto blks_idx = attr.blocks_idx(); + std::vector block_descs; + for (int blk_idx : blks_idx) { + block_descs.push_back(this->MutableBlock(blk_idx)); + } + op->SetBlocksAttr(attr.name(), block_descs); } } } diff --git a/paddle/fluid/framework/program_desc.h b/paddle/fluid/framework/program_desc.h index 65fa0a0cfd..f3afc85eb9 100644 --- a/paddle/fluid/framework/program_desc.h +++ b/paddle/fluid/framework/program_desc.h @@ -76,6 +76,8 @@ class ProgramDesc { void SetFetchHolderName(const std::string &fetch_holder_name); private: + void InitFromProto(); + proto::ProgramDesc desc_; std::vector> blocks_; diff --git a/paddle/fluid/framework/program_desc_test.cc b/paddle/fluid/framework/program_desc_test.cc index 6c46e9aad5..925ea98dbe 100644 --- a/paddle/fluid/framework/program_desc_test.cc +++ b/paddle/fluid/framework/program_desc_test.cc @@ -42,6 +42,19 @@ TEST(ProgramDesc, copy_ctor) { out->SetType(proto::VarType::LOD_TENSOR); op->SetOutput("Y", {out->Name()}); + BlockDesc* new_block = program.AppendBlock(*global_block); + op = new_block->AppendOp(); + op->SetType("mul"); + + op = global_block->AppendOp(); + op->SetType("op_with_subblock"); + op->SetAttr("sub_block", new_block); + + std::vector sub_blocks; + sub_blocks.push_back(program.AppendBlock(*global_block)); + sub_blocks.push_back(program.AppendBlock(*global_block)); + op->SetAttr("sub_blocks", sub_blocks); + ProgramDesc program_copy(program); auto* global_block_copy = program_copy.MutableBlock(0); @@ -64,6 +77,8 @@ TEST(ProgramDesc, copy_ctor) { assert_same_var("Y", y); assert_same_var("Out", out); + bool found_sub_block = false; + bool found_sub_blocks = false; for (size_t i = 0; i < global_block->OpSize(); ++i) { auto op_origin = global_block->Op(i); auto op_copy = global_block_copy->Op(i); @@ -74,8 +89,17 @@ TEST(ProgramDesc, copy_ctor) { ASSERT_EQ(op_copy->Proto()->SerializeAsString(), op_origin->Proto()->SerializeAsString()); - } + if (op->Type() == "op_with_subblock") { + ASSERT_EQ(1, op->GetBlockAttrId("sub_block")); + found_sub_block = true; + + ASSERT_EQ(2, op->GetBlocksAttrIds("sub_blocks").size()); + found_sub_blocks = true; + } + } + ASSERT_TRUE(found_sub_block); + ASSERT_TRUE(found_sub_blocks); // Not check block's protostr are same it because the order of vars could be // different and it is correct. } diff --git a/paddle/fluid/framework/rw_lock.h b/paddle/fluid/framework/rw_lock.h new file mode 100644 index 0000000000..1418fb5134 --- /dev/null +++ b/paddle/fluid/framework/rw_lock.h @@ -0,0 +1,48 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include + +#include "paddle/fluid/platform/enforce.h" + +namespace paddle { +namespace framework { + +struct RWLock { + RWLock() { pthread_rwlock_init(&lock_, nullptr); } + + ~RWLock() { pthread_rwlock_destroy(&lock_); } + + void RDLock() { + PADDLE_ENFORCE_EQ(pthread_rwlock_rdlock(&lock_), 0, + "acquire read lock failed"); + } + + void WRLock() { + PADDLE_ENFORCE_EQ(pthread_rwlock_wrlock(&lock_), 0, + "acquire write lock failed"); + } + + void UNLock() { + PADDLE_ENFORCE_EQ(pthread_rwlock_unlock(&lock_), 0, "unlock failed"); + } + + private: + pthread_rwlock_t lock_; +}; + +} // namespace framework +} // namespace paddle diff --git a/paddle/fluid/framework/rw_lock_test.cc b/paddle/fluid/framework/rw_lock_test.cc new file mode 100644 index 0000000000..16f9cbb652 --- /dev/null +++ b/paddle/fluid/framework/rw_lock_test.cc @@ -0,0 +1,81 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/framework/rw_lock.h" +#include +#include // NOLINT +#include // NOLINT +#include + +namespace f = paddle::framework; + +void f1(f::RWLock *lock) { + lock->RDLock(); + lock->UNLock(); +} + +TEST(RWLOCK, read_read) { + f::RWLock lock; + lock.RDLock(); + std::thread t1(f1, &lock); + std::thread t2(f1, &lock); + t1.join(); + t2.join(); + lock.UNLock(); +} + +void f2(f::RWLock *lock, std::vector *result) { + lock->RDLock(); + ASSERT_EQ(result->size(), 0UL); + lock->UNLock(); +} + +void f3(f::RWLock *lock, std::vector *result) { + lock->WRLock(); + result->push_back(1); + lock->UNLock(); +} + +TEST(RWLOCK, read_write) { + f::RWLock lock; + std::vector result; + + lock.RDLock(); + std::thread t1(f2, &lock, &result); + t1.join(); + std::thread t2(f3, &lock, &result); + std::this_thread::sleep_for(std::chrono::seconds(1)); + ASSERT_EQ(result.size(), 0UL); + lock.UNLock(); + t2.join(); + ASSERT_EQ(result.size(), 1UL); +} + +void f4(f::RWLock *lock, std::vector *result) { + lock->RDLock(); + ASSERT_EQ(result->size(), 1UL); + lock->UNLock(); +} + +TEST(RWLOCK, write_read) { + f::RWLock lock; + std::vector result; + + lock.WRLock(); + std::thread t1(f4, &lock, &result); + std::this_thread::sleep_for(std::chrono::seconds(1)); + result.push_back(1); + lock.UNLock(); + t1.join(); +} diff --git a/paddle/fluid/framework/selected_rows.cc b/paddle/fluid/framework/selected_rows.cc index 06ed87e7e8..c202b0a5be 100644 --- a/paddle/fluid/framework/selected_rows.cc +++ b/paddle/fluid/framework/selected_rows.cc @@ -120,66 +120,76 @@ bool SelectedRows::HasKey(int64_t key) const { : true; } -std::vector> SelectedRows::Get( - const std::vector& keys, framework::Tensor* value) const { +int64_t SelectedRows::AutoGrownIndex(int64_t key, bool auto_grown) { + rwlock_->RDLock(); + auto iter = id_to_index_.find(key); + if (iter == id_to_index_.end()) { + rwlock_->UNLock(); + if (!auto_grown) { + PADDLE_THROW("key %d not found", key); + } + rwlock_->WRLock(); + auto map_size = id_to_index_.size(); + auto vector_size = rows_.size(); + if (map_size != vector_size) { + rwlock_->UNLock(); + PADDLE_THROW( + "id_to_index_ size %d should have the same size with rows_ %d", + map_size, vector_size); + } + auto write_iter = id_to_index_.find(key); + if (write_iter == id_to_index_.end()) { + size_t row_num = rows_.size(); + if (row_num == value_->dims()[0]) { + rwlock_->UNLock(); + PADDLE_THROW("selected rows is full, then length exceed %d", row_num); + } + // key logic to put a key into id_to_index_ + rows_.push_back(key); + auto index = static_cast(rows_.size() - 1); + id_to_index_[key] = index; + rwlock_->UNLock(); + return index; + } else { + auto index = write_iter->second; + rwlock_->UNLock(); + return index; + } + } else { + auto index = iter->second; + rwlock_->UNLock(); + return index; + } +} + +void SelectedRows::SyncIndex() { + rwlock_->WRLock(); + id_to_index_.clear(); + for (size_t i = 0; i < rows_.size(); ++i) { + id_to_index_[rows_[i]] = i; + } + rwlock_->UNLock(); +} + +void SelectedRows::Get(const framework::Tensor& ids, framework::Tensor* value, + bool auto_grown) { PADDLE_ENFORCE(value->IsInitialized(), "The value tensor should be initialized."); - std::vector> non_keys_pair; - if (keys.empty()) { + if (ids.numel() == 0) { VLOG(3) << "keys is empty, please check data!"; } else { int64_t value_width = value_->numel() / value_->dims()[0]; PADDLE_ENFORCE_EQ(value_width, value->numel() / value->dims()[0], "output tensor should have the same shape with table " "except the dims[0]."); - - for (size_t i = 0; i < keys.size(); ++i) { - int64_t index = Index(keys[i]); - if (index == -1) { - non_keys_pair.push_back( - std::make_pair(keys[i], static_cast(i))); - } else { - framework::VisitDataType( - framework::ToDataType(value_->type()), - TensorCopyVisitor(value, i * value_width, *value_.get(), - index * value_width, value_width)); - } + for (size_t i = 0; i < ids.numel(); ++i) { + int64_t index = AutoGrownIndex(ids.data()[i], auto_grown); + framework::VisitDataType( + framework::ToDataType(value_->type()), + TensorCopyVisitor(value, i * value_width, *value_.get(), + index * value_width, value_width)); } } - return non_keys_pair; -} - -bool SelectedRows::Set(int64_t key, const framework::Tensor& value) { - PADDLE_ENFORCE(value.IsInitialized(), "The value should be initialized."); - if (value_->IsInitialized()) { - PADDLE_ENFORCE_EQ( - value.type(), value_->type(), - "The type of the value should be same with the original value"); - } - PADDLE_ENFORCE_EQ(value.dims()[0], static_cast(1), - "The first dim of value should be 1."); - std::lock_guard lock(*auto_grown_mutex_.get()); - auto index = Index(key); - bool is_new_key = false; - if (index == -1) { - rows_.push_back(key); - index = rows_.size() - 1; - is_new_key = true; - // whether need to resize the table - if (static_cast(rows_.size()) > value_->dims()[0]) { - auto dims = value_->dims(); - dims[0] = (dims[0] + 1) << 1; - framework::VisitDataType(framework::ToDataType(value.type()), - ReAllocateVisitor(dims, value_.get())); - } - } - - framework::VisitDataType( - framework::ToDataType(value.type()), - TensorCopyVisitor(value_.get(), - index * value_->numel() / value_->dims()[0], value, - static_cast(0), value.numel())); - return is_new_key; } } // namespace framework diff --git a/paddle/fluid/framework/selected_rows.h b/paddle/fluid/framework/selected_rows.h index 7160670ddd..daf5e95304 100644 --- a/paddle/fluid/framework/selected_rows.h +++ b/paddle/fluid/framework/selected_rows.h @@ -17,10 +17,12 @@ limitations under the License. */ #include #include #include // NOLINT +#include #include #include #include "paddle/fluid/framework/lod_tensor.h" +#include "paddle/fluid/framework/rw_lock.h" #include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/memory/memcpy.h" @@ -48,13 +50,13 @@ class SelectedRows { SelectedRows(const std::vector& rows, const int64_t& height) : rows_(rows), height_(height) { value_.reset(new Tensor()); - auto_grown_mutex_.reset(new std::mutex); + rwlock_.reset(new RWLock); } SelectedRows() { height_ = 0; value_.reset(new Tensor()); - auto_grown_mutex_.reset(new std::mutex); + rwlock_.reset(new RWLock); } platform::Place place() const { return value_->place(); } @@ -74,47 +76,51 @@ class SelectedRows { void set_rows(const Vector& rows) { rows_ = rows; } /* - * @brief wheter has the specified key in the table. + * @brief Get the index of key in rows + * + * @return -1 if the key does not exists. + */ + int64_t Index(int64_t key) const { + auto it = std::find(rows_.begin(), rows_.end(), key); + if (it == rows_.end()) { + PADDLE_THROW("id %s not in table", key); + } + return static_cast(std::distance(rows_.begin(), it)); + } + + /* + * @brief whether has the specified key in the table. * * @return true if the key is exists. */ bool HasKey(int64_t key) const; /* - * @brief Get value by the key list, if the + * @brief Get value by the key list. + * Note!!! this interface is only used when selected_rows is used as + * parameters + * for distribute lookup table. * * @return a list of pair which contains the non-exists key and the index in * the value */ - std::vector> Get(const std::vector& keys, - framework::Tensor* value) const; + void Get(const framework::Tensor& ids, framework::Tensor* value, + bool auto_grown = false); /* - * @brief Set a key-value pair into the table. - * This function will double the value memory if it's not engouth. + * @brief Get the index of the key from id_to_index_ map. If the key not + * exist, + * add the key into id_to_index_. * - * @note: - * 1. The first dim of the value should be 1 - * 2. The value should be initialized and the data type - * should be the same with the table. - * - * @return true if the key is a new one, otherwise false + * Note!!! this interface is only used when selected_rows is used as + * parameters + * for distribute lookup table. * + * @return index of the key. */ - bool Set(int64_t key, const Tensor& value); + int64_t AutoGrownIndex(int64_t key, bool auto_grown); - /* - * @brief Get the index of key in rows - * - * @return -1 if the key does not exists. - */ - int64_t Index(int64_t key) const { - auto it = std::find(rows_.begin(), rows_.end(), key); - if (it == rows_.end()) { - return static_cast(-1); - } - return static_cast(std::distance(rows_.begin(), it)); - } + void SyncIndex(); DDim GetCompleteDims() const { std::vector dims = vectorize(value_->dims()); @@ -127,9 +133,10 @@ class SelectedRows { // SelectedRows are simply concated when adding together. Until a // SelectedRows add a Tensor, will the duplicate rows be handled. Vector rows_; + std::unordered_map id_to_index_; std::unique_ptr value_{nullptr}; int64_t height_; - std::unique_ptr auto_grown_mutex_{nullptr}; + std::unique_ptr rwlock_{nullptr}; }; /* diff --git a/paddle/fluid/framework/selected_rows_test.cc b/paddle/fluid/framework/selected_rows_test.cc index eefcaa5672..5ca864cfdf 100644 --- a/paddle/fluid/framework/selected_rows_test.cc +++ b/paddle/fluid/framework/selected_rows_test.cc @@ -9,8 +9,11 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/framework/selected_rows.h" +#include +#include // NOLINT + #include "gtest/gtest.h" +#include "paddle/fluid/framework/selected_rows.h" namespace paddle { namespace framework { @@ -59,39 +62,129 @@ TEST_F(SelectedRowsTester, SerializeAndDeseralize) { ASSERT_EQ(selected_rows_->GetCompleteDims(), dst_tensor.GetCompleteDims()); } -TEST_F(SelectedRowsTester, SparseTable) { +TEST(SelectedRows, SparseTable) { platform::CPUPlace cpu; SelectedRows table; + + int64_t table_size = 100; + int64_t embedding_width = 8; // initialize a sparse table - table.mutable_value()->Resize(framework::make_ddim({1, 100})); - table.mutable_value()->mutable_data(cpu); - table.mutable_rows()->push_back(1); + table.mutable_value()->Resize( + framework::make_ddim({table_size, embedding_width})); + auto* data = table.mutable_value()->mutable_data(cpu); + for (int64_t i = 0; i < table_size; ++i) { + for (int64_t j = 0; j < embedding_width; ++j) { + data[i * embedding_width + j] = static_cast(i); + } + } + ASSERT_EQ(table.AutoGrownIndex(10, true), 0); + ASSERT_EQ(table.AutoGrownIndex(8, true), 1); + ASSERT_EQ(table.AutoGrownIndex(8, true), 1); + ASSERT_EQ(table.AutoGrownIndex(6, true), 2); + ASSERT_TRUE(table.HasKey(10)); + ASSERT_TRUE(table.HasKey(8)); + ASSERT_TRUE(table.HasKey(6)); + ASSERT_EQ(table.rows().size(), 3); + + framework::Tensor ids; + ids.Resize(framework::make_ddim({4})); + auto* ids_data = ids.mutable_data(cpu); + ids_data[0] = static_cast(6); + ids_data[1] = static_cast(6); + ids_data[2] = static_cast(8); + ids_data[3] = static_cast(10); - int64_t key = 10000; - int64_t non_key = 999; - framework::Tensor value; - value.Resize(framework::make_ddim({1, 100})); - auto ptr = value.mutable_data(cpu); - ptr[0] = static_cast(10); + framework::Tensor get_value; + auto* value_data = get_value.mutable_data( + framework::make_ddim({4, embedding_width}), cpu); + table.Get(ids, &get_value); - ASSERT_EQ(table.rows().size(), static_cast(1)); - ASSERT_EQ(table.HasKey(key), false); + for (int j = 0; j < embedding_width; ++j) { + ASSERT_EQ(value_data[0 * embedding_width + j], 2); + } + for (int j = 0; j < embedding_width; ++j) { + ASSERT_EQ(value_data[1 * embedding_width + j], 2); + } + for (int j = 0; j < embedding_width; ++j) { + ASSERT_EQ(value_data[2 * embedding_width + j], 1); + } + for (int j = 0; j < embedding_width; ++j) { + ASSERT_EQ(value_data[3 * embedding_width + j], 0); + } +} - table.Set(key, value); +void f1(SelectedRows* table, int table_size) { + for (int i = 1000000; i > 0; --i) { + auto id = i % table_size; + int64_t index1 = table->AutoGrownIndex(id, true); + int64_t index2 = table->AutoGrownIndex(id, false); + int64_t index3 = table->AutoGrownIndex(id, true); + ASSERT_EQ(index1, index2); + ASSERT_EQ(index2, index3); + } +} - ASSERT_EQ(table.rows().size(), static_cast(2)); - ASSERT_EQ(table.HasKey(key), true); - // check re-allocate - ASSERT_EQ(table.value().dims()[0], static_cast(4)); +void f2(SelectedRows* table, int table_size) { + for (int i = 0; i < 1000000; ++i) { + auto id = i % table_size; + int64_t index1 = table->AutoGrownIndex(id, true); + int64_t index2 = table->AutoGrownIndex(id, false); + int64_t index3 = table->AutoGrownIndex(id, true); + ASSERT_EQ(index1, index2); + ASSERT_EQ(index2, index3); + } +} - framework::Tensor get_value; - get_value.mutable_data(framework::make_ddim({2, 100}), cpu); - std::vector keys({non_key, key}); - auto non_key_pairs = table.Get(keys, &get_value); +void f3(SelectedRows* table, int table_size) { + clock_t t1 = clock(); + for (int i = 100000; i > 0; --i) { + auto id1 = table->AutoGrownIndex(i % table_size, true); + auto id2 = table->Index(i % table_size); + ASSERT_EQ(id1, id2); + } + clock_t t2 = clock(); + std::cout << "f3 run time:" << t2 - t1 << std::endl; +} + +void f4(SelectedRows* table, int table_size) { + clock_t t1 = clock(); + for (int i = 0; i < 100000; ++i) { + auto id1 = table->AutoGrownIndex(i % table_size, true); + auto id2 = table->Index(i % table_size); + ASSERT_EQ(id1, id2); + } + clock_t t2 = clock(); + std::cout << "f4 run time:" << t2 - t1 << std::endl; +} + +TEST(SelectedRows, MultiThreadAutoIndex) { + platform::CPUPlace cpu; + SelectedRows table; + + int64_t table_size = 100000; + int64_t embedding_width = 8; + // initialize a sparse table + table.mutable_value()->Resize( + framework::make_ddim({table_size, embedding_width})); + auto* data = table.mutable_value()->mutable_data(cpu); + for (int64_t i = 0; i < table_size; ++i) { + for (int64_t j = 0; j < embedding_width; ++j) { + data[i * embedding_width + j] = static_cast(i); + } + } - ASSERT_EQ(get_value.data()[100], static_cast(10)); - ASSERT_EQ(non_key_pairs.size(), static_cast(1)); - ASSERT_EQ(non_key_pairs[0].first, non_key); + std::thread t1(f1, &table, table_size); + std::thread t11(f1, &table, table_size); + std::thread t2(f2, &table, table_size); + std::thread t22(f2, &table, table_size); + t1.join(); + t11.join(); + t2.join(); + t22.join(); + std::thread t3(f3, &table, table_size); + std::thread t4(f4, &table, table_size); + t3.join(); + t4.join(); } } // namespace framework diff --git a/paddle/fluid/inference/analysis/CMakeLists.txt b/paddle/fluid/inference/analysis/CMakeLists.txt index 27fe575cb6..4feaed2b0d 100644 --- a/paddle/fluid/inference/analysis/CMakeLists.txt +++ b/paddle/fluid/inference/analysis/CMakeLists.txt @@ -1,13 +1,17 @@ +cc_library(ir_pass_manager SRCS ir_pass_manager.cc DEPS graph pass) cc_library(analysis SRCS pass_manager.cc dot.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc + analyzer.cc + helper.cc + # passes fluid_to_data_flow_graph_pass.cc data_flow_graph_to_fluid_pass.cc dfg_graphviz_draw_pass.cc tensorrt_subgraph_pass.cc tensorrt_subgraph_node_mark_pass.cc - analyzer.cc - helper.cc - model_store_pass.cc - DEPS framework_proto proto_desc) + fluid_to_ir_pass.cc + model_store_pass.cc + DEPS framework_proto proto_desc ir_pass_manager graph pass) + cc_test(test_node SRCS node_tester.cc DEPS analysis) cc_test(test_dot SRCS dot_tester.cc DEPS analysis) cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis) @@ -18,7 +22,7 @@ function (inference_analysis_test TARGET) if(WITH_TESTING) set(options "") set(oneValueArgs "") - set(multiValueArgs SRCS) + set(multiValueArgs SRCS EXTRA_DEPS) cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) set(mem_opt "") @@ -27,19 +31,51 @@ function (inference_analysis_test TARGET) endif() cc_test(${TARGET} SRCS "${analysis_test_SRCS}" - DEPS analysis + DEPS analysis graph fc_fuse_pass graph_viz_pass infer_clean_graph_pass graph_pattern_detecter pass ${analysis_test_EXTRA_DEPS} ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt}) set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec) endif(WITH_TESTING) endfunction(inference_analysis_test) +set(DITU_RNN_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fmodel.tar.gz") +set(DITU_RNN_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fdata.txt.tar.gz") +set(DITU_INSTALL_DIR "${THIRD_PARTY_PATH}/install/ditu_rnn" CACHE PATH "Ditu RNN model and data root." FORCE) +set(DITU_RNN_MODEL ${DITU_INSTALL_DIR}/model) +set(DITU_RNN_DATA ${DITU_INSTALL_DIR}/data.txt) + +function (inference_download_and_uncompress target url gz_filename) + message(STATUS "Download inference test stuff ${gz_filename} from ${url}") + execute_process(COMMAND bash -c "mkdir -p ${DITU_INSTALL_DIR}") + execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && wget -q ${url}") + execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && tar xzf ${gz_filename}") + message(STATUS "finish downloading ${gz_filename}") +endfunction(inference_download_and_uncompress) + +if (NOT EXISTS ${DITU_INSTALL_DIR}) + inference_download_and_uncompress(ditu_rnn_model ${DITU_RNN_MODEL_URL} "ditu_rnn_fluid%2Fmodel.tar.gz") + inference_download_and_uncompress(ditu_rnn_data ${DITU_RNN_DATA_URL} "ditu_rnn_fluid%2Fdata.txt.tar.gz") +endif() + +inference_analysis_test(test_analyzer SRCS analyzer_tester.cc + EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis + # ir + fc_fuse_pass + graph_viz_pass + infer_clean_graph_pass + graph_pattern_detecter + infer_clean_graph_pass + pass + ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model + --infer_ditu_rnn_model=${DITU_INSTALL_DIR}/model + --infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt) + inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc) inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc) +inference_analysis_test(test_fluid_to_ir_pass SRCS fluid_to_ir_pass_tester.cc) inference_analysis_test(test_fluid_to_data_flow_graph_pass SRCS fluid_to_data_flow_graph_pass_tester.cc) inference_analysis_test(test_subgraph_splitter SRCS subgraph_splitter_tester.cc) inference_analysis_test(test_dfg_graphviz_draw_pass SRCS dfg_graphviz_draw_pass_tester.cc) inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_tester.cc) inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc) inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc) -inference_analysis_test(test_analyzer SRCS analyzer_tester.cc) inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc) diff --git a/paddle/fluid/inference/analysis/analyzer.cc b/paddle/fluid/inference/analysis/analyzer.cc index 9318f10897..7d16364609 100644 --- a/paddle/fluid/inference/analysis/analyzer.cc +++ b/paddle/fluid/inference/analysis/analyzer.cc @@ -17,22 +17,23 @@ #include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h" #include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h" #include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h" +#include "paddle/fluid/inference/analysis/fluid_to_ir_pass.h" #include "paddle/fluid/inference/analysis/model_store_pass.h" #include "paddle/fluid/inference/analysis/pass_manager.h" #include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h" #include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h" -namespace paddle { - -DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, true, +DEFINE_bool(IA_enable_tensorrt_subgraph_engine, false, "Enable subgraph to TensorRT engine for acceleration"); -DEFINE_string(inference_analysis_graphviz_log_root, "./", +DEFINE_bool(IA_enable_ir, false, "Turn on IR support"); + +DEFINE_string(IA_graphviz_log_root, "./", "Graphviz debuger for data flow graphs."); -DEFINE_string(inference_analysis_output_storage_path, "", - "optimized model output path"); +DEFINE_string(IA_output_storage_path, "", "optimized model output path"); +namespace paddle { namespace inference { namespace analysis { @@ -40,11 +41,38 @@ class DfgPassManagerImpl final : public DfgPassManager { public: DfgPassManagerImpl() { // TODO(Superjomn) set the key with pass reprs. - AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass); - if (FLAGS_inference_analysis_enable_tensorrt_subgraph_engine) { + LOG(INFO) + << "-----------------------------------------------------------------"; + if (FLAGS_IA_enable_ir) { + AddPass("fluid-to-ir-pass", new FluidToIrPass); + } else { + AddPass("fluid-to-data-flow-graph", new FluidToDataFlowGraphPass); + } + TryAddTensorRtPass(); + AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass); + if (!FLAGS_IA_output_storage_path.empty()) { + AddPass("model-store-pass", new ModelStorePass); + } + LOG(INFO) + << "-----------------------------------------------------------------"; + } + + std::string repr() const override { return "dfg-pass-manager"; } + std::string description() const override { return "DFG pass manager."; } + + private: + void AddPass(const std::string& name, Pass* pass) { + VLOG(3) << "Adding pass " << name; + Register(name, pass); + AddGraphvizDebugerPass(pass); + } + + void TryAddTensorRtPass() { + if (FLAGS_IA_enable_tensorrt_subgraph_engine) { auto trt_teller = [&](const Node* node) { std::unordered_set teller_set( - {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax"}); + {"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax", + "depthwise_conv2d", "batch_norm"}); if (!node->IsFunction()) return false; const auto* func = static_cast(node); @@ -59,20 +87,6 @@ class DfgPassManagerImpl final : public DfgPassManager { new TensorRTSubgraphNodeMarkPass(trt_teller)); AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller)); } - AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass); - if (!FLAGS_inference_analysis_output_storage_path.empty()) { - AddPass("model-store-pass", new ModelStorePass); - } - } - - std::string repr() const override { return "dfg-pass-manager"; } - std::string description() const override { return "DFG pass manager."; } - - private: - void AddPass(const std::string& name, Pass* pass) { - LOG(INFO) << "Adding pass " << name; - Register(name, pass); - AddGraphvizDebugerPass(pass); } // Add the graphviz debuger pass if the parent pass has one. diff --git a/paddle/fluid/inference/analysis/analyzer.h b/paddle/fluid/inference/analysis/analyzer.h index c82fdfff86..2e107c82dd 100644 --- a/paddle/fluid/inference/analysis/analyzer.h +++ b/paddle/fluid/inference/analysis/analyzer.h @@ -39,14 +39,14 @@ limitations under the License. */ #include "paddle/fluid/inference/analysis/pass.h" #include "paddle/fluid/inference/analysis/pass_manager.h" -namespace paddle { - // TODO(Superjomn) add a definition flag like PADDLE_WITH_TENSORRT and hide this // flag if not available. -DECLARE_bool(inference_analysis_enable_tensorrt_subgraph_engine); -DECLARE_string(inference_analysis_graphviz_log_root); -DECLARE_string(inference_analysis_output_storage_path); +DECLARE_bool(IA_enable_tensorrt_subgraph_engine); +DECLARE_string(IA_graphviz_log_root); +DECLARE_string(IA_output_storage_path); +DECLARE_bool(IA_enable_ir); +namespace paddle { namespace inference { namespace analysis { diff --git a/paddle/fluid/inference/analysis/analyzer_tester.cc b/paddle/fluid/inference/analysis/analyzer_tester.cc index 24bfb3993c..52f5c4f5ae 100644 --- a/paddle/fluid/inference/analysis/analyzer_tester.cc +++ b/paddle/fluid/inference/analysis/analyzer_tester.cc @@ -13,15 +13,23 @@ // limitations under the License. #include "paddle/fluid/inference/analysis/analyzer.h" + #include +#include +#include "paddle/fluid/framework/ir/pass.h" #include "paddle/fluid/inference/analysis/ut_helper.h" +#include "paddle/fluid/inference/api/helper.h" +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +DEFINE_string(infer_ditu_rnn_model, "", "model path for ditu RNN"); +DEFINE_string(infer_ditu_rnn_data, "", "data path for ditu RNN"); namespace paddle { namespace inference { namespace analysis { TEST(Analyzer, analysis_without_tensorrt) { - FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = false; + FLAGS_IA_enable_tensorrt_subgraph_engine = false; Argument argument; argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir)); Analyzer analyser; @@ -29,13 +37,327 @@ TEST(Analyzer, analysis_without_tensorrt) { } TEST(Analyzer, analysis_with_tensorrt) { - FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = true; + FLAGS_IA_enable_tensorrt_subgraph_engine = true; Argument argument; argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir)); Analyzer analyser; analyser.Run(&argument); } +void TestWord2vecPrediction(const std::string &model_path) { + NativeConfig config; + config.model_dir = model_path; + config.use_gpu = false; + config.device = 0; + auto predictor = + ::paddle::CreatePaddlePredictor( + config); + + // One single batch + + int64_t data[4] = {1, 2, 3, 4}; + PaddleTensor tensor; + tensor.shape = std::vector({4, 1}); + tensor.data = PaddleBuf(data, sizeof(data)); + tensor.dtype = PaddleDType::INT64; + + // For simplicity, we set all the slots with the same data. + std::vector slots(4, tensor); + std::vector outputs; + CHECK(predictor->Run(slots, &outputs)); + + PADDLE_ENFORCE(outputs.size(), 1UL); + // Check the output buffer size and result of each tid. + PADDLE_ENFORCE(outputs.front().data.length(), 33168UL); + float result[5] = {0.00129761, 0.00151112, 0.000423564, 0.00108815, + 0.000932706}; + const size_t num_elements = outputs.front().data.length() / sizeof(float); + // The outputs' buffers are in CPU memory. + for (size_t i = 0; i < std::min(5UL, num_elements); i++) { + LOG(INFO) << "data: " + << static_cast(outputs.front().data.data())[i]; + PADDLE_ENFORCE(static_cast(outputs.front().data.data())[i], + result[i]); + } +} + +namespace { + +struct DataRecord { + std::vector>> link_step_data_all; + std::vector> week_data_all, minute_data_all; + std::vector lod1, lod2, lod3; + std::vector> rnn_link_data, rnn_week_datas, + rnn_minute_datas; + size_t batch_iter{0}; + size_t batch_size{1}; + DataRecord() = default; + DataRecord(const std::string &path, int batch_size = 1) + : batch_size(batch_size) { + Load(path); + } + DataRecord NextBatch() { + DataRecord data; + size_t batch_end = batch_iter + batch_size; + // NOTE skip the final batch, if no enough data is provided. + if (batch_end <= link_step_data_all.size()) { + data.link_step_data_all.assign(link_step_data_all.begin() + batch_iter, + link_step_data_all.begin() + batch_end); + data.week_data_all.assign(week_data_all.begin() + batch_iter, + week_data_all.begin() + batch_end); + data.minute_data_all.assign(minute_data_all.begin() + batch_iter, + minute_data_all.begin() + batch_end); + // Prepare LoDs + data.lod1.push_back(0); + data.lod2.push_back(0); + data.lod3.push_back(0); + CHECK(!data.link_step_data_all.empty()) << "empty"; + CHECK(!data.week_data_all.empty()); + CHECK(!data.minute_data_all.empty()); + CHECK_EQ(data.link_step_data_all.size(), data.week_data_all.size()); + CHECK_EQ(data.minute_data_all.size(), data.link_step_data_all.size()); + for (size_t j = 0; j < data.link_step_data_all.size(); j++) { + for (const auto &d : data.link_step_data_all[j]) { + data.rnn_link_data.push_back(d); + } + data.rnn_week_datas.push_back(data.week_data_all[j]); + data.rnn_minute_datas.push_back(data.minute_data_all[j]); + // calculate lod + data.lod1.push_back(data.lod1.back() + + data.link_step_data_all[j].size()); + data.lod3.push_back(data.lod3.back() + 1); + for (size_t i = 1; i < data.link_step_data_all[j].size() + 1; i++) { + data.lod2.push_back(data.lod2.back() + + data.link_step_data_all[j].size()); + } + } + } + batch_iter += batch_size; + return data; + } + void Load(const std::string &path) { + std::ifstream file(path); + std::string line; + int num_lines = 0; + while (std::getline(file, line)) { + num_lines++; + std::vector data; + split(line, ':', &data); + std::vector> link_step_data; + std::vector link_datas; + split(data[0], '|', &link_datas); + for (auto &step_data : link_datas) { + std::vector tmp; + split_to_float(step_data, ',', &tmp); + link_step_data.push_back(tmp); + } + // load week data + std::vector week_data; + split_to_float(data[2], ',', &week_data); + // load minute data + std::vector minute_data; + split_to_float(data[1], ',', &minute_data); + link_step_data_all.push_back(std::move(link_step_data)); + week_data_all.push_back(std::move(week_data)); + minute_data_all.push_back(std::move(minute_data)); + } + } +}; +void PrepareInputs(std::vector *input_slots, DataRecord *data, + int batch_size) { + // DataRecord data(FLAGS_datapath, batch_size); + PaddleTensor lod_attention_tensor, init_zero_tensor, lod_tensor_tensor, + week_tensor, minute_tensor; + lod_attention_tensor.name = "data_lod_attention"; + init_zero_tensor.name = "cell_init"; + lod_tensor_tensor.name = "data"; + week_tensor.name = "week"; + minute_tensor.name = "minute"; + auto one_batch = data->NextBatch(); + // clang-format off + std::vector rnn_link_data_shape + ({static_cast(one_batch.rnn_link_data.size()), static_cast(one_batch.rnn_link_data.front().size())}); + lod_attention_tensor.shape.assign({1, 2}); + lod_attention_tensor.lod.assign({one_batch.lod1, one_batch.lod2}); + init_zero_tensor.shape.assign({batch_size, 15}); + init_zero_tensor.lod.assign({one_batch.lod3}); + lod_tensor_tensor.shape = rnn_link_data_shape; + lod_tensor_tensor.lod.assign({one_batch.lod1}); + week_tensor.shape.assign({(int) one_batch.rnn_week_datas.size(), (int) one_batch.rnn_week_datas.front().size()}); + week_tensor.lod.assign({one_batch.lod3}); + minute_tensor.shape.assign({(int) one_batch.rnn_minute_datas.size(), + (int) one_batch.rnn_minute_datas.front().size()}); + minute_tensor.lod.assign({one_batch.lod3}); + // assign data + TensorAssignData(&lod_attention_tensor, std::vector>({{0, 0}})); + std::vector tmp_zeros(batch_size * 15, 0.); + TensorAssignData(&init_zero_tensor, {tmp_zeros}); + TensorAssignData(&lod_tensor_tensor, one_batch.rnn_link_data); + TensorAssignData(&week_tensor, one_batch.rnn_week_datas); + TensorAssignData(&minute_tensor, one_batch.rnn_minute_datas); + // clang-format on + // Set inputs. + auto init_zero_tensor1 = init_zero_tensor; + init_zero_tensor1.name = "hidden_init"; + input_slots->assign({week_tensor, init_zero_tensor, minute_tensor, + init_zero_tensor1, lod_attention_tensor, + lod_tensor_tensor}); + for (auto &tensor : *input_slots) { + tensor.dtype = PaddleDType::FLOAT32; + } +} + +std::string DescribeTensor(const PaddleTensor &tensor) { + std::stringstream os; + os << "Tensor [" << tensor.name << "]\n"; + os << " - type: "; + switch (tensor.dtype) { + case PaddleDType::FLOAT32: + os << "float32"; + break; + case PaddleDType::INT64: + os << "int64"; + break; + default: + os << "unset"; + } + os << '\n'; + + os << " - shape: " << to_string(tensor.shape) << '\n'; + os << " - lod: "; + for (auto &l : tensor.lod) { + os << to_string(l) << "; "; + } + os << "\n"; + os << " - data: "; + + // clang-format off + int dim = std::accumulate(tensor.shape.begin(), + tensor.shape.end(), + 1, + [](int a, int b) { return a * b; }); // clang-format on + for (size_t i = 0; i < dim; i++) { + os << static_cast(tensor.data.data())[i] << " "; + } + os << '\n'; + return os.str(); +} + +} // namespace + +const float ditu_rnn_target_data[] = { + 104.711, 11.2431, 1.35422, 0, 0, 0, 0, 0, + 27.7039, 1.41486, 7.09526, 0, 0, 0, 0, 0, + 7.6481, 6.5324, 56.383, 2.88018, 8.92918, 132.007, 4.27429, 2.02934, + 14.1727, 10.7461, 25.0616, 16.0197, 14.4163, 16.9199, 6.75517, 0, + 80.0249, 4.77739, 0, 0, 0, 0, 0, 0, + 47.5643, 2.67029, 8.76252, 0, 0, 0, 0, 0, + 51.8822, 4.4411, 0, 0, 0, 0, 0, 0, + 10.7286, 12.0595, 10.6672, 0, 0, 0, 0, 0, + 93.5771, 3.84641, 0, 0, 0, 0, 0, 0, + 169.426, 0, 0, 0, 0, 0, 0, 0}; +// Test with a really complicate model. +void TestDituRNNPrediction(const std::string &model_path, + const std::string &data_path, int batch_size, + bool use_analysis, bool activate_ir, + int num_times = 1) { + FLAGS_IA_enable_ir = activate_ir; + FLAGS_IA_enable_tensorrt_subgraph_engine = false; + FLAGS_IA_output_storage_path = "./analysis.out"; + + std::string model_out; + if (use_analysis) { + Argument argument(model_path); + argument.model_output_store_path.reset(new std::string("./analysis.out")); + + Analyzer analyzer; + analyzer.Run(&argument); + + // Should get the transformed model stored to ./analysis.out + model_out = "./analysis.out"; + ASSERT_TRUE(PathExists(model_out)); + } else { + model_out = FLAGS_infer_ditu_rnn_model; + } + + NativeConfig config; + config.prog_file = model_out + "/__model__"; + config.param_file = model_out + "/param"; + config.use_gpu = false; + config.device = 0; + config.specify_input_name = true; + + auto predictor = + CreatePaddlePredictor(config); + std::vector input_slots; + DataRecord data(data_path, batch_size); + // Prepare inputs. + PrepareInputs(&input_slots, &data, batch_size); + std::vector outputs; + + Timer timer; + timer.tic(); + for (int i = 0; i < num_times; i++) { + predictor->Run(input_slots, &outputs); + } + LOG(INFO) << "time/batch: " << timer.toc() / num_times; + + for (auto &out : outputs) { + size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1, + [](int a, int b) { return a * b; }); + float *data = static_cast(out.data.data()); + for (int i = 0; + i < std::min(sizeof(ditu_rnn_target_data) / sizeof(float), size); + i++) { + EXPECT_NEAR(data[i], ditu_rnn_target_data[i], 1e-3); + } + } +} + +// Turn on the IR pass supportion, run a real inference and check the result. +TEST(Analyzer, SupportIRPass) { + FLAGS_IA_enable_ir = true; + FLAGS_IA_enable_tensorrt_subgraph_engine = false; + FLAGS_IA_output_storage_path = "./analysis.out"; + + Argument argument(FLAGS_inference_model_dir); + argument.model_output_store_path.reset(new std::string("./analysis.out")); + + Analyzer analyzer; + analyzer.Run(&argument); + + // Should get the transformed model stored to ./analysis.out + ASSERT_TRUE(PathExists("./analysis.out")); + + // Inference from this path. + TestWord2vecPrediction("./analysis.out"); +} + +// Directly infer with the original model. +TEST(Analyzer, DituRNN_without_analysis) { + TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data, + 10, false, false); +} + +// Inference with the original model with the analysis turned on, the analysis +// module will transform the program to a data flow graph. +TEST(Analyzer, DituRNN_with_analysis) { + LOG(INFO) << "ditu rnn with analysis"; + TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data, + 10, true, false, 1); +} + +// Inference with analysis and IR. The IR module will fuse some large kernels. +TEST(Analyzer, DituRNN_with_analysis_with_IR) { + LOG(INFO) << "ditu rnn with analysis and IR fuse"; + TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data, + 10, true, true, 1); +} + } // namespace analysis } // namespace inference } // namespace paddle + +USE_PASS(fc_fuse_pass); +USE_PASS(graph_viz_pass); +USE_PASS(infer_clean_graph_pass); diff --git a/paddle/fluid/inference/analysis/data_flow_graph.cc b/paddle/fluid/inference/analysis/data_flow_graph.cc index 7f64bc75ae..100a7504b8 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph.cc @@ -19,14 +19,16 @@ limitations under the License. */ namespace paddle { namespace inference { namespace analysis { +using ir_node_t = framework::ir::Node; +using ir_graph_t = framework::ir::Graph; // It is a better idea that the inputs and outputs of this graph is set manually // before, but there must be a Pass that helps to prune the unnecessary ops that // do not contribute to the given targets, so in this pass, analysis and get the // inputs and outputs is OK. void DataFlowGraph::Build() { - inputs.clear(); - outputs.clear(); + inputs_.clear(); + outputs_.clear(); std::unordered_set ins; std::unordered_set outs; for (auto &node : nodes.nodes()) { @@ -42,18 +44,140 @@ void DataFlowGraph::Build() { // similarly, the nodes that in outs but not in ins is the graphs' outputs for (auto *in : ins) { if (!outs.count(in)) { - inputs.push_back(in); + inputs_.push_back(in); } } for (auto *out : outs) { - if (!outs.count(out)) { - outputs.push_back(out); + if (!ins.count(out)) { + outputs_.push_back(out); } } Clean(); } +void DataFlowGraph::Build(const framework::proto::ProgramDesc &prog) { + // insert vars + // The `var2id` keeps a map from a variable's name to its Node-id, the Node-id + // will keep updating to its latest alias during the graph-building. + std::unordered_map var2id; + auto &main_block = prog.blocks(framework::kRootBlockIndex); + for (int i = 0; i < main_block.vars_size(); i++) { + const auto &var = main_block.vars(i); + auto *v = nodes.Create(Node::Type::kValue); + v->SetName(var.name()); + v->SetPbDesc(const_cast(static_cast(&var))); + v->SetPbMsg(var.SerializeAsString()); + var2id[var.name()] = v->id(); + } + + // The variables in a SSA can only write once, so if a variable is written + // multiple times(quite common in our ProgramDesc design), multiple alias + // Nodes of this variable will be created, and each will just write once. + + // An set that keep all the names of the variables(the original, not alias) + // that have been written(as outputs). Once an Op's output variable hit the + // set, it should create a new alias and update the global alias for this + // variable. And that make a Data Flow Graph a SSA. + std::unordered_set unique_written_vars; + for (int i = 0; i < main_block.ops_size(); i++) { + const auto &op = main_block.ops(i); + auto *o = nodes.Create(Node::Type::kFunction); + o->SetName(op.type()); + static_cast(o)->SetFuncType(op.type()); + // Link to the original protobuf message's memory, make it easier to + // generate from a data flow graph to fluid ProgramDesc. + o->SetPbDesc(const_cast(static_cast(&op))); + o->SetPbMsg(op.SerializeAsString()); + + // set inputs and outputs + for (int j = 0; j < op.inputs_size(); j++) { + auto &in_var = op.inputs(j); + for (int k = 0; k < in_var.arguments_size(); k++) { + auto *in = nodes.GetMutable(var2id.at(in_var.arguments(k))); + in->outlinks.push_back(o); + o->inlinks.push_back(in); + unique_written_vars.insert(in); + } + } + for (int j = 0; j < op.outputs_size(); j++) { + auto &out_var = op.outputs(j); + for (int k = 0; k < out_var.arguments_size(); k++) { + auto *out = nodes.GetMutable(var2id[out_var.arguments(k)]); + if (unique_written_vars.count(out)) { + // Loop found, for example, a = op(a), use SSA, change to a1 = op(a). + auto *out_alias = nodes.Create(Node::Type::kValue); + out_alias->SetName(out->name()); + out_alias->SetPbDesc(out->pb_desc()); + out_alias->SetPbMsg(out->pb_msg()); + var2id[out_alias->name()] = + out_alias->id(); // update variable's alias Node + LOG(INFO) << "loop found in graph, create SSA alias node [" + << out_alias->repr() << "] for [" << out->repr() << "]"; + out = out_alias; + } + out->inlinks.push_back(o); + o->outlinks.push_back(out); + } + } + } + // Analysis and extract the inputs and outputs of this graph. + Build(); +} + +void DataFlowGraph::Build(const framework::ir::Graph &graph) { + // Create nodes + std::unordered_map ir_node_map; + for (auto *ir_node : graph.Nodes()) { + Node *x{nullptr}; + if (ir_node->IsOp()) { + PADDLE_ENFORCE(ir_node->Op()); + VLOG(4) << "get op " << ir_node << " " << ir_node->Name(); + x = nodes.Create(Node::Type::kFunction); + x->attr("ir_node").Pointer() = ir_node; + PADDLE_ENFORCE(ir_node->Op()->Proto()); + x->SetName(ir_node->Op()->Proto()->type()); + x->SetPbMsg(ir_node->Op()->Proto()->SerializeAsString()); + } else if (ir_node->IsVar()) { + // Not create a Node for IR ControlDepVar, considering Inference currently + // just used in single thread scenerio. + VLOG(4) << "get var " << ir_node->Name(); + x = nodes.Create(Node::Type::kValue); + x->attr("ir_node").Pointer() = ir_node; + x->SetName(ir_node->Name()); + // x->SetPbMsg(ir_node->Var()->Proto()->SerializeAsString()); + } else { + PADDLE_THROW("Failed to create an Node from IR, unknown type"); + } + ir_node_map.emplace(ir_node, x); + } + VLOG(4) << "finish creating Nodes"; + + VLOG(4) << "to create edge"; + // Create links + for (auto *ir_node : graph.Nodes()) { + auto it = ir_node_map.find(ir_node); + // Skip ControlDepVar. + if (it == ir_node_map.end()) continue; + auto *node = it->second; + for (auto *x : ir_node->inputs) { + if (!ir_node_map.count(x)) continue; + node->inlinks.push_back(ir_node_map.at(x)); + } + for (auto *x : ir_node->outputs) { + if (!ir_node_map.count(x)) continue; + node->outlinks.push_back(ir_node_map.at(x)); + } + } + + Build(); + PADDLE_ENFORCE(!inputs_.empty(), + "Can't deduce any inputs from the graph, Is the graph empty?"); + + ir_graph = &graph; + VLOG(3) << "finished build from IR"; +} + void DataFlowGraph::Clean() { for (auto &node : nodes.nodes()) { std::unordered_set inlinks_set(node->inlinks.begin(), @@ -61,11 +185,9 @@ void DataFlowGraph::Clean() { std::unordered_set outlinks_set(node->outlinks.begin(), node->outlinks.end()); if (inlinks_set.size() < node->inlinks.size()) { - LOG(INFO) << "Clean: node " << node->repr() << " prune duplicate inputs"; node->inlinks.assign(inlinks_set.begin(), inlinks_set.end()); } if (outlinks_set.size() < node->outlinks.size()) { - LOG(INFO) << "Clean: node " << node->repr() << " prune duplicate inputs"; node->outlinks.assign(outlinks_set.begin(), outlinks_set.end()); } } @@ -112,10 +234,10 @@ GraphTraits::NodesBFSIterator::NodesBFSIterator( const std::vector &source) : queue_(source.begin(), source.end()) {} -// GraphTraits::NodesBFSIterator::NodesBFSIterator( -// GraphTraits::NodesBFSIterator &&other) noexcept -// : queue_(std::move(other.queue_)), -// visited_(std::move(other.visited_)) {} +GraphTraits::NodesBFSIterator::NodesBFSIterator( + GraphTraits::NodesBFSIterator &&other) noexcept + : queue_(std::move(other.queue_)), + visited_(std::move(other.visited_)) {} GraphTraits::NodesBFSIterator::NodesBFSIterator( const GraphTraits::NodesBFSIterator &other) @@ -159,7 +281,7 @@ bool GraphTraits::NodesBFSIterator::operator==( if (queue_.empty()) return other.queue_.empty(); if ((!queue_.empty()) && (!other.queue_.empty())) { return queue_.front() == other.queue_.front() && - visited_.size() == other.visited_.size(); // here need to check the + visited_.size() == other.visited_.size(); // equality of queue and // visited. Just a light but week implementation. } @@ -174,10 +296,10 @@ GraphTraits::NodesDFSIterator::NodesDFSIterator( for (auto *x : source) stack_.push(x); } -// GraphTraits::NodesDFSIterator::NodesDFSIterator( -// GraphTraits::NodesDFSIterator &&other) noexcept -// : stack_(std::move(other.stack_)), -// visited_(std::move(other.visited_)) {} +GraphTraits::NodesDFSIterator::NodesDFSIterator( + GraphTraits::NodesDFSIterator &&other) noexcept + : stack_(std::move(other.stack_)), + visited_(std::move(other.visited_)) {} GraphTraits::NodesDFSIterator::NodesDFSIterator( const GraphTraits::NodesDFSIterator &other) @@ -339,7 +461,7 @@ ExtractInputAndOutputOfSubGraph(std::vector &graph) { // NOLINT void FilterRedundantOutputOfSubGraph(DataFlowGraph *graph) { std::vector op_nodes; - for (auto &node : GraphTraits(graph).nodes_in_TS()) { + for (auto &node : GraphTraits(*graph).nodes_in_TS()) { if (node.type() == Node::Type::kValue || node.deleted()) { continue; } diff --git a/paddle/fluid/inference/analysis/data_flow_graph.h b/paddle/fluid/inference/analysis/data_flow_graph.h index bb3ec6bbc1..437e097acd 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph.h +++ b/paddle/fluid/inference/analysis/data_flow_graph.h @@ -26,6 +26,7 @@ limitations under the License. */ #include #include +#include "paddle/fluid/framework/ir/graph.h" #include "paddle/fluid/inference/analysis/graph_traits.h" #include "paddle/fluid/inference/analysis/node.h" #include "paddle/fluid/platform/enforce.h" @@ -41,19 +42,43 @@ namespace analysis { */ struct DataFlowGraph { NodeMap nodes; - std::vector inputs; - std::vector outputs; + // inputs and outputs are deduced from the graph. + // Used to interact with IR. + const framework::ir::Graph *ir_graph{nullptr}; // Extract inputs and outputs of the graph. void Build(); + void Build(const framework::proto::ProgramDesc &prog); + + // Build a graph from ir::Graph. + void Build(const framework::ir::Graph &graph); + + // Get an attribute. + AnyAttr &Attr(const std::string &key) { return attrs_[key]; } + // Output a DOT graph file for debug. std::string DotString() const; std::string HumanReadableInfo(bool show_values = true, bool show_functions = true) const; + const std::vector &inputs() const { + PADDLE_ENFORCE(!inputs_.empty(), + "No inputs are deduced, need to Build() first."); + return inputs_; + } + const std::vector &outputs() const { + PADDLE_ENFORCE(!outputs_.empty(), + "No outputs are deduced, need to Build() first."); + return outputs_; + } + private: + mutable std::vector inputs_; + mutable std::vector outputs_; + std::unordered_map attrs_; + // Remove duplicate edges and so on. void Clean(); }; @@ -70,7 +95,7 @@ struct GraphTraits { : public std::iterator { NodesBFSIterator() = default; explicit NodesBFSIterator(const std::vector &source); - // NodesBFSIterator(NodesBFSIterator &&other) noexcept; + NodesBFSIterator(NodesBFSIterator &&other) noexcept; // NOTE Heavy to use. NodesBFSIterator(const NodesBFSIterator &other); @@ -93,8 +118,8 @@ struct GraphTraits { struct NodesDFSIterator : public std::iterator { NodesDFSIterator() = default; - explicit NodesDFSIterator(const std::vector &source); - // NodesDFSIterator(NodesDFSIterator &&other) noexcept; + NodesDFSIterator(const std::vector &source); + NodesDFSIterator(NodesDFSIterator &&other) noexcept; NodesDFSIterator(const NodesDFSIterator &other); Node &operator*(); @@ -116,7 +141,7 @@ struct GraphTraits { struct NodesTSIterator : public std::iterator { NodesTSIterator() = default; - explicit NodesTSIterator(const std::vector &source); + NodesTSIterator(const std::vector &source); NodesTSIterator(NodesTSIterator &&other) : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) { other.cursor_ = 0; @@ -138,7 +163,7 @@ struct GraphTraits { size_t cursor_{0}; }; - explicit GraphTraits(DataFlowGraph *graph) : graph_(graph) {} + explicit GraphTraits(const DataFlowGraph &graph) : graph_(graph) {} // default use BFS to visit the nodes. iterator_range nodes() { @@ -156,20 +181,20 @@ struct GraphTraits { private: NodesBFSIterator nodes_bfs_begin() { - return NodesBFSIterator(graph_->inputs); + return NodesBFSIterator(graph_.inputs()); } NodesBFSIterator nodes_bfs_end() { return NodesBFSIterator(); } NodesDFSIterator nodes_dfs_begin() { - return NodesDFSIterator(graph_->inputs); + return NodesDFSIterator(graph_.inputs()); } NodesDFSIterator nodes_dfs_end() { return NodesDFSIterator(); } - NodesTSIterator nodes_ts_begin() { return NodesTSIterator(graph_->inputs); } + NodesTSIterator nodes_ts_begin() { return NodesTSIterator(graph_.inputs()); } NodesTSIterator nodes_ts_end() { return NodesTSIterator(); } private: - DataFlowGraph *graph_; + const DataFlowGraph &graph_; }; // Extract the inputs and outputs of a graph. The inputs and outputs of a diff --git a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc index a881262665..1682011c3d 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_tester.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_tester.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/inference/analysis/data_flow_graph.h" +#include "paddle/fluid/framework/program_desc.h" #include "paddle/fluid/inference/analysis/ut_helper.h" namespace paddle { @@ -24,20 +25,18 @@ TEST(DataFlowGraph, BFS) { auto dfg = ProgramDescToDFG(desc); dfg.Build(); - for (auto *in : dfg.inputs) { + for (auto* in : dfg.inputs()) { LOG(INFO) << "inputs: " << in->name() << " " << static_cast(in->type()); } - for (auto *out : dfg.outputs) { + for (auto* out : dfg.outputs()) { LOG(INFO) << "outputs: " << out->name() << " " << static_cast(out->type()); } - GraphTraits trait(&dfg); - auto nodes = trait.nodes(); size_t count = 0; - for (auto it = nodes.begin(); it != nodes.end(); ++it) { - LOG(INFO) << "visiting " << it->name(); + for (auto& node : GraphTraits(dfg).nodes()) { + LOG(INFO) << "visiting " << node.name(); ++count; } ASSERT_EQ(count, dfg.nodes.size()); @@ -45,13 +44,11 @@ TEST(DataFlowGraph, BFS) { TEST(DataFlowGraph, DFS) { auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); - auto dfg = ProgramDescToDFG(desc); - dfg.Build(); - GraphTraits trait(&dfg); - auto nodes = trait.nodes_in_DFS(); + DataFlowGraph dfg; + dfg.Build(desc); size_t count = 0; - for (auto it = nodes.begin(); it != nodes.end(); ++it) { - LOG(INFO) << "visiting " << it->name(); + for (auto& node : GraphTraits(dfg).nodes_in_DFS()) { + LOG(INFO) << "visiting " << node.name(); ++count; } ASSERT_EQ(count, dfg.nodes.size()); @@ -74,21 +71,17 @@ TEST(DataFlowGraph, TS) { DataFlowGraph graph; for (int i = 0; i < 8; i++) { - auto *node = graph.nodes.Create(Node::Type::kValue); + auto* node = graph.nodes.Create(Node::Type::kValue); node->SetName("node-" + std::to_string(i)); } auto add_link = [&](int i, int j) { - Node *source = graph.nodes.GetMutable(i); - Node *target = graph.nodes.GetMutable(j); + Node* source = graph.nodes.GetMutable(i); + Node* target = graph.nodes.GetMutable(j); target->inlinks.push_back(source); source->outlinks.push_back(target); }; - graph.inputs.push_back(graph.nodes.GetMutable(0)); - graph.inputs.push_back(graph.nodes.GetMutable(1)); - graph.inputs.push_back(graph.nodes.GetMutable(2)); - add_link(0, 4); add_link(0, 5); add_link(1, 6); @@ -97,8 +90,9 @@ TEST(DataFlowGraph, TS) { add_link(4, 7); add_link(4, 3); add_link(7, 3); + graph.Build(); - auto its = GraphTraits(&graph).nodes_in_TS(); + auto its = GraphTraits(graph).nodes_in_TS(); std::vector sorted_ids; for (auto it = its.begin(); it != its.end(); ++it) { LOG(INFO) << it->name(); @@ -122,6 +116,50 @@ TEST(DataFlowGraph, TS) { assert_positive_sequence_pair(4, 7); } +TEST(DataFlowGraph, Build_ProgramDesc) { + auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__"); + DataFlowGraph graph; + graph.Build(desc); + ASSERT_EQ(graph.nodes.size(), 38UL); +} + +void SetOp(framework::ProgramDesc* prog, const std::string& type, + const std::vector& inputs, + const std::vector& outputs) { + auto* op = prog->MutableBlock(0)->AppendOp(); + op->SetType(type); + op->SetInput("Xs", inputs); + op->SetOutput("Xs", outputs); +} + +TEST(DataFlowGraph, Build_IR_Graph) { + framework::ProgramDesc prog; + for (auto& v : std::vector({"a", "b", "c", "d", "e", "f"})) { + auto* var = prog.MutableBlock(0)->Var(v); + var->SetType(framework::proto::VarType::SELECTED_ROWS); + if (v == "c") { + var->SetPersistable(true); + } + } + + SetOp(&prog, "OP0", std::vector({"a"}), + std::vector({"b"})); + SetOp(&prog, "OP1", std::vector({"a"}), + std::vector({"c"})); + SetOp(&prog, "mul", std::vector({"b", "c"}), + std::vector({"d"})); + SetOp(&prog, "elementwise_add", std::vector({"d", "e"}), + std::vector({"f"})); + + DataFlowGraph graph; + + framework::ir::Graph ir_graph(prog); + + graph.Build(ir_graph); + + ASSERT_EQ(graph.nodes.size(), ir_graph.Nodes().size()); +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc index 18c32fa091..8c7dd146e4 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc @@ -23,9 +23,6 @@ namespace paddle { namespace inference { -DEFINE_int32(tensorrt_max_batchsize, 3, "TensorRT maximum batch size"); -DEFINE_int32(tensorrt_workspace_size, 2048, "TensorRT workspace size"); - namespace analysis { using framework::proto::ProgramDesc; @@ -52,19 +49,15 @@ bool DataFlowGraphToFluidPass::Initialize(Argument *argument) { bool DataFlowGraphToFluidPass::Finalize() { return true; } void DataFlowGraphToFluidPass::Run(DataFlowGraph *graph) { - FilterRedundantOutputOfSubGraph(graph); - LOG(INFO) << "graph.inputs " << graph->inputs.size(); - for (auto &node : GraphTraits(graph).nodes_in_TS()) { + // FilterRedundantOutputOfSubGraph(graph); + for (auto &node : GraphTraits(*graph).nodes_in_TS()) { if (node.deleted()) continue; switch (node.type()) { case Node::Type::kFunction: { - LOG(INFO) << "add function " << node.repr(); AddFluidOp(&node); } break; case Node::Type::kFunctionBlock: { - LOG(INFO) << "add engine op " << node.repr() << " , " - << static_cast(&node)->subgraph.size(); AddEngineOp(&node); } break; default: @@ -76,15 +69,27 @@ void DataFlowGraphToFluidPass::Run(DataFlowGraph *graph) { } void DataFlowGraphToFluidPass::AddFluidOp(Node *node) { - auto *ori_op = static_cast(node->pb_desc()); + PADDLE_ENFORCE(node); + PADDLE_ENFORCE(node->IsFunction()); + PADDLE_ENFORCE(node->pb_desc() || !node->pb_msg().empty(), + "node has invalid protobuf repr."); + // currently only the main block is analyzed. + PADDLE_ENFORCE(desc_); auto *main_block = desc_->mutable_blocks(framework::kRootBlockIndex); auto *op = main_block->add_ops(); - *op = *ori_op; // copy the attributes, by default, these will not be changed - // by analysis phrase. - // The inputs and outputs of the existing ops are not changed by tensorrt - // subgraph pass. - // NOTE It might be changed by other passes in the long run. + + if (node->pb_desc()) { + auto *ori_op = static_cast(node->pb_desc()); + *op = + *ori_op; // copy the attributes, by default, these will not be changed + // by analysis phrase. + // The inputs and outputs of the existing ops are not changed by tensorrt + // subgraph pass. + // NOTE It might be changed by other passes in the long run. + } else { + op->ParseFromString(node->pb_msg()); + } } void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph, @@ -191,8 +196,6 @@ void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph, // Set attrs SetAttr(desc.Proto(), "subgraph", block->SerializeAsString()); SetAttr(desc.Proto(), "engine_uniq_key", "trt-" + std::to_string(counter++)); - SetAttr(desc.Proto(), "max_batch", FLAGS_tensorrt_max_batchsize); - SetAttr(desc.Proto(), "max_workspace", FLAGS_tensorrt_workspace_size); SetAttr(desc.Proto(), "parameters", ExtractParameters(graph.nodes.nodes())); SetAttr(desc.Proto(), "output_name_mapping", output_mapping); node->SetPbMsg(desc.Proto()->SerializeAsString()); @@ -221,10 +224,9 @@ void DataFlowGraphToFluidPass::AddEngineOp(Node *node) { framework::BlockDesc block_desc(nullptr, &proto); block_desc.Proto()->set_parent_idx(-1); block_desc.Proto()->set_idx(0); - LOG(INFO) << "origin variable size: " - << argument_->origin_program_desc->blocks(0).vars().size(); - LOG(INFO) << "transformed variable size: " - << block_desc.Proto()->vars().size(); + VLOG(4) << "origin variable size: " + << argument_->origin_program_desc->blocks(0).vars().size(); + VLOG(4) << "transformed variable size: " << block_desc.Proto()->vars().size(); // copy ops. for (auto *node : block_node->subgraph) { @@ -258,7 +260,7 @@ class DFG_DebuggerPass : public DFG_GraphvizDrawPass { Pass *DataFlowGraphToFluidPass::CreateGraphvizDebugerPass() const { return new DFG_DebuggerPass(DFG_GraphvizDrawPass::Config( - FLAGS_inference_analysis_graphviz_log_root, + FLAGS_IA_graphviz_log_root, "data_flow_graph_to_fluid_graphviz_debugger")); } diff --git a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h index 59c47365aa..0c9a8a0b7c 100644 --- a/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h +++ b/paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h @@ -27,9 +27,6 @@ namespace paddle { namespace inference { -DECLARE_int32(tensorrt_max_batchsize); -DECLARE_int32(tensorrt_workspace_size); - namespace analysis { class DataFlowGraphToFluidPass final : public DataFlowGraphPass { public: diff --git a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc index c05b0e5d46..648b8f7d6a 100644 --- a/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc +++ b/paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc @@ -29,7 +29,7 @@ void DFG_GraphvizDrawPass::Run(DataFlowGraph *graph) { auto png_path = dot_path.substr(0, dot_path.size() - 4) + ".png"; std::string message; - LOG(INFO) << "draw to " << png_path; + VLOG(3) << "draw to " << png_path; ExecShellCommand("dot -Tpng " + dot_path + " -o " + png_path, &message); } diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc index 511631d3e0..51bd0ac42d 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc @@ -52,72 +52,7 @@ bool FluidToDataFlowGraphPass::Finalize() { return true; } void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) { PADDLE_ENFORCE(graph); PADDLE_ENFORCE(desc_); - // insert vars - // The `var2id` keeps a map from a variable's name to its Node-id, the Node-id - // will keep updating to its latest alias during the graph-building. - std::unordered_map var2id; - auto &main_block = desc_->blocks(framework::kRootBlockIndex); - for (int i = 0; i < main_block.vars_size(); i++) { - const auto &var = main_block.vars(i); - auto *v = graph->nodes.Create(Node::Type::kValue); - v->SetName(var.name()); - v->SetPbDesc(const_cast(static_cast(&var))); - v->SetPbMsg(var.SerializeAsString()); - var2id[var.name()] = v->id(); - } - - // The variables in a SSA can only write once, so if a variable is written - // multiple times(quite common in our ProgramDesc design), multiple alias - // Nodes of this variable will be created, and each will just write once. - - // An set that keep all the names of the variables(the original, not alias) - // that have been written(as outputs). Once an Op's output variable hit the - // set, it should create a new alias and update the global alias for this - // variable. And that make a Data Flow Graph a SSA. - std::unordered_set unique_written_vars; - for (int i = 0; i < main_block.ops_size(); i++) { - const auto &op = main_block.ops(i); - auto *o = graph->nodes.Create(Node::Type::kFunction); - o->SetName(op.type()); - static_cast(o)->SetFuncType(op.type()); - // Link to the original protobuf message's memory, make it easier to - // generate from a data flow graph to fluid ProgramDesc. - o->SetPbDesc(const_cast(static_cast(&op))); - o->SetPbMsg(op.SerializeAsString()); - - // set inputs and outputs - for (int j = 0; j < op.inputs_size(); j++) { - auto &in_var = op.inputs(j); - for (int k = 0; k < in_var.arguments_size(); k++) { - auto *in = graph->nodes.GetMutable(var2id.at(in_var.arguments(k))); - in->outlinks.push_back(o); - o->inlinks.push_back(in); - } - } - for (int j = 0; j < op.outputs_size(); j++) { - auto &out_var = op.outputs(j); - for (int k = 0; k < out_var.arguments_size(); k++) { - auto *out = graph->nodes.GetMutable(var2id[out_var.arguments(k)]); - if (unique_written_vars.count(out)) { - // Loop found, for example, a = op(a), use SSA, change to a1 = op(a). - auto *out_alias = graph->nodes.Create(Node::Type::kValue); - out_alias->SetName(out->name()); - out_alias->SetPbDesc(out->pb_desc()); - out_alias->SetPbMsg(out->pb_msg()); - var2id[out_alias->name()] = - out_alias->id(); // update variable's alias Node - LOG(INFO) << "loop found in graph, create SSA alias node [" - << out_alias->repr() << "] for [" << out->repr() << "]"; - out = out_alias; - } - out->inlinks.push_back(o); - o->outlinks.push_back(out); - unique_written_vars.insert(out); - } - } - } - // Analysis and extract the inputs and outputs of this graph. - graph->Build(); + graph->Build(*desc_); } namespace { @@ -133,7 +68,7 @@ class DFG_DebuggerPass : public DFG_GraphvizDrawPass { Pass *FluidToDataFlowGraphPass::CreateGraphvizDebugerPass() const { return new DFG_DebuggerPass(DFG_GraphvizDrawPass::Config( - FLAGS_inference_analysis_graphviz_log_root, "fluid-to-dfg-debuger")); + FLAGS_IA_graphviz_log_root, "fluid-to-dfg-debuger")); } } // namespace analysis diff --git a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc index d218dcd050..267a0a84eb 100644 --- a/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc +++ b/paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc @@ -30,7 +30,7 @@ TEST(FluidToDataFlowGraphPass, Test) { ASSERT_EQ(argument.main_dfg->nodes.size(), 38UL); pass.Finalize(); ASSERT_FALSE(argument.main_dfg->DotString().empty()); - EXPECT_FALSE(argument.main_dfg->inputs.empty()); + EXPECT_FALSE(argument.main_dfg->inputs().empty()); } } // namespace analysis diff --git a/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc b/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc new file mode 100644 index 0000000000..073f497528 --- /dev/null +++ b/paddle/fluid/inference/analysis/fluid_to_ir_pass.cc @@ -0,0 +1,15 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/fluid_to_ir_pass.h" diff --git a/paddle/fluid/inference/analysis/fluid_to_ir_pass.h b/paddle/fluid/inference/analysis/fluid_to_ir_pass.h new file mode 100644 index 0000000000..fa3f8d313b --- /dev/null +++ b/paddle/fluid/inference/analysis/fluid_to_ir_pass.h @@ -0,0 +1,82 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include "paddle/fluid/inference/analysis/ir_pass_manager.h" +#include "paddle/fluid/inference/analysis/pass.h" + +namespace paddle { +namespace inference { +namespace analysis { + +class FluidToIrPass final : public DataFlowGraphPass { + public: + FluidToIrPass() = default; + + bool Initialize(Argument *argument) override { + ANALYSIS_ARGUMENT_CHECK_FIELD(argument); + if (argument->origin_program_desc) { + LOG(WARNING) << "argument's origin_program_desc is already set, might " + "duplicate called"; + } + // set fluid model program path + if (!argument->fluid_model_program_path) { + ANALYSIS_ARGUMENT_CHECK_FIELD(argument->fluid_model_dir); + argument->fluid_model_program_path.reset( + new std::string(*argument->fluid_model_dir + "/__model__")); + } + ANALYSIS_ARGUMENT_CHECK_FIELD(argument->fluid_model_program_path); + // Load program. + auto program = LoadProgramDesc(*argument->fluid_model_program_path); + argument->origin_program_desc.reset( + new framework::proto::ProgramDesc(program)); + // Create main data flow graph. + if (!argument->main_dfg) { + argument->main_dfg.reset(new DataFlowGraph); + } + // Persist the ProgramDesc in graph's attribute. The IR graph just keep the + // address, will segfault if the original ProgramDesc destroys. + auto &ir_program_p = argument->main_dfg->Attr("ir_program_desc").Pointer(); + ir_program_p = new framework::ProgramDesc(program); + + argument_ = argument; + return true; + } + + bool Finalize() override { return true; } + + void Run(DataFlowGraph *graph) override { + // Call all the IR Passes + IRPassManager ir_passes(*static_cast( + argument_->main_dfg->Attr("ir_program_desc").Pointer())); + ir_passes.Apply(std::vector( + {// Manual update the passes here. + "graph_viz_pass", "infer_clean_graph_pass", "graph_viz_pass", + "fc_fuse_pass", "graph_viz_pass"})); + + PADDLE_ENFORCE(argument_->main_dfg.get()); + argument_->main_dfg->Build(ir_passes.graph()); + // PADDLE_ENFORCE(argument_->main_dfg->IsFullyConnected()); + } + + std::string repr() const override { return "fluid-to-ir-pass"; } + + private: + Argument *argument_{nullptr}; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/fluid_to_ir_pass_tester.cc b/paddle/fluid/inference/analysis/fluid_to_ir_pass_tester.cc new file mode 100644 index 0000000000..af934f261b --- /dev/null +++ b/paddle/fluid/inference/analysis/fluid_to_ir_pass_tester.cc @@ -0,0 +1,37 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/fluid_to_ir_pass.h" + +#include +#include "paddle/fluid/inference/analysis/ut_helper.h" + +namespace paddle { +namespace inference { +namespace analysis { + +TEST(FluidToIrPass, Test) { + FluidToIrPass pass; + Argument argument(FLAGS_inference_model_dir); + pass.Initialize(&argument); + pass.Run(argument.main_dfg.get()); +} + +} // namespace analysis +} // namespace inference +} // namespace paddle + +USE_PASS(fc_fuse_pass); +USE_PASS(graph_viz_pass); +USE_PASS(infer_clean_graph_pass); diff --git a/paddle/fluid/inference/analysis/helper.h b/paddle/fluid/inference/analysis/helper.h index a0f912b251..5151e2b69a 100644 --- a/paddle/fluid/inference/analysis/helper.h +++ b/paddle/fluid/inference/analysis/helper.h @@ -14,6 +14,7 @@ limitations under the License. */ #pragma once +#include #include #include #include @@ -151,6 +152,23 @@ static framework::proto::ProgramDesc LoadProgramDesc( return program_desc; } +static bool FileExists(const std::string &filepath) { + std::ifstream file(filepath); + bool exists = file.is_open(); + file.close(); + return exists; +} + +static bool PathExists(const std::string &path) { + struct stat statbuf; + if (stat(path.c_str(), &statbuf) != -1) { + if (S_ISDIR(statbuf.st_mode)) { + return true; + } + } + return false; +} + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.cc b/paddle/fluid/inference/analysis/ir_pass_manager.cc new file mode 100644 index 0000000000..d849b637bc --- /dev/null +++ b/paddle/fluid/inference/analysis/ir_pass_manager.cc @@ -0,0 +1,45 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#include "paddle/fluid/inference/analysis/ir_pass_manager.h" +#include + +namespace paddle { +namespace inference { +namespace analysis { + +IRPassManager::IRPassManager(const ProgramDesc& program) { + graph_.reset(new framework::ir::Graph(program)); +} + +void IRPassManager::Apply(const std::vector& passes) { + graph_->Set("graph_viz_path", new std::string("./1.dot")); + // Apply all the passes + std::string pre_pass; + for (const std::string& pass_name : passes) { + LOG(WARNING) << "Running IR pass [" << pass_name << "]"; + auto pass = framework::ir::PassRegistry::Instance().Get(pass_name); + if (pass_name == "graph_viz_pass") { + std::string dot_file_path = + "ir_" + (pre_pass.empty() ? "origin" : pre_pass) + ".dot"; + pass->Set("graph_viz_path", new std::string(std::move(dot_file_path))); + } + graph_ = pass->Apply(std::move(graph_)); + pre_pass = pass_name; + } +} + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/ir_pass_manager.h b/paddle/fluid/inference/analysis/ir_pass_manager.h new file mode 100644 index 0000000000..3338e37ecf --- /dev/null +++ b/paddle/fluid/inference/analysis/ir_pass_manager.h @@ -0,0 +1,46 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +/* + * This file defines IRPassManager, it helps control the passes in IR. Inference + * phrase will load the model program and parameters from disk, that is quite + * different from the training phase. + * This manager will control the Passes and make the passes in IR work smoothly + * for inference. + */ + +#include "paddle/fluid/framework/ir/graph.h" +#include "paddle/fluid/framework/ir/pass.h" +#include "paddle/fluid/framework/program_desc.h" + +namespace paddle { +namespace inference { +namespace analysis { +using framework::ProgramDesc; + +class IRPassManager final { + public: + IRPassManager(const ProgramDesc& program); + + void Apply(const std::vector& passes); + + framework::ir::Graph& graph() const { return *graph_; } + + private: + std::unique_ptr graph_; +}; + +} // namespace analysis +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/analysis/model_store_pass.cc b/paddle/fluid/inference/analysis/model_store_pass.cc index 1c42917642..c313db0887 100644 --- a/paddle/fluid/inference/analysis/model_store_pass.cc +++ b/paddle/fluid/inference/analysis/model_store_pass.cc @@ -35,19 +35,21 @@ void ModelStorePass::Run(DataFlowGraph *x) { std::stringstream ss; // NOTE these commands only works on linux. ss << "mkdir -p " << *argument_->model_output_store_path; - LOG(INFO) << "run command: " << ss.str(); + VLOG(3) << "run command: " << ss.str(); PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0); ss.str(""); ss << "cp " << *argument_->fluid_model_dir << "/*" << " " << *argument_->model_output_store_path; - LOG(INFO) << "run command: " << ss.str(); + VLOG(3) << "run command: " << ss.str(); PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0); // Store program PADDLE_ENFORCE_NOT_NULL(argument_->transformed_program_desc, "program desc is not transformed, should call " "DataFlowGraphToFluidPass first."); + VLOG(3) << "store analyzed program to " + << *argument_->model_output_store_path; const std::string program_output_path = *argument_->model_output_store_path + "/__model__"; std::ofstream file(program_output_path, std::ios::binary); @@ -58,6 +60,8 @@ void ModelStorePass::Run(DataFlowGraph *x) { file.write(serialized_message.c_str(), serialized_message.size()); } +bool ModelStorePass::Finalize() { return true; } + } // namespace analysis } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/analysis/model_store_pass.h b/paddle/fluid/inference/analysis/model_store_pass.h index fac7083925..3a2869e30b 100644 --- a/paddle/fluid/inference/analysis/model_store_pass.h +++ b/paddle/fluid/inference/analysis/model_store_pass.h @@ -44,6 +44,8 @@ class ModelStorePass : public DataFlowGraphPass { model in the disk, and that model can be reloaded for prediction again.)DD"; } + bool Finalize() override; + private: Argument* argument_{nullptr}; }; diff --git a/paddle/fluid/inference/analysis/model_store_pass_tester.cc b/paddle/fluid/inference/analysis/model_store_pass_tester.cc index 5f3526dd50..d6493fc25e 100644 --- a/paddle/fluid/inference/analysis/model_store_pass_tester.cc +++ b/paddle/fluid/inference/analysis/model_store_pass_tester.cc @@ -30,7 +30,7 @@ TEST(DFG_StorePass, test) { argument.model_output_store_path.reset( new std::string("./_dfg_store_pass_tmp")); // disable storage in alalyzer - FLAGS_inference_analysis_output_storage_path = ""; + FLAGS_IA_output_storage_path = ""; analyzer.Run(&argument); ModelStorePass pass; diff --git a/paddle/fluid/inference/analysis/node.cc b/paddle/fluid/inference/analysis/node.cc index f2e918f3ff..3339b5044d 100644 --- a/paddle/fluid/inference/analysis/node.cc +++ b/paddle/fluid/inference/analysis/node.cc @@ -20,17 +20,6 @@ namespace paddle { namespace inference { namespace analysis { -template <> -std::string &NodeAttr::As() { - if (data_.empty()) { - type_index_ = std::type_index(typeid(std::string)); - } - PADDLE_ENFORCE_EQ(type_index_, std::type_index(typeid(std::string))); - return data_; -} - -std::string &NodeAttr::String() { return As(); } - std::vector Value::dot_attrs() const { return std::vector({Dot::Attr("style", "filled,rounded"), Dot::Attr("shape", "box"), diff --git a/paddle/fluid/inference/analysis/node.h b/paddle/fluid/inference/analysis/node.h index 47e524bc5c..af34156bc2 100644 --- a/paddle/fluid/inference/analysis/node.h +++ b/paddle/fluid/inference/analysis/node.h @@ -29,6 +29,7 @@ limitations under the License. */ #include "paddle/fluid/inference/analysis/device.h" #include "paddle/fluid/inference/analysis/dot.h" #include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/platform/variant.h" namespace paddle { namespace inference { @@ -37,41 +38,36 @@ namespace analysis { class NodeMap; // A helper class to maintain the status from Pass. -struct NodeAttr { +struct AnyAttr { + using any_t = + boost::variant; // NOTE T should be a primary type or a struct combined by several primary // types. // NOTE the STL containers should not use here. // Some usages // Attr attr; // attr.Bool() = true; - bool &Bool() { return As(); } float &Float() { return As(); } int32_t &Int32() { return As(); } int64_t &Int64() { return As(); } void *&Pointer() { return As(); } - std::string &String(); + std::string &String() { return As(); } - private: template T &As() { - // init storage in the first usage. - if (data_.empty()) { - VLOG(4) << "resize data to " << sizeof(T); - type_index_ = std::type_index(typeid(T)); - data_.resize(sizeof(T)); + if (type_index_ == typeid(AnyAttr)) { + type_index_ = typeid(T); + any_data_ = T(); + } else { + PADDLE_ENFORCE(type_index_ == typeid(T), "fetch error type"); } - PADDLE_ENFORCE(framework::IsType(type_index_), - "type not matched, origin is %s, want %s", - DataTypeNamer::Global().repr(type_index_), - DataTypeNamer::Global().repr()); - PADDLE_ENFORCE_EQ(data_.size(), sizeof(T), "Node attr type recast error"); - return *reinterpret_cast(&data_[0]); + return boost::get(any_data_); } private: - std::string data_; - std::type_index type_index_{typeid(NodeAttr)}; + any_t any_data_; + std::type_index type_index_{typeid(AnyAttr)}; }; /* @@ -108,7 +104,7 @@ class Node { // Get an additional attribute and convert it to T data type. NOTE this will // silently create a new attribute if not exists. - NodeAttr &attr(const std::string &name) const { return attrs_[name]; } + AnyAttr &attr(const std::string &name) const { return attrs_[name]; } int id() const { return id_; } @@ -153,7 +149,7 @@ class Node { Type type_{Type::kNone}; // Mark this node is deleted by some pass. bool deleted_{false}; - mutable std::unordered_map attrs_; + mutable std::unordered_map attrs_; }; class Function; diff --git a/paddle/fluid/inference/analysis/node_tester.cc b/paddle/fluid/inference/analysis/node_tester.cc index ea832a3a7e..9207c15373 100644 --- a/paddle/fluid/inference/analysis/node_tester.cc +++ b/paddle/fluid/inference/analysis/node_tester.cc @@ -20,6 +20,24 @@ namespace paddle { namespace inference { namespace analysis { +TEST(NodeAttr, bool) { + AnyAttr x; + x.Bool() = true; + ASSERT_EQ(x.Bool(), true); +} + +TEST(NodeAttr, int32) { + AnyAttr x; + x.Int32() = 32; + ASSERT_EQ(x.Int32(), 32); +} + +TEST(NodeAttr, string) { + AnyAttr x; + x.String() = "Hello"; + ASSERT_EQ(x.String(), "Hello"); +} + TEST(Node, Attr) { // Node is an abstract class, use Value instead for they share the same Attr // logic. @@ -27,6 +45,9 @@ TEST(Node, Attr) { auto* node = nodes.Create(Node::Type::kValue); node->attr("v0").Int32() = 2008; ASSERT_EQ(node->attr("v0").Int32(), 2008); + + node->attr("str").String() = "hello world"; + ASSERT_EQ(node->attr("str").String(), "hello world"); } } // namespace analysis diff --git a/paddle/fluid/inference/analysis/pass.h b/paddle/fluid/inference/analysis/pass.h index 6806f9ff7d..7719c6f5ff 100644 --- a/paddle/fluid/inference/analysis/pass.h +++ b/paddle/fluid/inference/analysis/pass.h @@ -63,7 +63,7 @@ class Pass { // Human-readable short representation. virtual std::string repr() const = 0; // Human-readable long description. - virtual std::string description() const = 0; + virtual std::string description() const { return "No DOC"; } }; // NodePass process on any Node types. diff --git a/paddle/fluid/inference/analysis/pass_manager.cc b/paddle/fluid/inference/analysis/pass_manager.cc index b428bb22b1..cfdca33882 100644 --- a/paddle/fluid/inference/analysis/pass_manager.cc +++ b/paddle/fluid/inference/analysis/pass_manager.cc @@ -22,7 +22,7 @@ namespace analysis { bool PassManager::Initialize(Argument* argument) { argument_ = argument; for (auto& pass : data_) { - LOG(INFO) << "Initializing pass " << pass->repr(); + LOG(WARNING) << "Initializing pass [" << pass->repr() << "]"; if (!pass->Initialize(argument)) { LOG(ERROR) << "Failed to initialize pass [" << pass->repr() << "]"; return false; @@ -33,8 +33,9 @@ bool PassManager::Initialize(Argument* argument) { void DfgPassManager::RunAll() { PADDLE_ENFORCE(argument_); + LOG(INFO) << "Total " << data_.size() << " passes"; for (auto& pass : data_) { - VLOG(4) << "Running pass [" << pass->repr() << "]"; + LOG(WARNING) << "Running pass [" << pass->repr() << "]"; pass->Run(argument_->main_dfg.get()); } } @@ -42,8 +43,7 @@ void DfgPassManager::RunAll() { void NodePassManager::RunAll() { PADDLE_ENFORCE(argument_); PADDLE_ENFORCE(argument_->main_dfg.get()); - auto trait = - GraphTraits(argument_->main_dfg.get()).nodes_in_DFS(); + auto trait = GraphTraits(*argument_->main_dfg).nodes_in_DFS(); for (auto& node : trait) { for (auto& pass : data_) { pass->Run(&node); diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc index 80809d4c43..670a8de667 100644 --- a/paddle/fluid/inference/analysis/subgraph_splitter.cc +++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc @@ -34,7 +34,7 @@ inline void MarkOutLinksInSubGraph(const Function *func) { } void SubGraphSplitter::MarkNodesInsideSubGraph() { - for (auto &node : GraphTraits(graph_).nodes()) { + for (auto &node : GraphTraits(*graph_).nodes()) { if (node_inside_subgraph_teller_(&node)) { node.attr(kMarkerAttrName).Bool() = true; if (node.type() == Node::Type::kFunction) { @@ -76,7 +76,7 @@ void UnionFindCombine(const node_map_t &node_map, size_t a, size_t b) { std::vector> SubGraphSplitter::ExtractSubGraphs() { std::vector marked_nodes; - for (auto &node : GraphTraits(graph_).nodes_in_TS()) { + for (auto &node : GraphTraits(*graph_).nodes_in_TS()) { if (node.attr(kMarkerAttrName).Bool()) { marked_nodes.push_back(&node); } @@ -153,6 +153,7 @@ void SubGraphFuse::ReplaceNodesWithSubGraphs() { inlink_or_outlink_cleaner(o->inlinks); } } + FilterRedundantOutputOfSubGraph(graph_); } } // namespace analysis diff --git a/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc index f736e385c1..9f51fafe0b 100644 --- a/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc +++ b/paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.cc @@ -69,8 +69,8 @@ class DfgDebuggerPass : public DFG_GraphvizDrawPass { }; Pass *TensorRTSubgraphNodeMarkPass::CreateGraphvizDebugerPass() const { - DFG_GraphvizDrawPass::Config config( - FLAGS_inference_analysis_graphviz_log_root, "tensorrt_marked_node"); + DFG_GraphvizDrawPass::Config config(FLAGS_IA_graphviz_log_root, + "tensorrt_marked_node"); return new DfgDebuggerPass(config); } bool TensorRTSubgraphNodeMarkPass::Finalize() { return true; } diff --git a/paddle/fluid/inference/api/CMakeLists.txt b/paddle/fluid/inference/api/CMakeLists.txt index 83867e0a2c..0ca1af455c 100644 --- a/paddle/fluid/inference/api/CMakeLists.txt +++ b/paddle/fluid/inference/api/CMakeLists.txt @@ -18,7 +18,10 @@ if(APPLE) endif(APPLE) -set(inference_deps paddle_inference_api paddle_fluid_api) +set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager + graph_viz_pass fc_fuse_pass + infer_clean_graph_pass + ) if(WITH_GPU AND TENSORRT_FOUND) set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine) @@ -60,19 +63,22 @@ cc_library(paddle_inference_tensorrt_subgraph_engine inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec) endif() -if (WITH_ANAKIN) # only needed in CI +if (WITH_ANAKIN AND WITH_GPU) # only needed in CI # compile the libinference_anakin_api.a and anakin.so. - nv_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber) - #nv_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin) + cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber mklml) + cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber) function(anakin_target target_name) target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) endfunction() anakin_target(inference_anakin_api) - #anakin_target(inference_anakin_api_shared) + anakin_target(inference_anakin_api_shared) if (WITH_TESTING) - cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc + cc_test(api_anakin_engine_tester SRCS api_anakin_engine_tester.cc ARGS --model=${ANAKIN_SOURCE_DIR}/mobilenet_v2.anakin.bin - DEPS inference_anakin_api dynload_cuda SERIAL) - target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS}) + DEPS inference_anakin_api_shared dynload_cuda SERIAL) + cc_test(api_anakin_engine_rnn_tester SRCS api_anakin_engine_rnn_tester.cc + ARGS --model=${ANAKIN_SOURCE_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin + --datapath=${ANAKIN_SOURCE_DIR}/anakin_test%2Fditu_rnn_data.txt + DEPS inference_anakin_api_shared dynload_cuda SERIAL) endif(WITH_TESTING) endif() diff --git a/paddle/fluid/inference/api/api.cc b/paddle/fluid/inference/api/api.cc index 63c3f0d7b3..5f1e1b548c 100644 --- a/paddle/fluid/inference/api/api.cc +++ b/paddle/fluid/inference/api/api.cc @@ -1,11 +1,8 @@ /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. - Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at - http://www.apache.org/licenses/LICENSE-2.0 - Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. diff --git a/paddle/fluid/inference/api/api_anakin_engine.cc b/paddle/fluid/inference/api/api_anakin_engine.cc index 6b374ceefb..ea66aa89b8 100644 --- a/paddle/fluid/inference/api/api_anakin_engine.cc +++ b/paddle/fluid/inference/api/api_anakin_engine.cc @@ -13,9 +13,22 @@ // limitations under the License. #include "paddle/fluid/inference/api/api_anakin_engine.h" + +#ifdef PADDLE_WITH_CUDA #include +#endif + +#include +#include +#include +#include +#include #include +#include "framework/core/net/net.h" +#include "framework/operators/ops.h" +#include "saber/funcs/timer.h" + namespace paddle { template @@ -23,16 +36,24 @@ PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor( const AnakinConfig &config) { CHECK(Init(config)); } - +template <> +PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor( + const AnakinConfig &config) { + omp_set_dynamic(0); + omp_set_num_threads(1); + mkl_set_num_threads(1); + CHECK(Init(config)); +} template bool PaddleInferenceAnakinPredictor::Init(const AnakinConfig &config) { if (!(graph_.load(config.model_file))) { - LOG(FATAL) << "fail to load graph from " << config.model_file; + VLOG(3) << "fail to load graph from " << config.model_file; return false; } auto inputs = graph_.get_ins(); for (auto &input_str : inputs) { graph_.ResetBatchSize(input_str, config.max_batch_size); + max_batch_size_ = config.max_batch_size; } // optimization for graph if (!(graph_.Optimize())) { @@ -52,15 +73,15 @@ bool PaddleInferenceAnakinPredictor::Run( std::vector *output_data, int batch_size) { for (const auto &input : inputs) { if (input.dtype != PaddleDType::FLOAT32) { - LOG(ERROR) << "Only support float type inputs. " << input.name - << "'s type is not float"; + VLOG(3) << "Only support float type inputs. " << input.name + << "'s type is not float"; return false; } auto d_tensor_in_p = executor_p_->get_in(input.name); - auto net_shape = d_tensor_in_p->valid_shape(); + auto net_shape = d_tensor_in_p->shape(); if (net_shape.size() != input.shape.size()) { - LOG(ERROR) << " input " << input.name - << "'s shape size should be equal to that of net"; + VLOG(3) << " input " << input.name + << "'s shape size should be equal to that of net"; return false; } int sum = 1; @@ -79,21 +100,45 @@ bool PaddleInferenceAnakinPredictor::Run( } d_tensor_in_p->reshape(tmp_shape); + if (input.lod.size() > 0) { + if (input.lod.size() > 1) { + VLOG(3) << " input lod first dim should <=1, but you set " + << input.lod.size(); + return false; + } + std::vector offset(input.lod[0].begin(), input.lod[0].end()); + d_tensor_in_p->set_seq_offset(offset); + VLOG(3) << "offset.size(): " << offset.size(); + for (int i = 0; i < offset.size(); i++) { + VLOG(3) << offset[i]; + } + } + float *d_data_p = d_tensor_in_p->mutable_data(); - if (cudaMemcpy(d_data_p, static_cast(input.data.data()), - d_tensor_in_p->valid_size() * sizeof(float), - cudaMemcpyHostToDevice) != 0) { - LOG(ERROR) << "copy data from CPU to GPU error"; - return false; + +#ifdef PADDLE_WITH_CUDA + if (std::is_same::value) { + if (cudaMemcpy(d_data_p, static_cast(input.data.data()), + d_tensor_in_p->valid_size() * sizeof(float), + cudaMemcpyHostToDevice) != 0) { + VLOG(3) << "copy data from CPU to GPU error"; + return false; + } + } +#endif + if (std::is_same::value) { + memcpy(d_data_p, static_cast(input.data.data()), + d_tensor_in_p->valid_size() * sizeof(float)); } - cudaStreamSynchronize(NULL); } +#ifdef PADDLE_WITH_CUDA cudaDeviceSynchronize(); executor_p_->prediction(); cudaDeviceSynchronize(); +#endif if (output_data->empty()) { - LOG(ERROR) << "At least one output should be set with tensors' names."; + VLOG(3) << "At least one output should be set with tensors' names."; return false; } for (auto &output : *output_data) { @@ -102,14 +147,22 @@ bool PaddleInferenceAnakinPredictor::Run( if (output.data.length() < tensor->valid_size() * sizeof(float)) { output.data.Resize(tensor->valid_size() * sizeof(float)); } - // Copy data from GPU -> CPU - if (cudaMemcpy(output.data.data(), tensor->mutable_data(), - tensor->valid_size() * sizeof(float), - cudaMemcpyDeviceToHost) != 0) { - LOG(ERROR) << "copy data from GPU to CPU error"; - return false; + +#if PADDLE_WITH_CUDA + if (std::is_same::value) { + // Copy data from GPU -> CPU + if (cudaMemcpy(output.data.data(), tensor->mutable_data(), + tensor->valid_size() * sizeof(float), + cudaMemcpyDeviceToHost) != 0) { + VLOG(3) << "copy data from GPU to CPU error"; + return false; + } + } +#endif + if (std::is_same::value) { + memcpy(output.data.data(), tensor->mutable_data(), + tensor->valid_size() * sizeof(float)); } - cudaStreamSynchronize(NULL); } return true; } @@ -132,7 +185,7 @@ PaddleInferenceAnakinPredictor::Clone() { auto anakin_predictor_p = dynamic_cast *>(cls.get()); if (!anakin_predictor_p) { - LOG(ERROR) << "fail to call Init"; + VLOG(3) << "fail to call Init"; return nullptr; } anakin_predictor_p->get_executer().init(graph_); @@ -162,6 +215,44 @@ std::unique_ptr CreatePaddlePredictor< VLOG(3) << "Anakin Predictor create on unknown platform."; return nullptr; } -}; +} + +#ifdef PADDLE_ANAKIN_ENABLE_OP_TIMER +template +using executor_t = + anakin::Net; + +template +void DisplayOpTimer(executor_t *net_executor, int epoch) { + std::vector op_time = net_executor->get_op_time(); + auto exec_funcs = net_executor->get_exec_funcs(); + auto op_param = net_executor->get_op_param(); + for (int i = 0; i < op_time.size(); i++) { + LOG(INFO) << "name: " << exec_funcs[i].name + << " op_type: " << exec_funcs[i].op_name + << " op_param: " << op_param[i] << " time " << op_time[i] / epoch; + } + std::map op_map; + for (int i = 0; i < op_time.size(); i++) { + auto it = op_map.find(op_param[i]); + if (it != op_map.end()) + op_map[op_param[i]] += op_time[i]; + else + op_map.insert(std::pair(op_param[i], op_time[i])); + } + for (auto it = op_map.begin(); it != op_map.end(); ++it) { + LOG(INFO) << it->first << " " << (it->second) / epoch << " ms"; + } +} +#endif + +template +PaddleInferenceAnakinPredictor::~PaddleInferenceAnakinPredictor() { +#ifdef PADDLE_ANAKIN_ENABLE_OP_TIMER + DisplayOpTimer(executor_p_, max_batch_size_); +#endif + delete executor_p_; + executor_p_ = nullptr; +} } // namespace paddle diff --git a/paddle/fluid/inference/api/api_anakin_engine.h b/paddle/fluid/inference/api/api_anakin_engine.h index 836badd979..dd08661880 100644 --- a/paddle/fluid/inference/api/api_anakin_engine.h +++ b/paddle/fluid/inference/api/api_anakin_engine.h @@ -47,10 +47,7 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor { anakin::Net& get_executer(); - ~PaddleInferenceAnakinPredictor() override { - delete executor_p_; - executor_p_ = nullptr; - }; + ~PaddleInferenceAnakinPredictor() override; private: bool Init(const AnakinConfig& config); @@ -60,6 +57,7 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor { anakin::Net* executor_p_{nullptr}; AnakinConfig config_; + int max_batch_size_{0}; }; } // namespace paddle diff --git a/paddle/fluid/inference/api/api_anakin_engine_rnn_tester.cc b/paddle/fluid/inference/api/api_anakin_engine_rnn_tester.cc new file mode 100644 index 0000000000..6183864234 --- /dev/null +++ b/paddle/fluid/inference/api/api_anakin_engine_rnn_tester.cc @@ -0,0 +1,315 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include +#include +#include +#include +#include +#include // NOLINT +#include +#include "framework/core/net/net.h" +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +DEFINE_string(model, "", "Directory of the inference model."); +DEFINE_string(datapath, "", "Path of the dataset."); +DEFINE_int32(batch_size, 1, "batch size."); +DEFINE_int32(repeat, 1, "Running the inference program repeat times."); + +// Timer for timer +class Timer { + public: + double start; + double startu; + void tic() { + struct timeval tp; + gettimeofday(&tp, NULL); + start = tp.tv_sec; + startu = tp.tv_usec; + } + double toc() { + struct timeval tp; + gettimeofday(&tp, NULL); + double used_time_ms = + (tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0; + return used_time_ms; + } +}; + +std::vector string_split(std::string in_str, + std::string delimiter) { + std::vector seq; + int found = in_str.find(delimiter); + int pre_found = -1; + while (found != std::string::npos) { + if (pre_found == -1) { + seq.push_back(in_str.substr(0, found)); + } else { + seq.push_back(in_str.substr(pre_found + delimiter.length(), + found - delimiter.length() - pre_found)); + } + pre_found = found; + found = in_str.find(delimiter, pre_found + delimiter.length()); + } + seq.push_back( + in_str.substr(pre_found + 1, in_str.length() - (pre_found + 1))); + return seq; +} +std::vector string_split( + std::string in_str, std::vector& delimiter) { // NOLINT + std::vector in; + std::vector out; + out.push_back(in_str); + for (auto del : delimiter) { + in = out; + out.clear(); + for (auto s : in) { + auto out_s = string_split(s, del); + for (auto o : out_s) { + out.push_back(o); + } + } + } + return out; +} + +class Data { + public: + Data(std::string file_name, int batch_size) + : _batch_size(batch_size), _total_length(0) { + _file.open(file_name); + _file.seekg(_file.end); + _total_length = _file.tellg(); + _file.seekg(_file.beg); + } + void get_batch_data(std::vector>& fea, // NOLINT + std::vector>& week_fea, // NOLINT + std::vector>& time_fea, // NOLINT + std::vector& seq_offset); // NOLINT + + private: + std::fstream _file; + int _total_length; + int _batch_size; +}; + +void Data::get_batch_data( + std::vector>& fea, // NOLINT + std::vector>& week_fea, // NOLINT + std::vector>& time_fea, // NOLINT + std::vector& seq_offset) { // NOLINT + int seq_num = 0; + long unsigned int cum = 0; // NOLINT + + char buf[10000]; + seq_offset.clear(); + seq_offset.push_back(0); + fea.clear(); + week_fea.clear(); + time_fea.clear(); + while (_file.getline(buf, 10000)) { + std::string s = buf; + std::vector deli_vec = {":"}; + std::vector data_vec = string_split(s, deli_vec); + + std::vector seq; + seq = string_split(data_vec[0], {"|"}); + + for (auto link : seq) { + std::vector data = string_split(link, ","); + std::vector vec; + for (int i = 0; i < data.size(); i++) { + vec.push_back(atof(data[i].c_str())); + } + fea.push_back(vec); + } + std::vector week_data; + std::vector time_data; + + week_data = string_split(data_vec[2], ","); + std::vector vec_w; + for (int i = 0; i < week_data.size(); i++) { + vec_w.push_back(atof(week_data[i].c_str())); + } + week_fea.push_back(vec_w); + + time_data = string_split(data_vec[1], ","); + std::vector vec_t; + for (int i = 0; i < time_data.size(); i++) { + vec_t.push_back(atof(time_data[i].c_str())); + } + time_fea.push_back(vec_t); + + cum += seq.size(); + seq_offset.push_back(cum); + + seq_num++; + if (seq_num >= _batch_size) { + break; + } + } +} + +namespace paddle { + +AnakinConfig GetConfig() { + AnakinConfig config; + // using AnakinConfig::X86 if you need to use cpu to do inference + config.target_type = AnakinConfig::X86; + config.model_file = FLAGS_model; + config.device = 0; + config.max_batch_size = 1000; // the max number of token + return config; +} + +void set_tensor(std::string name, std::vector shape, + std::vector& vec) { // NOLINT + int sum = 1; + std::for_each(shape.begin(), shape.end(), [&](int n) { sum *= n; }); + float* data = new float[sum]; + PaddleTensor tensor; + tensor.name = name; + tensor.shape = shape; + tensor.data = PaddleBuf(data, sum); + tensor.dtype = PaddleDType::FLOAT32; + vec.push_back(tensor); +} + +void single_test() { + AnakinConfig config = GetConfig(); + auto predictor = + CreatePaddlePredictor(config); + + int max_batch_size = 1000; + std::string feature_file = FLAGS_datapath; + Data map_data(feature_file, FLAGS_batch_size); + std::vector> fea; + std::vector> week_fea; + std::vector> time_fea; + std::vector seq_offset; // NOLINT + + paddle::PaddleTensor tensor_0, tensor_1, tensor_2; + tensor_0.name = "input_0"; + tensor_1.name = "input_4"; + tensor_2.name = "input_5"; + + PaddleTensor tensor_out; + tensor_out.name = "final_output.tmp_1_gout"; + tensor_out.shape = std::vector({}); + tensor_out.data = PaddleBuf(); + tensor_out.dtype = PaddleDType::FLOAT32; + + std::vector inputs; + std::vector outputs(1, tensor_out); + + int data_0_dim = 38; + int data_1_dim = 10; + int data_2_dim = 10; + float data_0[max_batch_size * data_0_dim]; // NOLINT + float data_1[max_batch_size * data_1_dim]; // NOLINT + float data_2[max_batch_size * data_2_dim]; // NOLINT + + int count = 0; + while (true) { + if (count++ > 0) break; // only run the first batch in ci. + seq_offset.clear(); + map_data.get_batch_data(fea, week_fea, time_fea, seq_offset); + if (seq_offset.size() <= 1) { + LOG(FATAL) << "seq_offset.size() <= 1, exit."; + break; + } + + std::vector> seq_offset_vec; // NOLINT + seq_offset_vec.push_back(seq_offset); + tensor_0.lod = seq_offset_vec; + + int p_shape_0[] = {(int)fea.size(), 1, 1, data_0_dim}; // NOLINT + int p_shape_1[] = {(int)week_fea.size(), data_1_dim, 1, 1}; // NOLINT + int p_shape_2[] = {(int)time_fea.size(), data_2_dim, 1, 1}; // NOLINT + + std::vector shape_0(p_shape_0, p_shape_0 + 4); + std::vector shape_1(p_shape_1, p_shape_1 + 4); + std::vector shape_2(p_shape_2, p_shape_2 + 4); + + tensor_0.shape = shape_0; + tensor_1.shape = shape_1; + tensor_2.shape = shape_2; + + for (int i = 0; i < fea.size(); i++) { + memcpy(data_0 + i * data_0_dim, &fea[i][0], sizeof(float) * data_0_dim); + } + for (int i = 0; i < week_fea.size(); i++) { + memcpy(data_1 + i * data_1_dim, &week_fea[i][0], + sizeof(float) * data_1_dim); + } + for (int i = 0; i < time_fea.size(); i++) { + memcpy(data_2 + i * data_2_dim, &time_fea[i][0], + sizeof(float) * data_2_dim); + } + + tensor_0.data = + paddle::PaddleBuf(data_0, fea.size() * sizeof(float) * data_0_dim); + tensor_1.data = + paddle::PaddleBuf(data_1, week_fea.size() * sizeof(float) * data_1_dim); + tensor_2.data = + paddle::PaddleBuf(data_2, time_fea.size() * sizeof(float) * data_2_dim); + + tensor_0.dtype = paddle::PaddleDType::FLOAT32; + tensor_1.dtype = paddle::PaddleDType::FLOAT32; + tensor_2.dtype = paddle::PaddleDType::FLOAT32; + + inputs.clear(); + inputs.push_back(tensor_1); + inputs.push_back(tensor_2); + inputs.push_back(tensor_0); + + Timer timer; + timer.tic(); + for (int i = 0; i < FLAGS_repeat; i++) predictor->Run(inputs, &outputs); + + LOG(INFO) << "batch_size = " << FLAGS_batch_size + << ", repeat = " << FLAGS_repeat + << ", sequence_length = " << seq_offset[seq_offset.size() - 1] + << ", latency: " << timer.toc() / FLAGS_repeat << "ms"; + + float* data_o = static_cast(outputs[0].data.data()); + VLOG(3) << "outputs[0].data.length() = " << outputs[0].data.length(); + for (size_t j = 0; j < outputs[0].data.length(); ++j) { + VLOG(3) << "output[" << j << "]: " << data_o[j]; + } + } +} +} // namespace paddle + +int main(int argc, char** argv) { + google::ParseCommandLineFlags(&argc, &argv, true); + logger::init(argv[0]); + + paddle::single_test(); + /* multi-threads + std::vector threads; + int num = 1; + for (int i = 0; i < num; i++) { + LOG(INFO) << " thread id : " << i; + threads.emplace_back(paddle::single_test); + } + for (int i = 0; i < num; i++) { + threads[i].join(); + } + threads.clear(); + */ + + return 0; +} diff --git a/paddle/fluid/inference/api/api_impl.cc b/paddle/fluid/inference/api/api_impl.cc index e31c637e96..32a691b81f 100644 --- a/paddle/fluid/inference/api/api_impl.cc +++ b/paddle/fluid/inference/api/api_impl.cc @@ -137,8 +137,11 @@ bool NativePaddlePredictor::Run(const std::vector &inputs, return false; } for (size_t i = 0; i < feed_target_names_.size(); ++i) { - VLOG(4) << "setting " << i << "-th target"; - feed_targets[feed_target_names_[i]] = &feeds[i]; + if (config_.specify_input_name) { + feed_targets[inputs[i].name] = &feeds[i]; + } else { + feed_targets[feed_target_names_[i]] = &feeds[i]; + } } // get fetch variable std::map fetch_targets; diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc index 45b5a7638b..9ac0372971 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc @@ -15,6 +15,7 @@ #include "paddle/fluid/inference/analysis/analyzer.h" #include "paddle/fluid/inference/api/api_impl.h" #include "paddle/fluid/inference/api/paddle_inference_api.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/utils/singleton.h" #include "paddle/fluid/operators/tensorrt_engine_op.h" @@ -32,7 +33,8 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor { bool Init(const std::shared_ptr& parent_scope) { VLOG(3) << "Predictor::init()"; - + FLAGS_tensorrt_max_batch_size = config_.max_batch_size; + FLAGS_tensorrt_workspace_size = config_.workspace_size; if (config_.use_gpu) { place_ = paddle::platform::CUDAPlace(config_.device); } else { @@ -150,3 +152,12 @@ CreatePaddlePredictor( } } // namespace paddle + +USE_TRT_CONVERTER(elementwise_add_weight); +USE_TRT_CONVERTER(mul); +USE_TRT_CONVERTER(conv2d); +USE_TRT_CONVERTER(relu); +USE_TRT_CONVERTER(fc); +USE_TRT_CONVERTER(pool2d); +USE_TRT_CONVERTER(softmax); +USE_TRT_CONVERTER(batch_norm); diff --git a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc index fcbf9b89d6..8f1a72316d 100644 --- a/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc +++ b/paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc @@ -23,7 +23,7 @@ namespace paddle { DEFINE_string(dirname, "", "Directory of the inference model."); void CompareTensorRTWithFluid(bool enable_tensorrt) { - FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = enable_tensorrt; + FLAGS_IA_enable_tensorrt_subgraph_engine = enable_tensorrt; //# 1. Create PaddlePredictor with a config. NativeConfig config0; diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh index 3e829dd726..7824ef2649 100755 --- a/paddle/fluid/inference/api/demo_ci/run.sh +++ b/paddle/fluid/inference/api/demo_ci/run.sh @@ -13,16 +13,22 @@ else use_gpu_list='false' fi +PREFIX=inference-vis-demos%2F +URL_ROOT=http://paddlemodels.bj.bcebos.com/${PREFIX} + # download vis_demo data function download() { dir_name=$1 mkdir -p $dir_name cd $dir_name - wget -q ${URL_ROOT}$dir_name.tar.gz - tar xzf *.tar.gz + if [[ -e "${PREFIX}${dir_name}.tar.gz" ]]; then + echo "${PREFIX}{dir_name}.tar.gz has been downloaded." + else + wget -q ${URL_ROOT}$dir_name.tar.gz + tar xzf *.tar.gz + fi cd .. } -URL_ROOT=http://paddlemodels.bj.bcebos.com/inference-vis-demos%2F mkdir -p data cd data vis_demo_list='se_resnext50 ocr mobilenet' diff --git a/paddle/fluid/inference/api/helper.h b/paddle/fluid/inference/api/helper.h new file mode 100644 index 0000000000..2c166cc062 --- /dev/null +++ b/paddle/fluid/inference/api/helper.h @@ -0,0 +1,110 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include +#include +#include +#include +#include "paddle/fluid/inference/api/paddle_inference_api.h" + +namespace paddle { +namespace inference { + +// Timer for timer +class Timer { + public: + double start; + double startu; + void tic() { + struct timeval tp; + gettimeofday(&tp, NULL); + start = tp.tv_sec; + startu = tp.tv_usec; + } + double toc() { + struct timeval tp; + gettimeofday(&tp, NULL); + double used_time_ms = + (tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0; + return used_time_ms; + } +}; + +void split(const std::string &str, char sep, std::vector *pieces) { + pieces->clear(); + if (str.empty()) { + return; + } + size_t pos = 0; + size_t next = str.find(sep, pos); + while (next != std::string::npos) { + pieces->push_back(str.substr(pos, next - pos)); + pos = next + 1; + next = str.find(sep, pos); + } + if (!str.substr(pos).empty()) { + pieces->push_back(str.substr(pos)); + } +} +void split_to_float(const std::string &str, char sep, std::vector *fs) { + std::vector pieces; + split(str, sep, &pieces); + std::transform(pieces.begin(), pieces.end(), std::back_inserter(*fs), + [](const std::string &v) { return std::stof(v); }); +} +template +std::string to_string(const std::vector &vec) { + std::stringstream ss; + for (const auto &c : vec) { + ss << c << " "; + } + return ss.str(); +} +template <> +std::string to_string>( + const std::vector> &vec) { + std::stringstream ss; + for (const auto &piece : vec) { + ss << to_string(piece) << "\n"; + } + return ss.str(); +} +template <> +std::string to_string>>( + const std::vector>> &vec) { + std::stringstream ss; + for (const auto &line : vec) { + for (const auto &rcd : line) { + ss << to_string(rcd) << ";\t"; + } + ss << '\n'; + } + return ss.str(); +} +// clang-format off +void TensorAssignData(PaddleTensor *tensor, const std::vector> &data) { + // Assign buffer + int dim = std::accumulate(tensor->shape.begin(), tensor->shape.end(), 1, [](int a, int b) { return a * b; }); + tensor->data.Resize(sizeof(float) * dim); + int c = 0; + for (const auto &f : data) { + for (float v : f) { static_cast(tensor->data.data())[c++] = v; } + } +} + +} // namespace inference +} // namespace paddle diff --git a/paddle/fluid/inference/api/high_level_api_cn.md b/paddle/fluid/inference/api/high_level_api_cn.md index 2fb914592c..442c598978 100644 --- a/paddle/fluid/inference/api/high_level_api_cn.md +++ b/paddle/fluid/inference/api/high_level_api_cn.md @@ -65,13 +65,13 @@ config.model_dir = "xxx"; config.use_gpu = false; // 创建一个原生的 PaddlePredictor auto predictor = - paddle::CreatePaddlePredictor(config); + paddle::CreatePaddlePredictor(config); // 创建输入 tensor int64_t data[4] = {1, 2, 3, 4}; paddle::PaddleTensor tensor{.name = "", .shape = std::vector({4, 1}), - .data = PaddleBuf(data, sizeof(data)), - .dtype = PaddleDType::INT64}; + .data = paddle::PaddleBuf(data, sizeof(data)), + .dtype = paddle::PaddleDType::INT64}; // 创建输出 tensor，输出 tensor 的内存可以复用 std::vector outputs; // 执行预测 diff --git a/paddle/fluid/inference/api/paddle_inference_api.h b/paddle/fluid/inference/api/paddle_inference_api.h index 794534467b..36fd0727aa 100644 --- a/paddle/fluid/inference/api/paddle_inference_api.h +++ b/paddle/fluid/inference/api/paddle_inference_api.h @@ -45,7 +45,7 @@ class PaddleBuf { PaddleBuf(void* data, size_t length) : data_(data), length_(length), memory_owned_{false} {} // Own memory. - explicit PaddleBuf(size_t length) + PaddleBuf(size_t length) : data_(new char[length]), length_(length), memory_owned_(true) {} // Resize to `length` bytes. void Resize(size_t length); @@ -70,7 +70,7 @@ struct PaddleTensor { std::vector shape; PaddleBuf data; // blob of data. PaddleDType dtype; - std::vector> lod; // lod data + std::vector> lod; // Tensor+LoD equals LoDTensor }; enum class PaddleEngineKind { @@ -120,6 +120,8 @@ struct NativeConfig : public PaddlePredictor::Config { bool use_gpu{false}; int device{0}; float fraction_of_gpu_memory{-1.f}; // Negative to notify initialization. + // Specify the variable's name of each input. + bool specify_input_name{false}; std::string prog_file; std::string param_file; @@ -137,6 +139,14 @@ struct AnakinConfig : public PaddlePredictor::Config { struct TensorRTConfig : public NativeConfig { // Determine whether a subgraph will be executed by TRT. int min_subgraph_size{1}; + // While TensorRT allows an engine optimized for a given max batch size + // to run at any smaller size, the performance for those smaller + // sizes may not be as well-optimized. Therefore, Max batch is best + // equivalent to the runtime batch size. + int max_batch_size{1}; + // For workspace_size, refer it from here: + // https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting + int workspace_size{1 << 30}; }; // A factory to help create different predictors. diff --git a/paddle/fluid/inference/tensorrt/CMakeLists.txt b/paddle/fluid/inference/tensorrt/CMakeLists.txt index b52d083f28..a610687a5b 100644 --- a/paddle/fluid/inference/tensorrt/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/CMakeLists.txt @@ -1,4 +1,4 @@ -nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto) +nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto device_context) nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader) nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine) add_subdirectory(convert) diff --git a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt index 6863b035d8..2a449eb95e 100644 --- a/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt +++ b/paddle/fluid/inference/tensorrt/convert/CMakeLists.txt @@ -1,7 +1,7 @@ # Add TRT tests nv_library(tensorrt_converter SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc -activation_op.cc softmax_op.cc +batch_norm_op.cc activation_op.cc softmax_op.cc DEPS tensorrt_engine operator scope framework_proto op_registry) nv_test(test_op_converter SRCS test_op_converter.cc DEPS @@ -24,3 +24,6 @@ nv_test(test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc nv_test(test_trt_softmax_op SRCS test_softmax_op.cc softmax_op.cc DEPS ${FLUID_CORE_MODULES} tensorrt_engine softmax_op SERIAL) + +nv_test(test_trt_batch_norm_op SRCS test_batch_norm_op.cc batch_norm_op.cc + DEPS ${FLUID_CORE_MODULES} tensorrt_engine batch_norm_op SERIAL) diff --git a/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc new file mode 100644 index 0000000000..94f8b0ae56 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/batch_norm_op.cc @@ -0,0 +1,136 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +class BatchNormOpConverter : public OpConverter { + public: + void operator()(const framework::proto::OpDesc& op, + const framework::Scope& scope, bool test_mode) override { + LOG(INFO) << "convert a fluid batch norm op to tensorrt batch_norm"; + + framework::OpDesc op_desc(op, nullptr); + PADDLE_ENFORCE_EQ(op_desc.Input("X").size(), 1); + PADDLE_ENFORCE_EQ(op_desc.Input("Bias").size(), 1); // Bias is a weight + PADDLE_ENFORCE_EQ(op_desc.Input("Mean").size(), 1); // Mean is a weight + PADDLE_ENFORCE_EQ(op_desc.Input("Scale").size(), 1); // Scale is a weight + PADDLE_ENFORCE_EQ(op_desc.Input("Variance").size(), + 1); // Variance is a weight + PADDLE_ENFORCE_EQ(op_desc.Output("Y").size(), 1); + + auto* X = engine_->GetITensor(op_desc.Input("X").front()); + // Declare weights + auto* Bias_v = scope.FindVar(op_desc.Input("Bias").front()); + auto* Mean_v = scope.FindVar(op_desc.Input("Mean").front()); + auto* Scale_v = scope.FindVar(op_desc.Input("Scale").front()); + auto* Variance_v = scope.FindVar(op_desc.Input("Variance").front()); + const float eps = boost::get(op_desc.GetAttr("epsilon")); + + PADDLE_ENFORCE_NOT_NULL(Bias_v); + PADDLE_ENFORCE_NOT_NULL(Mean_v); + PADDLE_ENFORCE_NOT_NULL(Scale_v); + PADDLE_ENFORCE_NOT_NULL(Variance_v); + + // get tensor + auto* Bias_t = Bias_v->GetMutable(); + auto* Mean_t = Mean_v->GetMutable(); + auto* Scale_t = Scale_v->GetMutable(); + auto* Variance_t = Variance_v->GetMutable(); + + // create temp tensor for weights + framework::LoDTensor bias_tensor; + framework::LoDTensor mean_tensor; + framework::LoDTensor scale_tensor; + framework::LoDTensor variance_tensor; + + bias_tensor.Resize(Bias_t->dims()); + mean_tensor.Resize(Mean_t->dims()); + scale_tensor.Resize(Scale_t->dims()); + variance_tensor.Resize(Variance_t->dims()); + + platform::CPUPlace cpu_place; + // copy data from gpu to cpu + TensorCopySync((*Bias_t), cpu_place, &bias_tensor); + TensorCopySync((*Mean_t), cpu_place, &mean_tensor); + TensorCopySync((*Scale_t), cpu_place, &scale_tensor); + TensorCopySync((*Variance_t), cpu_place, &variance_tensor); + + auto* bias_data = bias_tensor.mutable_data(platform::CPUPlace()); + auto* mean_data = mean_tensor.mutable_data(platform::CPUPlace()); + auto* scale_data = scale_tensor.mutable_data(platform::CPUPlace()); + auto* variance_data = + variance_tensor.mutable_data(platform::CPUPlace()); + + std::unique_ptr combile_scale_tensor( + new framework::LoDTensor()); + std::unique_ptr combile_bias_tensor( + new framework::LoDTensor()); + + combile_scale_tensor->Resize(scale_tensor.dims()); + combile_bias_tensor->Resize(bias_tensor.dims()); + + auto* combile_scale_data = + combile_scale_tensor->mutable_data(platform::CPUPlace()); + auto* combile_bias_data = + combile_bias_tensor->mutable_data(platform::CPUPlace()); + + size_t ele_num = combile_scale_tensor->memory_size() / sizeof(float); + + for (size_t i = 0; i < ele_num; i++) { + float scale = scale_data[i]; + float bias = bias_data[i]; + float mean = mean_data[i]; + float variance = variance_data[i]; + combile_scale_data[i] = scale / sqrtf(variance + eps); + combile_bias_data[i] = bias - mean * combile_scale_data[i]; + } + + TensorRTEngine::Weight scale_weights{ + nvinfer1::DataType::kFLOAT, static_cast(combile_scale_data), + combile_scale_tensor->memory_size() / sizeof(float)}; + TensorRTEngine::Weight shift_weights{ + nvinfer1::DataType::kFLOAT, static_cast(combile_bias_data), + combile_bias_tensor->memory_size() / sizeof(float)}; + TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr, + 0}; + + nvinfer1::IScaleLayer* layer = + TRT_ENGINE_ADD_LAYER(engine_, Scale, *const_cast(X), + nvinfer1::ScaleMode::kCHANNEL, shift_weights.get(), + scale_weights.get(), power_weights.get()); + + auto output_name = op_desc.Output("Y").front(); + engine_->weight_map[op_desc.Input("Bias").front()] = + std::move(combile_bias_tensor); + engine_->weight_map[op_desc.Input("Scale").front()] = + std::move(combile_scale_tensor); + + engine_->SetITensor(output_name, layer->getOutput(0)); + + if (test_mode) { + engine_->DeclareOutput(output_name); + } + } +}; + +} // namespace tensorrt +} // namespace inference +} // namespace paddle + +REGISTER_TRT_OP_CONVERTER(batch_norm, BatchNormOpConverter); diff --git a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc index dba1d50b2d..841a95db38 100644 --- a/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/conv2d_op.cc @@ -35,12 +35,20 @@ class Conv2dOpConverter : public OpConverter { auto* Y_v = scope.FindVar(op_desc.Input("Filter").front()); PADDLE_ENFORCE_NOT_NULL(Y_v); auto* Y_t = Y_v->GetMutable(); - auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); - PADDLE_ENFORCE_EQ(Y_t->dims().size(), 4UL); - const int n_output = Y_t->dims()[0]; - const int filter_h = Y_t->dims()[2]; - const int filter_w = Y_t->dims()[3]; + platform::CPUPlace cpu_place; + std::unique_ptr weight_tensor( + new framework::LoDTensor()); + weight_tensor->Resize(Y_t->dims()); + TensorCopySync((*Y_t), cpu_place, weight_tensor.get()); + + auto* weight_data = + weight_tensor->mutable_data(platform::CPUPlace()); + + PADDLE_ENFORCE_EQ(weight_tensor->dims().size(), 4UL); + const int n_output = weight_tensor->dims()[0]; + const int filter_h = weight_tensor->dims()[2]; + const int filter_w = weight_tensor->dims()[3]; const int groups = boost::get(op_desc.GetAttr("groups")); const std::vector dilations = @@ -57,7 +65,7 @@ class Conv2dOpConverter : public OpConverter { TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, static_cast(weight_data), - Y_t->memory_size() / sizeof(float)}; + weight_tensor->memory_size() / sizeof(float)}; TensorRTEngine::Weight bias{nvinfer1::DataType::kFLOAT, nullptr, 0}; auto* layer = TRT_ENGINE_ADD_LAYER( @@ -70,6 +78,8 @@ class Conv2dOpConverter : public OpConverter { layer->setNbGroups(groups); auto output_name = op_desc.Output("Output").front(); + engine_->weight_map[op_desc.Input("Filter").front()] = + std::move(weight_tensor); engine_->SetITensor(output_name, layer->getOutput(0)); if (test_mode) { engine_->DeclareOutput(output_name); diff --git a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc index 3744550f60..60a72b4eb5 100644 --- a/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/elementwise_op.cc @@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" namespace paddle { @@ -40,10 +39,17 @@ class ElementwiseWeightOpConverter : public OpConverter { auto* Y_v = scope.FindVar(op_desc.Input("Y").front()); PADDLE_ENFORCE_NOT_NULL(Y_v); auto* Y_t = Y_v->GetMutable(); - auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); + + platform::CPUPlace cpu_place; + std::unique_ptr weight_tensor( + new framework::LoDTensor()); + weight_tensor->Resize(Y_t->dims()); + TensorCopySync((*Y_t), cpu_place, weight_tensor.get()); + auto* weight_data = + weight_tensor->mutable_data(platform::CPUPlace()); auto scale_mode = nvinfer1::ScaleMode::kELEMENTWISE; - std::vector dims_y = framework::vectorize2int(Y_t->dims()); + std::vector dims_y = framework::vectorize2int(weight_tensor->dims()); if (static_cast(dims_y.size()) == dims_x.nbDims + 1) { if (dims_y[0] == 1) dims_y.erase(dims_y.begin()); } @@ -70,9 +76,9 @@ class ElementwiseWeightOpConverter : public OpConverter { PADDLE_THROW("TensorRT unsupported weight Shape for Elementwise op!"); } - TensorRTEngine::Weight shift_weights{nvinfer1::DataType::kFLOAT, - static_cast(weight_data), - Y_t->memory_size() / sizeof(float)}; + TensorRTEngine::Weight shift_weights{ + nvinfer1::DataType::kFLOAT, static_cast(weight_data), + weight_tensor->memory_size() / sizeof(float)}; TensorRTEngine::Weight scale_weights{nvinfer1::DataType::kFLOAT, nullptr, 0}; TensorRTEngine::Weight power_weights{nvinfer1::DataType::kFLOAT, nullptr, @@ -82,6 +88,8 @@ class ElementwiseWeightOpConverter : public OpConverter { engine_, Scale, *const_cast(X), scale_mode, shift_weights.get(), scale_weights.get(), power_weights.get()); auto output_name = op_desc.Output("Out")[0]; + + engine_->weight_map[op_desc.Input("Y").front()] = std::move(weight_tensor); engine_->SetITensor(output_name, layer->getOutput(0)); if (test_mode) { // the test framework can not determine which is the // output, so place the declaration inside. diff --git a/paddle/fluid/inference/tensorrt/convert/fc_op.cc b/paddle/fluid/inference/tensorrt/convert/fc_op.cc index 39fe1f609d..ad98d85aae 100644 --- a/paddle/fluid/inference/tensorrt/convert/fc_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/fc_op.cc @@ -12,12 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/framework/eigen.h" -#include "paddle/fluid/framework/lod_tensor.h" -#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" -#include "paddle/fluid/inference/tensorrt/engine.h" -#include "paddle/fluid/platform/place.h" namespace paddle { namespace inference { @@ -73,19 +68,26 @@ class FcOpConverter : public OpConverter { auto* Y_t = Y_v->GetMutable(); // This may trigger a GPU->CPU copy, because TRT's weight can only be // assigned from CPU memory, that can't be avoided. - auto* weight_data = Y_t->mutable_data(platform::CPUPlace()); - PADDLE_ENFORCE_EQ(Y_t->dims().size(), 2UL); // a matrix - size_t n_output = Y_t->dims()[1]; + platform::CPUPlace cpu_place; + framework::LoDTensor weight_tensor; + weight_tensor.Resize(Y_t->dims()); + TensorCopySync((*Y_t), cpu_place, &weight_tensor); - framework::LoDTensor tmp; - tmp.Resize(Y_t->dims()); - memcpy(tmp.mutable_data(platform::CPUPlace()), weight_data, + auto* weight_data = weight_tensor.mutable_data(platform::CPUPlace()); + + PADDLE_ENFORCE_EQ(weight_tensor.dims().size(), 2UL); // a matrix + size_t n_output = weight_tensor.dims()[1]; + + std::unique_ptr tmp(new framework::LoDTensor()); + tmp->Resize(weight_tensor.dims()); + + memcpy(tmp->mutable_data(platform::CPUPlace()), weight_data, Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float)); TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT, static_cast(weight_data), Y_t->memory_size() / sizeof(float)}; TensorRTEngine::Weight tmp_weight(nvinfer1::DataType::kFLOAT, - static_cast(tmp.data()), + static_cast(tmp->data()), Y_t->memory_size() / sizeof(float)); weight.dims.assign({Y_t->dims()[0], Y_t->dims()[1]}); tmp_weight.dims = weight.dims; @@ -106,6 +108,7 @@ class FcOpConverter : public OpConverter { auto output_name = op_desc.Output("Out").front(); engine_->SetITensor(output_name, layer->getOutput(0)); + engine_->weight_map[op_desc.Input("Y").front()] = std::move(tmp); if (test_mode) { engine_->DeclareOutput(output_name); } diff --git a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc index 11cad95361..73f1b28ddf 100644 --- a/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/pool2d_op.cc @@ -33,6 +33,7 @@ class Pool2dOpConverter : public OpConverter { PADDLE_ENFORCE_EQ(op_desc.Output("Out").size(), 1); auto* input1 = engine_->GetITensor(op_desc.Input("X")[0]); + bool global_pooling = boost::get(op_desc.GetAttr("global_pooling")); std::string pool_type = boost::get(op_desc.GetAttr("pooling_type")); std::vector ksize = @@ -42,7 +43,13 @@ class Pool2dOpConverter : public OpConverter { std::vector paddings = boost::get>(op_desc.GetAttr("paddings")); - const nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); + nvinfer1::DimsHW nv_ksize(ksize[0], ksize[1]); + if (global_pooling == true) { + nvinfer1::Dims input_shape = input1->getDimensions(); + int nbDims = input_shape.nbDims; + nv_ksize.d[0] = input_shape.d[nbDims - 2]; + nv_ksize.d[1] = input_shape.d[nbDims - 1]; + } const nvinfer1::DimsHW nv_strides(strides[0], strides[1]); const nvinfer1::DimsHW nv_paddings(paddings[0], paddings[1]); diff --git a/paddle/fluid/inference/tensorrt/convert/test_batch_norm_op.cc b/paddle/fluid/inference/tensorrt/convert/test_batch_norm_op.cc new file mode 100644 index 0000000000..41412cb079 --- /dev/null +++ b/paddle/fluid/inference/tensorrt/convert/test_batch_norm_op.cc @@ -0,0 +1,71 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" +#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" + +namespace paddle { +namespace inference { +namespace tensorrt { + +TEST(batch_norm_op, test) { + std::unordered_set parameters( + {"batch_norm_scale", "batch_norm_bias", "batch_norm_mean", + "batch_norm_variance"}); + framework::Scope scope; + TRTConvertValidation validator(5, parameters, scope, 1 << 15); + std::vector param_shape{2}; + + validator.DeclInputVar("batch_norm_X", nvinfer1::DimsCHW(2, 5, 5)); + validator.DeclParamVar("batch_norm_scale", param_shape); + validator.DeclParamVar("batch_norm_bias", param_shape); + validator.DeclParamVar("batch_norm_mean", param_shape); + validator.DeclParamVar("batch_norm_variance", param_shape); + validator.DeclOutputVar("batch_norm_Y", nvinfer1::DimsCHW(2, 5, 5)); + validator.DeclOutputVar("batch_norm_save_mean", param_shape); + validator.DeclOutputVar("batch_norm_save_variance", param_shape); + + // Prepare Op description + framework::OpDesc desc; + + desc.SetType("batch_norm"); + desc.SetInput("X", {"batch_norm_X"}); + desc.SetInput("Scale", {"batch_norm_scale"}); + desc.SetInput("Bias", {"batch_norm_bias"}); + desc.SetInput("Mean", {"batch_norm_mean"}); + desc.SetInput("Variance", {"batch_norm_variance"}); + desc.SetOutput("Y", {"batch_norm_Y"}); + desc.SetOutput("MeanOut", {"batch_norm_mean"}); + desc.SetOutput("VarianceOut", {"batch_norm_variance"}); + desc.SetOutput("SavedMean", {"batch_norm_save_mean"}); + desc.SetOutput("SavedVariance", {"batch_norm_save_variance"}); + + float eps = 1e-5f; + bool is_test = true; + desc.SetAttr("epsilon", eps); + desc.SetAttr("is_test", is_test); + + validator.SetOp(*desc.Proto()); + + std::unordered_set neglected_output = { + "batch_norm_save_mean", "batch_norm_save_variance", "batch_norm_mean", + "batch_norm_variance"}; + validator.Execute(3, neglected_output); +} + +} // namespace tensorrt +} // namespace inference +} // namespace paddle +USE_OP(batch_norm); diff --git a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc index d6651a5b24..01d7f700da 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_op_converter.cc @@ -57,6 +57,7 @@ TEST(OpConverter, ConvertBlock) { auto* x = scope.Var("conv2d-Y"); auto* x_tensor = x->GetMutable(); x_tensor->Resize(framework::make_ddim(dim_vec)); + x_tensor->mutable_data(platform::CUDAPlace(0)); OpConverter converter; converter.ConvertBlock(*block->Proto(), {"conv2d-Y"}, scope, diff --git a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc index c5dddbc8cd..aedd6b62df 100644 --- a/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc +++ b/paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc @@ -20,7 +20,7 @@ namespace paddle { namespace inference { namespace tensorrt { -TEST(Pool2dOpConverter, main) { +void test_pool2d(bool global_pooling) { framework::Scope scope; std::unordered_set parameters; TRTConvertValidation validator(5, parameters, scope, 1 << 15); @@ -28,7 +28,10 @@ TEST(Pool2dOpConverter, main) { // The ITensor's Dims should not contain the batch size. // So, the ITensor's Dims of input and output should be C * H * W. validator.DeclInputVar("pool2d-X", nvinfer1::Dims3(3, 4, 4)); - validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2)); + if (global_pooling) + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 1, 1)); + else + validator.DeclOutputVar("pool2d-Out", nvinfer1::Dims3(3, 2, 2)); // Prepare Op description framework::OpDesc desc; @@ -45,6 +48,7 @@ TEST(Pool2dOpConverter, main) { desc.SetAttr("ksize", ksize); desc.SetAttr("strides", strides); desc.SetAttr("paddings", paddings); + desc.SetAttr("global_pooling", global_pooling); LOG(INFO) << "set OP"; validator.SetOp(*desc.Proto()); @@ -53,6 +57,10 @@ TEST(Pool2dOpConverter, main) { validator.Execute(3); } +TEST(Pool2dOpConverter, normal) { test_pool2d(false); } + +TEST(Pool2dOpConverter, test_global_pooling) { test_pool2d(true); } + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/convert/ut_helper.h b/paddle/fluid/inference/tensorrt/convert/ut_helper.h index 4265f33f28..0a6f171fc4 100644 --- a/paddle/fluid/inference/tensorrt/convert/ut_helper.h +++ b/paddle/fluid/inference/tensorrt/convert/ut_helper.h @@ -24,6 +24,7 @@ limitations under the License. */ #include "paddle/fluid/framework/lod_tensor.h" #include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/framework/tensor_util.h" #include "paddle/fluid/inference/analysis/helper.h" #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" @@ -48,11 +49,17 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place, auto dims = tensor->dims(); size_t num_elements = analysis::AccuDims(dims, dims.size()); PADDLE_ENFORCE_GT(num_elements, 0); - auto* data = tensor->mutable_data(place); + + platform::CPUPlace cpu_place; + framework::LoDTensor temp_tensor; + temp_tensor.Resize(dims); + auto* temp_data = temp_tensor.mutable_data(cpu_place); for (size_t i = 0; i < num_elements; i++) { - *(data + i) = random(0., 1.); + *(temp_data + i) = random(0., 1.); } + + TensorCopySync(temp_tensor, place, tensor); } /* @@ -91,18 +98,26 @@ class TRTConvertValidation { engine_->DeclareInput(name, nvinfer1::DataType::kFLOAT, dims); } + void DeclParamVar(const std::string& name, const std::vector dim_vec) { + DeclVar(name, dim_vec); + } + // Declare a parameter varaible in the scope. void DeclParamVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims, true); } + void DeclOutputVar(const std::string& name, const std::vector dim_vec) { + DeclVar(name, dim_vec); + } + void DeclOutputVar(const std::string& name, const nvinfer1::Dims& dims) { DeclVar(name, dims); } void DeclVar(const std::string& name, const std::vector dim_vec) { - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); auto* x = scope_.Var(name); auto* x_tensor = x->GetMutable(); @@ -141,18 +156,22 @@ class TRTConvertValidation { PADDLE_ENFORCE(var); auto tensor = var->GetMutable(); - engine_->SetInputFromCPU( + engine_->SetInputFromGPU( input, static_cast(tensor->data()), sizeof(float) * analysis::AccuDims(tensor->dims(), tensor->dims().size())); } } - void Execute(int batch_size) { + // We use the set 'neglected_output' here, because some Ops like batch norm, + // the outputs specified in the op des are only used during training, + // so we should neglect those output during inference. + void Execute(int batch_size, + std::unordered_set neglected_output = {}) { // Execute Fluid Op PADDLE_ENFORCE_LE(batch_size, max_batch_size_); - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); op_->Run(scope_, place); // Execute TRT. engine_->Execute(batch_size); @@ -161,6 +180,7 @@ class TRTConvertValidation { ASSERT_FALSE(op_desc_->OutputArgumentNames().empty()); const size_t output_space_size = 3000; for (const auto& output : op_desc_->OutputArgumentNames()) { + if (neglected_output.count(output)) continue; std::vector fluid_out; std::vector trt_out(output_space_size); engine_->GetOutputInCPU(output, &trt_out[0], output_space_size); diff --git a/paddle/fluid/inference/tensorrt/engine.cc b/paddle/fluid/inference/tensorrt/engine.cc index b821c3d0bf..14e9e14d33 100644 --- a/paddle/fluid/inference/tensorrt/engine.cc +++ b/paddle/fluid/inference/tensorrt/engine.cc @@ -33,6 +33,7 @@ void TensorRTEngine::Build(const DescType &paddle_model) { } void TensorRTEngine::Execute(int batch_size) { + freshDeviceId(); batch_size_ = batch_size; std::vector buffers; for (auto &buf : buffers_) { @@ -60,6 +61,7 @@ TensorRTEngine::~TensorRTEngine() { } void TensorRTEngine::FreezeNetwork() { + freshDeviceId(); PADDLE_ENFORCE(infer_builder_ != nullptr, "Call InitNetwork first to initialize network."); PADDLE_ENFORCE(infer_network_ != nullptr, @@ -241,6 +243,13 @@ void TensorRTEngine::SetRuntimeBatch(size_t batch_size) { int TensorRTEngine::GetRuntimeBatch() { return runtime_batch_; } +void TensorRTEngine::freshDeviceId() { + int count; + cudaGetDeviceCount(&count); + PADDLE_ENFORCE_LT(device_, count); + cudaSetDevice(device_); +} + } // namespace tensorrt } // namespace inference } // namespace paddle diff --git a/paddle/fluid/inference/tensorrt/engine.h b/paddle/fluid/inference/tensorrt/engine.h index 694468c419..bd3ba4cea6 100644 --- a/paddle/fluid/inference/tensorrt/engine.h +++ b/paddle/fluid/inference/tensorrt/engine.h @@ -19,6 +19,7 @@ limitations under the License. */ #include #include #include +#include "paddle/fluid/framework/tensor.h" #include "paddle/fluid/inference/engine.h" #include "paddle/fluid/inference/tensorrt/helper.h" #include "paddle/fluid/inference/utils/singleton.h" @@ -52,13 +53,15 @@ class TensorRTEngine : public EngineBase { }; TensorRTEngine(int max_batch, int max_workspace, - cudaStream_t* stream = nullptr, + cudaStream_t* stream = nullptr, int device = 0, nvinfer1::ILogger& logger = NaiveLogger::Global()) : max_batch_(max_batch), max_workspace_(max_workspace), stream_(stream ? stream : &default_stream_), - logger_(logger) { - cudaStreamCreate(&default_stream_); + logger_(logger), + device_(device) { + freshDeviceId(); + cudaStreamCreate(stream_); } virtual ~TensorRTEngine(); @@ -119,6 +122,15 @@ class TensorRTEngine : public EngineBase { nvinfer1::INetworkDefinition* network() { return infer_network_.get(); } void SetRuntimeBatch(size_t batch_size); int GetRuntimeBatch(); + int GetDevice() { return device_; } + + // A pointer to CPU memory is needed of the TRT weight. + // Before TRT runs, fluid loads weight into GPU storage. + // so we need to copy the weights from GPU to CPU in our op converter. + // We use a map to store these weights for the weight memory is not released + // in advance, which affecting the construction of TRT Op. + std::unordered_map> + weight_map; private: // the max batch size @@ -140,6 +152,8 @@ class TensorRTEngine : public EngineBase { std::unordered_map buffer_sizes_; std::unordered_map itensor_map_; + // The specific GPU id that the TensorRTEngine bounded to. + int device_; // TensorRT related internal members template @@ -156,6 +170,10 @@ class TensorRTEngine : public EngineBase { infer_ptr infer_network_; infer_ptr infer_engine_; infer_ptr infer_context_; + // Each ICudaEngine object is bound to a specific GPU when it is instantiated, + // ensure that the thread is associated with the correct device by calling + // freshDeviceId(). + void freshDeviceId(); }; // class TensorRTEngine // Add an layer__ into engine__ with args ARGS. @@ -188,8 +206,8 @@ class TRT_EngineManager { // Create or get an engine called `name` TensorRTEngine* Create(int max_batch, int max_workspace, cudaStream_t* stream, - const std::string& name) { - auto* p = new TensorRTEngine(max_batch, max_workspace, stream); + const std::string& name, int gpu_device = 0) { + auto* p = new TensorRTEngine(max_batch, max_workspace, stream, gpu_device); engines_[name].reset(p); return p; } diff --git a/paddle/fluid/inference/tensorrt/test_engine.cc b/paddle/fluid/inference/tensorrt/test_engine.cc index dc03702990..da1f6535cb 100644 --- a/paddle/fluid/inference/tensorrt/test_engine.cc +++ b/paddle/fluid/inference/tensorrt/test_engine.cc @@ -27,7 +27,7 @@ namespace tensorrt { class TensorRTEngineTest : public ::testing::Test { protected: void SetUp() override { - ASSERT_EQ(0, cudaStreamCreate(&stream_)); + // ASSERT_EQ(0, cudaStreamCreate(&stream_)); engine_ = new TensorRTEngine(10, 1 << 10, &stream_); engine_->InitNetwork(); } diff --git a/paddle/fluid/operators/CMakeLists.txt b/paddle/fluid/operators/CMakeLists.txt index e8b5dec9d4..68fbde2c09 100644 --- a/paddle/fluid/operators/CMakeLists.txt +++ b/paddle/fluid/operators/CMakeLists.txt @@ -9,7 +9,6 @@ function(op_library TARGET) # op_library is a function to create op library. The interface is same as # cc_library. But it handle split GPU/CPU code and link some common library # for ops. - set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE) set(cc_srcs) set(cu_srcs) set(hip_cu_srcs) @@ -84,6 +83,16 @@ function(op_library TARGET) message(FATAL_ERROR "The op library ${TARGET} should contains at least one .cc file") endif() + #remove windows unsupported op + if (WIN32) + foreach(windows_unsupport_op "nccl_op" "gen_nccl_id_op") + if ("${TARGET}" STREQUAL "${windows_unsupport_op}") + return() + endif() + endforeach() + endif(WIN32) + set(OP_LIBRARY ${TARGET} ${OP_LIBRARY} PARENT_SCOPE) + list(LENGTH op_library_DEPS op_library_DEPS_len) if (${op_library_DEPS_len} GREATER 0) set(DEPS_OPS ${TARGET} ${DEPS_OPS} PARENT_SCOPE) @@ -100,7 +109,8 @@ function(op_library TARGET) endif() # Define operators that don't need pybind here. - foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" "tensor_array_read_write_op") + foreach(manual_pybind_op "compare_op" "logical_op" "nccl_op" +"tensor_array_read_write_op" "tensorrt_engine_op") if ("${TARGET}" STREQUAL "${manual_pybind_op}") set(pybind_flag 1) endif() @@ -180,19 +190,19 @@ function(op_library TARGET) endfunction() add_subdirectory(math) +if (NOT WIN32) add_subdirectory(nccl) - if(WITH_GPU) op_library(nccl_op DEPS nccl_common) file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(ncclAllReduce);\n") else() set(DEPS_OPS ${DEPS_OPS} nccl_op) endif() +endif() # NOT WIN32 set(DISTRIBUTE_DEPS "") if(WITH_DISTRIBUTE) add_subdirectory(distributed) - set(DISTRIBUTE_DEPS "") if(WITH_GRPC) set(DISTRIBUTE_DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf node) @@ -221,7 +231,7 @@ if(WITH_DISTRIBUTE) #set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) #cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op # listen_and_serv_op sum_op executor SERIAL) - if(WITH_GPU) + if(WITH_GPU AND NOT WIN32) set_source_files_properties(test_send_nccl_id.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) cc_test(test_send_nccl_id SRCS test_send_nccl_id.cc DEPS listen_and_serv_op ${DISTRIBUTE_DEPS} executor SERIAL) if(WITH_GRPC) @@ -232,7 +242,7 @@ if(WITH_DISTRIBUTE) set_source_files_properties(gen_nccl_id_op.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS}) else() set(DEPS_OPS ${DEPS_OPS} gen_nccl_id_op) - endif() + endif() # WITH_GPU AND NOT WIN32 else() set(DEPS_OPS ${DEPS_OPS} checkpoint_notify_op prefetch_op recv_op listen_and_serv_op send_op send_barrier_op fetch_barrier_op gen_nccl_id_op) endif() @@ -248,6 +258,7 @@ op_library(softmax_op DEPS softmax) op_library(sequence_softmax_op DEPS softmax) if (WITH_GPU AND TENSORRT_FOUND) op_library(tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter) + file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(tensorrt_engine);\n") nv_test(test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc DEPS tensorrt_engine_op analysis) @@ -329,5 +340,7 @@ cc_test(beam_search_op_test SRCS beam_search_op_test.cc DEPS lod_tensor beam_sea cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory) cc_test(save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op) cc_test(save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op) +if(NOT WIN32) nv_test(nccl_op_test SRCS nccl_op_test.cu.cc DEPS nccl_op gpu_info device_context) +endif() nv_test(dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor) diff --git a/paddle/fluid/operators/activation_op.cu b/paddle/fluid/operators/activation_op.cu index d3a7ceed46..27487b396c 100644 --- a/paddle/fluid/operators/activation_op.cu +++ b/paddle/fluid/operators/activation_op.cu @@ -26,8 +26,6 @@ namespace plat = paddle::platform; act_type##_grad, ops::ActivationGradKernel>, \ ops::ActivationGradKernel>, \ - ops::ActivationGradKernel>); + ops::grad_functor>); FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL); diff --git a/paddle/fluid/operators/activation_op.h b/paddle/fluid/operators/activation_op.h index 48f3b5a5bc..9124151926 100644 --- a/paddle/fluid/operators/activation_op.h +++ b/paddle/fluid/operators/activation_op.h @@ -333,7 +333,8 @@ struct SqrtGradFunctor : public BaseActivationFunctor { template void operator()(Device d, X x, Out out, dOut dout, dX dx) const { - dx.device(d) = static_cast(0.5) * dout / out; + const Out out_conj = Eigen::numext::conj(out); + dx.device(d) = static_cast(0.5) * dout / out_conj; } }; @@ -739,7 +740,7 @@ struct PowGradFunctor : public BaseActivationFunctor { typename dX> void operator()(Device d, X x, Out out, dOut dout, dX dx) const { dx.device(d) = dout * static_cast(factor) * - x.pow(static_cast(factor) - static_cast(1)); + x.pow(static_cast(factor - static_cast(1))); } }; @@ -862,11 +863,10 @@ struct SwishGradFunctor : public BaseActivationFunctor { template void operator()(Device d, X x, Out out, dOut dout, dX dx) const { - T b = static_cast(beta); auto temp1 = static_cast(1) / - (static_cast(1) + (static_cast(-b) * x).exp()); - auto temp2 = temp1 * (static_cast(1) - (b * out)); - dx.device(d) = dout * ((b * out) + temp2); + (static_cast(1) + (static_cast(-beta) * x).exp()); + auto temp2 = temp1 * (static_cast(1) - (beta * out)); + dx.device(d) = dout * ((beta * out) + temp2); } }; diff --git a/paddle/fluid/operators/assign_value_op.cu.cc b/paddle/fluid/operators/assign_value_op.cu.cc index 0ff174b388..08bfde5dc9 100644 --- a/paddle/fluid/operators/assign_value_op.cu.cc +++ b/paddle/fluid/operators/assign_value_op.cu.cc @@ -13,10 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/assign_value_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(assign_value, ops::AssignValueKernel, - ops::AssignValueKernel, - ops::AssignValueKernel); + ops::AssignValueKernel); diff --git a/paddle/fluid/operators/conditional_block_op.cc b/paddle/fluid/operators/conditional_block_op.cc index 580fde7538..135254ce6b 100644 --- a/paddle/fluid/operators/conditional_block_op.cc +++ b/paddle/fluid/operators/conditional_block_op.cc @@ -29,9 +29,9 @@ class ConditionalOp : public framework::OperatorBase { protected: std::vector InputTensors( - const framework::Scope &scope) const { + const framework::Scope &scope, const std::string &in_name) const { std::vector retv; - auto xs = Inputs("X"); + auto xs = Inputs(in_name); retv.resize(xs.size(), nullptr); std::transform( xs.begin(), xs.end(), retv.begin(), @@ -81,12 +81,18 @@ class ConditionalBlockOp : public ConditionalOp { private: void RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const override { - auto xs = InputTensors(scope); - bool need_run; if (Attr("is_scalar_condition")) { + // When is_scalar_condition is True, the conditional variable is a scalar, + // whether need to execute the operators in sub-block depends on the + // conditional variable (Cond). + auto xs = InputTensors(scope, "Cond"); need_run = ScalarCondition(xs); } else { + // When is_scalar_condition is False, the conditional variable maybe a + // vector or tensor, whether need to execute the operators in sub-block + // depends on the input variables (Input). + auto xs = InputTensors(scope, "Input"); need_run = std::all_of( xs.begin(), xs.end(), [](const framework::LoDTensor *t) { return t->numel() != 0; }); @@ -110,11 +116,11 @@ class ConditionalBlockOp : public ConditionalOp { class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { - AddInput("X", - "The conditional variable of this operator. If X is empty, the " + AddInput("Cond", + "The conditional variable of this operator. If Cond is empty, the " "whole sub-block will not be executed.") .AsDuplicable(); - AddInput("Params", "The input variables of the sub-block.").AsDuplicable(); + AddInput("Input", "The input variables of the sub-block.").AsDuplicable(); AddOutput("Out", "The output variables of the sub-block.").AsDuplicable(); AddOutput("Scope", "(std::vector) The step scope of conditional block. To " @@ -123,13 +129,18 @@ class ConditionalBlockOpProtoMaker : public framework::OpProtoAndCheckerMaker { AddAttr( "sub_block", "The step block of conditional block operator"); AddAttr("is_scalar_condition", - "the input X is used as scalar " - "condition") + "The conditional variable (Cond) is used as scalar " + "condition.") .SetDefault(false); AddComment(R"DOC(Conditional block operator -Run the sub-block if X is not empty. Params is the other inputs and Out is the -outputs of the sub-block. +If `is_scalar_condition` is True, the conditional variable (Cond) is a scalar, +run the operators in sub-block if Cond is True. + +If `is_scalar_condition` is False, the conditional variable (Cond) is a vector or +tensor, run the operators in sub-block if all of input variables are not empty. + + )DOC"); } }; @@ -145,12 +156,12 @@ class ConditionalBlockGradOp : public ConditionalOp { private: void RunImpl(const framework::Scope &scope, const platform::Place &dev_place) const override { - auto xs = this->InputTensors(scope); - bool need_run; if (Attr("is_scalar_condition")) { + auto xs = this->InputTensors(scope, "Cond"); need_run = ScalarCondition(xs); } else { + auto xs = this->InputTensors(scope, "Input"); need_run = std::all_of( xs.begin(), xs.end(), [](const framework::LoDTensor *t) { return t->numel() != 0; }); @@ -166,11 +177,11 @@ class ConditionalBlockGradOp : public ConditionalOp { auto *block = Attr("sub_block"); exec.Run(*block->Program(), &cur_scope, block->ID(), false); - AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Params"), - Outputs(framework::GradVarName("Params"))); + AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Input"), + Outputs(framework::GradVarName("Input"))); - AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("X"), - Outputs(framework::GradVarName("X"))); + AssignLocalGradientToGlobal(dev_place, cur_scope, Inputs("Cond"), + Outputs(framework::GradVarName("Cond"))); } } @@ -199,15 +210,15 @@ class ConditionalBlockGradOp : public ConditionalOp { class ConditionalBlockGradInferShape : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext *context) const override { - PADDLE_ENFORCE(context->HasInputs("X")); - if (context->HasInputs("Params")) { - PADDLE_ENFORCE(context->HasOutputs(framework::GradVarName("Params"))); - context->SetOutputsDim(framework::GradVarName("Params"), - context->GetInputsDim("Params")); + PADDLE_ENFORCE(context->HasInputs("Cond")); + if (context->HasInputs("Input")) { + PADDLE_ENFORCE(context->HasOutputs(framework::GradVarName("Input"))); + context->SetOutputsDim(framework::GradVarName("Input"), + context->GetInputsDim("Input")); } - if (context->HasOutputs(framework::GradVarName("X"))) { - context->SetOutputsDim(framework::GradVarName("X"), - context->GetInputsDim("X")); + if (context->HasOutputs(framework::GradVarName("Cond"))) { + context->SetOutputsDim(framework::GradVarName("Cond"), + context->GetInputsDim("Cond")); } } }; @@ -220,14 +231,15 @@ class ConditionalBlockGradMaker : public framework::SingleGradOpDescMaker { std::unique_ptr Apply() const override { auto grad_op = new framework::OpDesc(); grad_op->SetType("conditional_block_grad"); - grad_op->SetInput("X", Input("X")); - grad_op->SetInput("Params", Input("Params")); + grad_op->SetInput("Cond", Input("Cond")); + grad_op->SetInput("Input", Input("Input")); grad_op->SetInput("Out", Output("Out")); grad_op->SetInput(framework::GradVarName("Out"), OutputGrad("Out")); grad_op->SetInput("Scope", Output("Scope")); - grad_op->SetOutput(framework::GradVarName("X"), InputGrad("X", false)); - grad_op->SetOutput(framework::GradVarName("Params"), - InputGrad("Params", false)); + grad_op->SetOutput(framework::GradVarName("Cond"), + InputGrad("Cond", false)); + grad_op->SetOutput(framework::GradVarName("Input"), + InputGrad("Input", false)); grad_op->SetBlockAttr("sub_block", this->grad_block_[0]); grad_op->SetAttr("is_scalar_condition", GetAttr("is_scalar_condition")); return std::unique_ptr(grad_op); diff --git a/paddle/fluid/operators/conv_cudnn_op.cu.cc b/paddle/fluid/operators/conv_cudnn_op.cu.cc index 59bfe8f61d..22cbf680c0 100644 --- a/paddle/fluid/operators/conv_cudnn_op.cu.cc +++ b/paddle/fluid/operators/conv_cudnn_op.cu.cc @@ -39,27 +39,6 @@ using ScalingParamType = typename platform::CudnnDataType::ScalingParamType; static constexpr size_t kCONV_CUDNN_WORKSPACE_LIMIT_BYTES = static_cast(1024) * 1024 * 1024; -template -// bool EnableFp16(const T& dummy, const DeviceContext& dev_ctx, -bool EnableFp16(const DeviceContext& dev_ctx, - cudnnConvolutionDescriptor_t cudnn_conv_desc) { -#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1) - // Tensor core is supported since the volta GPU and - // is only enabled when input and filter data are float16 - if (dev_ctx.GetComputeCapability() >= 70 && - std::type_index(typeid(T)) == - std::type_index(typeid(platform::float16))) { - PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( - cudnn_conv_desc, CUDNN_TENSOR_OP_MATH)); - return true; - } else { - PADDLE_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( - cudnn_conv_desc, CUDNN_DEFAULT_MATH)); - } -#endif - return false; -} - template class CUDNNConvOpKernel : public framework::OpKernel { public: @@ -149,14 +128,27 @@ class CUDNNConvOpKernel : public framework::OpKernel { cudnnConvolutionFwdAlgo_t algo; auto& dev_ctx = ctx.template device_context(); auto handle = dev_ctx.cudnn_handle(); - if (EnableFp16(dev_ctx, cudnn_conv_desc)) { + + CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( + handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, + cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, + workspace_size_limit, &algo)); + +#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1) + // Tensor core is supported since the volta GPU and + // is only enabled when input and filter data are float16 + if (dev_ctx.GetComputeCapability() >= 70 && + std::type_index(typeid(T)) == + std::type_index(typeid(platform::float16))) { + CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( + cudnn_conv_desc, CUDNN_TENSOR_OP_MATH)); + // Currently tensor core is only enabled using this algo algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM; } else { - PADDLE_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm( - handle, cudnn_input_desc, cudnn_filter_desc, cudnn_conv_desc, - cudnn_output_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, - workspace_size_limit, &algo)); + CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType( + cudnn_conv_desc, CUDNN_DEFAULT_MATH)); } +#endif // get workspace size able to allocate CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize( @@ -296,9 +288,6 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } else { data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1; } - if (EnableFp16(dev_ctx, cudnn_conv_desc)) { - data_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1; - } CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize( @@ -318,9 +307,6 @@ class CUDNNConvGradOpKernel : public framework::OpKernel { } else { filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1; } - if (EnableFp16(dev_ctx, cudnn_conv_desc)) { - filter_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1; - } CUDNN_ENFORCE( platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize( @@ -376,8 +362,7 @@ REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel); REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, - paddle::operators::CUDNNConvGradOpKernel, - paddle::operators::CUDNNConvGradOpKernel); + paddle::operators::CUDNNConvGradOpKernel); REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel, @@ -385,5 +370,4 @@ REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvOpKernel); REGISTER_OP_KERNEL(conv3d_grad, CUDNN, plat::CUDAPlace, paddle::operators::CUDNNConvGradOpKernel, - paddle::operators::CUDNNConvGradOpKernel, - paddle::operators::CUDNNConvGradOpKernel) + paddle::operators::CUDNNConvGradOpKernel); diff --git a/paddle/fluid/operators/conv_mkldnn_op.cc b/paddle/fluid/operators/conv_mkldnn_op.cc index f07ab5a33b..527a87db53 100644 --- a/paddle/fluid/operators/conv_mkldnn_op.cc +++ b/paddle/fluid/operators/conv_mkldnn_op.cc @@ -126,6 +126,15 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler { pipeline); } + std::shared_ptr AcquireBiasMemoryFromPrimitive( + const std::shared_ptr user_bias_memory_p, + std::vector& pipeline) { // NOLINT + auto user_bias_pd = user_bias_memory_p->get_primitive_desc(); + auto bias_pd = conv_pd_->bias_primitive_desc(); + return this->AcquireMemory(bias_pd, user_bias_pd, user_bias_memory_p, + "@bias_mem_p", pipeline); + } + std::shared_ptr AcquireConvolution( std::shared_ptr src_memory_p, std::shared_ptr weights_memory_p, @@ -147,6 +156,28 @@ class ConvMKLDNNHandler : public platform::MKLDNNHandler { return conv_p; } + std::shared_ptr AcquireConvolution( + std::shared_ptr src_memory_p, + std::shared_ptr weights_memory_p, + std::shared_ptr bias_memory_p, + std::shared_ptr dst_memory_p) { + auto prim_key = key_ + "@conv_p"; + auto conv_p = std::static_pointer_cast( + dev_ctx_.GetBlob(prim_key)); + PADDLE_ENFORCE((conv_p != nullptr) || (is_reusing_ == false), + "Fail to find convolution primitive in device context"); + if (conv_p == nullptr) { + conv_p = std::make_shared( + *conv_pd_, *(src_memory_p), *(weights_memory_p.get()), + *(bias_memory_p.get()), *(dst_memory_p.get())); + + dev_ctx_.SetBlob(prim_key, conv_p); + } else { + is_reusing_ = true; + } + return conv_p; + } + std::shared_ptr AcquireConvolutionBackwardWeights( std::shared_ptr src_memory_p, @@ -229,6 +260,7 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { auto* input = ctx.Input("Input"); auto* filter = ctx.Input("Filter"); + auto* bias = ctx.HasInput("Bias") ? ctx.Input("Bias") : nullptr; auto* output = ctx.Output("Output"); PADDLE_ENFORCE(input->layout() == DataLayout::kMKLDNN && @@ -237,6 +269,17 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { PADDLE_ENFORCE(filter->layout() == DataLayout::kMKLDNN && filter->format() != memory::format::format_undef, "Wrong layout/format set for Filter tensor"); + PADDLE_ENFORCE(input->dims().size() == 4, + "Input must be with 4 dimensions, i.e. NCHW"); + PADDLE_ENFORCE(filter->dims().size() == 4, + "Filter must be with 4 dimensions, i.e. OIHW"); + if (bias) { + PADDLE_ENFORCE(bias->layout() == DataLayout::kMKLDNN && + bias->format() != memory::format::format_undef, + "Wrong layout/format set for Bias tensor"); + PADDLE_ENFORCE(bias->dims().size() == 1, + "Bias must only have 1 dimension, i.e. X"); + } std::vector strides = ctx.Attr>("strides"); std::vector paddings = ctx.Attr>("paddings"); @@ -253,11 +296,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { const T* filter_data = filter->data(); T* output_data = output->mutable_data(ctx.GetPlace()); - PADDLE_ENFORCE(input->dims().size() == 4, - "Input must be with 4 dimensions, i.e. NCHW"); - PADDLE_ENFORCE(filter->dims().size() == 4, - "Filter must be with 4 dimensions, i.e. OIHW"); - std::vector src_tz = paddle::framework::vectorize2int(input->dims()); std::vector weights_tz = paddle::framework::vectorize2int(filter->dims()); @@ -288,13 +326,23 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { src_tz, platform::MKLDNNGetDataType(), chosen_memory_format); auto weights_md = platform::MKLDNNMemDesc( weights_tz, platform::MKLDNNGetDataType(), chosen_memory_format); + std::vector bias_tz; // TODO(mgallus): avoid empty vector creation. + // Currently used whenever bias is != nullptr. auto dst_md = platform::MKLDNNMemDesc( dst_tz, platform::MKLDNNGetDataType(), chosen_memory_format); // create a conv primitive descriptor and save it for usage in backward - std::shared_ptr conv_pd = - ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, paddings, - mkldnn_engine); + std::shared_ptr conv_pd; + if (bias) { + bias_tz = paddle::framework::vectorize2int(bias->dims()); + auto bias_md = platform::MKLDNNMemDesc( + bias_tz, platform::MKLDNNGetDataType(), memory::format::x); + conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, bias_md, dst_md, + strides, paddings, mkldnn_engine); + } else { + conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides, + paddings, mkldnn_engine); + } // Save conv_pd/src_memory/weights_memory for backward pass dev_ctx.SetBlob(key_conv_pd, conv_pd); @@ -315,8 +363,22 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { handler.AcquireDstMemoryFromPrimitive(to_void_cast(output_data)); // create convolution op primitive - auto conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p, - dst_memory_p); + std::shared_ptr conv_p; + if (bias) { + const T* bias_data = bias->data(); + auto user_bias_md = platform::MKLDNNMemDesc( + {bias_tz}, platform::MKLDNNGetDataType(), memory::format::x); + auto user_bias_memory_p = + handler.AcquireBiasMemory(user_bias_md, to_void_cast(bias_data)); + + auto bias_memory_p = + handler.AcquireBiasMemoryFromPrimitive(user_bias_memory_p, pipeline); + conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p, + bias_memory_p, dst_memory_p); + } else { + conv_p = handler.AcquireConvolution(src_memory_p, weights_memory_p, + dst_memory_p); + } // push primitive to stream and wait until it's executed pipeline.push_back(*conv_p); @@ -346,6 +408,27 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel { return std::unique_ptr( p_conv_pd); } + + std::unique_ptr + ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights, + const memory::desc& bias, const memory::desc& dst, + const std::vector& strides, + const std::vector& paddings, + const mkldnn::engine& engine) const { + memory::dims stride_dims = {strides[0], strides[1]}; + memory::dims padding_dims = {paddings[0], paddings[1]}; + + auto conv_desc = mkldnn::convolution_forward::desc( + mkldnn::prop_kind::forward, mkldnn::convolution_direct, src, weights, + bias, dst, stride_dims, padding_dims, padding_dims, + mkldnn::padding_kind::zero); + + auto p_conv_pd = + new mkldnn::convolution_forward::primitive_desc(conv_desc, engine); + + return std::unique_ptr( + p_conv_pd); + } }; template diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc index 37153d5843..61ca80877a 100644 --- a/paddle/fluid/operators/conv_op.cc +++ b/paddle/fluid/operators/conv_op.cc @@ -37,6 +37,7 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { auto in_dims = ctx->GetInputDim("Input"); auto filter_dims = ctx->GetInputDim("Filter"); + std::vector strides = ctx->Attrs().Get>("strides"); std::vector paddings = ctx->Attrs().Get>("paddings"); int groups = ctx->Attrs().Get("groups"); @@ -57,7 +58,6 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const { PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[1] * groups, "The number of input channels should be equal to filter " "channels * groups."); - PADDLE_ENFORCE_EQ( filter_dims[0] % groups, 0, "The number of output channels should be divided by groups."); @@ -122,6 +122,11 @@ void Conv2DOpMaker::Make() { "H is the height of the filter, and W is the width of the filter. " "If the groups attribute is greater than 1, C equals the number of " "input image channels divided by the groups."); + AddInput("Bias", + "(Tensor) Bias to be added to each output of filter application." + "The format of output tensor is X (one-dimensional) of size equal" + "to the number of output channels. Only used with MKL-DNN.") + .AsDispensable(); AddOutput("Output", "(Tensor) The output tensor of convolution operator. " "The format of output tensor is also NCHW.") diff --git a/paddle/fluid/operators/crf_decoding_op.h b/paddle/fluid/operators/crf_decoding_op.h index 3f5fab3b38..8181897c3d 100644 --- a/paddle/fluid/operators/crf_decoding_op.h +++ b/paddle/fluid/operators/crf_decoding_op.h @@ -85,6 +85,199 @@ class CRFDecodingOpKernel : public framework::OpKernel { int* track_value = track.mutable_data(emission_dims, platform::CPUPlace()); +#ifdef __AVX__ +// It use the AVX or AVX512 instruction to deal the data as the vector of 8 or +// 16 elements per iteration. Then it can implement the parallel processing. +// Only optimize for float type. +#ifdef __AVX512F__ + size_t step_size = 16; +#else + size_t step_size = 8; +#endif + if (std::is_same::value && (tag_num >= step_size)) { + size_t steps = tag_num / step_size; + size_t remain = tag_num % step_size; + int last_offset = static_cast(remain) - static_cast(step_size); + + // Setup the alpha initial value. + size_t i_offset = 0; + for (size_t i = 0; i <= steps; ++i) { +#ifdef __AVX512F__ + // Declare the variable for the content of weights, input and alpha + // values. + __m512 w_content, x_content, alpha_content; + + // Load the relevant data into the variables from un-aligned address. + w_content = _mm512_loadu_ps((const float*)(w + i_offset)); + x_content = _mm512_loadu_ps((const float*)(x + i_offset)); + alpha_content = _mm512_add_ps(w_content, x_content); + + // Save the alpha value. + _mm512_storeu_ps(reinterpret_cast(alpha_value + i_offset), + alpha_content); +#else + // Declare the variable for the content of weights, input and alpha + // values. + __m256 w_content, x_content, alpha_content; + + // Load the relevant data into the variables from un-aligned address. + w_content = _mm256_loadu_ps((const float*)(w + i_offset)); + x_content = _mm256_loadu_ps((const float*)(x + i_offset)); + alpha_content = _mm256_add_ps(w_content, x_content); + + // Save the alpha value. + _mm256_storeu_ps(reinterpret_cast(alpha_value + i_offset), + alpha_content); +#endif + i_offset += step_size; + if (i == steps - 1) { + if (remain > 0) { + i_offset += last_offset; + } else { + break; + } + } + } + + // Use the column-major strategy to get the location of maximum score. + size_t seq_offset = 0; + for (size_t k = 1; k < seq_len; ++k) { + size_t j_offset = 0; + for (size_t j = 0; j <= steps; ++j) { +#ifdef __AVX512F__ + // Initialize the variables of maximum score and location. + __m512 max_score = _mm512_set1_ps(-std::numeric_limits::max()); + __m512i max_j = _mm512_setzero_si512(); +#else + // Initialize the variables of maximum score and location. + __m256 max_score = _mm256_set1_ps(-std::numeric_limits::max()); + __m256i max_j = _mm256_set1_epi32(0); +#endif + // Calculate the offset of transition_weights. + size_t trans_offset = state_trans_base_idx * tag_num + j_offset; + for (size_t i = 0; i < tag_num; ++i) { +#ifdef __AVX512F__ + // Initalize the content of alpha variable with related offset. + __m512 alpha_content = + _mm512_set1_ps(*(const float*)(alpha_value + seq_offset + i)); + // Obtain the content of weights from un-aligned address. + __m512 w_content = + _mm512_loadu_ps((const float*)(w + trans_offset)); + + __m512 score_v = _mm512_add_ps(alpha_content, w_content); + + __mmask16 mask = _mm512_cmp_ps_mask(score_v, max_score, _CMP_GT_OS); + + // According to the mask value, it update the index of the max_score + // location. + max_j = _mm512_mask_set1_epi32(max_j, mask, i); + + // Update the max_score value. + max_score = _mm512_max_ps(max_score, score_v); +#else + // Initalize the content of alpha variable with related offset. + __m256 alpha_content = _mm256_broadcast_ss( + (const float*)(alpha_value + seq_offset + i)); + // Obtain the content of weights from un-aligned address. + __m256 w_content = + _mm256_loadu_ps((const float*)(w + trans_offset)); + __m256 score_v = _mm256_add_ps(alpha_content, w_content); + + __m256 mask = _mm256_cmp_ps(score_v, max_score, _CMP_GT_OS); + +#ifdef __AVX2__ + // According to the mask value, it update the index of the max_score + // location. + max_j = _mm256_or_si256( + _mm256_andnot_si256((__m256i)mask, max_j), + _mm256_and_si256((__m256i)mask, _mm256_set1_epi32(i))); +#else + __m128i lo_max_j = _mm256_extractf128_si256(max_j, 0); + __m128i hi_max_j = _mm256_extractf128_si256(max_j, 1); + __m128i lo_mask = _mm256_extractf128_si256((__m256i)mask, 0); + __m128i hi_mask = _mm256_extractf128_si256((__m256i)mask, 1); + + lo_max_j = _mm_andnot_si128(lo_mask, lo_max_j); + hi_max_j = _mm_andnot_si128(hi_mask, hi_max_j); + lo_mask = _mm_and_si128(lo_mask, _mm_set1_epi32(i)); + hi_mask = _mm_and_si128(hi_mask, _mm_set1_epi32(i)); + + lo_max_j = _mm_or_si128(lo_mask, lo_max_j); + hi_max_j = _mm_or_si128(hi_mask, hi_max_j); + + // According to the mask value, it update the index of the max_score + // location. + max_j = _mm256_insertf128_si256(max_j, lo_max_j, 0); + max_j = _mm256_insertf128_si256(max_j, hi_max_j, 1); +#endif + + // Update the max_score value. + max_score = _mm256_max_ps(max_score, score_v); +#endif + trans_offset += tag_num; + } + +#ifdef __AVX512F__ + // Update the alpha and track values. + __m512 x_content = _mm512_loadu_ps( + (const float*)(x + seq_offset + tag_num + j_offset)); + max_score = _mm512_add_ps(max_score, x_content); + _mm512_storeu_ps(reinterpret_cast(alpha_value + seq_offset + + tag_num + j_offset), + max_score); + _mm512_storeu_si512( + reinterpret_cast<__m512i*>(track_value + seq_offset + tag_num + + j_offset), + max_j); +#else + // Update the alpha and track values. + __m256 x_content = _mm256_loadu_ps( + (const float*)(x + seq_offset + tag_num + j_offset)); + max_score = _mm256_add_ps(max_score, x_content); + _mm256_storeu_ps(reinterpret_cast(alpha_value + seq_offset + + tag_num + j_offset), + max_score); + _mm256_storeu_si256( + reinterpret_cast<__m256i*>(track_value + seq_offset + tag_num + + j_offset), + max_j); +#endif + + // Calculate the offset of next step + j_offset += step_size; + if (j == steps - 1) { + if (remain > 0) { + j_offset += last_offset; + } else { + break; + } + } + } + + seq_offset += tag_num; + } + } else { + for (size_t i = 0; i < tag_num; ++i) alpha_value[i] = w[i] + x[i]; + + for (size_t k = 1; k < seq_len; ++k) { + for (size_t i = 0; i < tag_num; ++i) { + T max_score = -std::numeric_limits::max(); + int max_j = 0; + for (size_t j = 0; j < tag_num; ++j) { + T score = alpha_value[(k - 1) * tag_num + j] + + w[(j + state_trans_base_idx) * tag_num + i]; + if (score > max_score) { + max_score = score; + max_j = j; + } + } + + alpha_value[k * tag_num + i] = max_score + x[k * tag_num + i]; + track_value[k * tag_num + i] = max_j; + } + } + } +#else for (size_t i = 0; i < tag_num; ++i) alpha_value[i] = w[i] + x[i]; for (size_t k = 1; k < seq_len; ++k) { @@ -105,6 +298,7 @@ class CRFDecodingOpKernel : public framework::OpKernel { } } +#endif T max_score = -std::numeric_limits::max(); int max_i = 0; for (size_t i = 0; i < tag_num; ++i) { diff --git a/paddle/fluid/operators/cross_entropy_op.cu b/paddle/fluid/operators/cross_entropy_op.cu index 65fd3a5dbc..30dbd5bd3d 100644 --- a/paddle/fluid/operators/cross_entropy_op.cu +++ b/paddle/fluid/operators/cross_entropy_op.cu @@ -13,16 +13,12 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/cross_entropy_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; using CUDACtx = paddle::platform::CUDADeviceContext; REGISTER_OP_CUDA_KERNEL(cross_entropy, ops::CrossEntropyOpKernel, - ops::CrossEntropyOpKernel, - ops::CrossEntropyOpKernel); -REGISTER_OP_CUDA_KERNEL( - cross_entropy_grad, ops::CrossEntropyGradientOpKernel, - ops::CrossEntropyGradientOpKernel, - ops::CrossEntropyGradientOpKernel); + ops::CrossEntropyOpKernel); +REGISTER_OP_CUDA_KERNEL(cross_entropy_grad, + ops::CrossEntropyGradientOpKernel, + ops::CrossEntropyGradientOpKernel); diff --git a/paddle/fluid/operators/distributed/request_handler_impl.cc b/paddle/fluid/operators/distributed/request_handler_impl.cc index de1a503154..66784f0b51 100644 --- a/paddle/fluid/operators/distributed/request_handler_impl.cc +++ b/paddle/fluid/operators/distributed/request_handler_impl.cc @@ -130,12 +130,13 @@ bool RequestCheckpointHandler::Handle(const std::string& varname, checkpoint_notify_id != -1, "when checkpoint_notify_id = -1, there should be no RPC invoke."); - auto* lt_var = scope->FindVar(LOOKUP_TABLE_PATH)->GetMutable(); + // TODO(tangwei12): find out why scope will be error. + auto* lt_var = scope_->FindVar(LOOKUP_TABLE_PATH)->GetMutable(); lt_var->clear(); lt_var->append(out_var_name); VLOG(4) << "RequestCheckpointHandler update var kLookupTablePath to: " << out_var_name; - executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope); + executor_->RunPreparedContext(checkpoint_prepared_ctx_.get(), scope_); return true; } diff --git a/paddle/fluid/operators/distributed/rpc_server_test.cc b/paddle/fluid/operators/distributed/rpc_server_test.cc index b50830c362..d6176e1443 100644 --- a/paddle/fluid/operators/distributed/rpc_server_test.cc +++ b/paddle/fluid/operators/distributed/rpc_server_test.cc @@ -78,10 +78,9 @@ void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place, int64_t rows_numel) { CreateVarsOnScope(scope, place); auto w = scope->Var("w")->GetMutable(); - auto rows = w->mutable_rows(); - for (int64_t i = 0; i < rows_numel; ++i) rows->push_back(i); auto w_value = w->mutable_value(); w_value->Resize({rows_numel, 10}); + for (int64_t i = 0; i < rows_numel; ++i) w->AutoGrownIndex(i, true); auto ptr = w_value->mutable_data(*place); diff --git a/paddle/fluid/operators/elementwise_add_op.cu b/paddle/fluid/operators/elementwise_add_op.cu index f9f5c66d34..dfff518f17 100644 --- a/paddle/fluid/operators/elementwise_add_op.cu +++ b/paddle/fluid/operators/elementwise_add_op.cu @@ -30,5 +30,4 @@ REGISTER_OP_CUDA_KERNEL( ops::ElementwiseAddGradKernel, ops::ElementwiseAddGradKernel, ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel, - ops::ElementwiseAddGradKernel); + ops::ElementwiseAddGradKernel); diff --git a/paddle/fluid/operators/elementwise_div_op.cu b/paddle/fluid/operators/elementwise_div_op.cu index 4cc7ba0f43..588d1f7420 100644 --- a/paddle/fluid/operators/elementwise_div_op.cu +++ b/paddle/fluid/operators/elementwise_div_op.cu @@ -14,24 +14,19 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/elementwise_div_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( elementwise_div, ops::ElementwiseDivKernel, ops::ElementwiseDivKernel, ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel, - ops::ElementwiseDivKernel); + ops::ElementwiseDivKernel); REGISTER_OP_CUDA_KERNEL( elementwise_div_grad, ops::ElementwiseDivGradKernel, ops::ElementwiseDivGradKernel, ops::ElementwiseDivGradKernel, - ops::ElementwiseDivGradKernel, ops::ElementwiseDivGradKernel); + int64_t>); diff --git a/paddle/fluid/operators/elementwise_mul_op.cu b/paddle/fluid/operators/elementwise_mul_op.cu index 350d43168d..2fb1b4bee6 100644 --- a/paddle/fluid/operators/elementwise_mul_op.cu +++ b/paddle/fluid/operators/elementwise_mul_op.cu @@ -14,25 +14,19 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/elementwise_mul_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( elementwise_mul, ops::ElementwiseMulKernel, ops::ElementwiseMulKernel, ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel, - ops::ElementwiseMulKernel); + ops::ElementwiseMulKernel); REGISTER_OP_CUDA_KERNEL( elementwise_mul_grad, ops::ElementwiseMulGradKernel, ops::ElementwiseMulGradKernel, ops::ElementwiseMulGradKernel, - ops::ElementwiseMulGradKernel, ops::ElementwiseMulGradKernel); diff --git a/paddle/fluid/operators/elementwise_mul_op.h b/paddle/fluid/operators/elementwise_mul_op.h index 82c5fa0472..4437da4d95 100644 --- a/paddle/fluid/operators/elementwise_mul_op.h +++ b/paddle/fluid/operators/elementwise_mul_op.h @@ -15,6 +15,8 @@ limitations under the License. */ #pragma once #include "paddle/fluid/operators/elementwise_op.h" #include "paddle/fluid/operators/elementwise_op_function.h" +#include "paddle/fluid/operators/math/blas.h" + namespace paddle { namespace operators { @@ -23,6 +25,37 @@ struct MulFunctor { inline HOSTDEVICE T operator()(T a, T b) const { return a * b; } }; +template +void default_elementwise_mul(const framework::ExecutionContext& ctx, + const framework::Tensor* x, + const framework::Tensor* y, framework::Tensor* z) { + int axis = ctx.Attr("axis"); + ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, + MulFunctor(), z); +} + +template +typename std::enable_if< + std::is_floating_point::value && + std::is_same::value>::type +elementwise_mul(const framework::ExecutionContext& ctx, + const framework::Tensor* x, const framework::Tensor* y, + framework::Tensor* z) { + auto blas = math::GetBlas(ctx); + blas.VMUL(x->numel(), x->data(), y->data(), + z->mutable_data(ctx.GetPlace())); +} + +template +typename std::enable_if< + !std::is_floating_point::value || + !std::is_same::value>::type +elementwise_mul(const framework::ExecutionContext& ctx, + const framework::Tensor* x, const framework::Tensor* y, + framework::Tensor* z) { + default_elementwise_mul(ctx, x, y, z); +} + template class ElementwiseMulKernel : public framework::OpKernel { public: @@ -33,9 +66,11 @@ class ElementwiseMulKernel : public framework::OpKernel { auto* y = ctx.Input("Y"); auto* z = ctx.Output("Out"); z->mutable_data(ctx.GetPlace()); - int axis = ctx.Attr("axis"); - ElementwiseComputeEx, DeviceContext, T>(ctx, x, y, axis, - MulFunctor(), z); + if (x->numel() == y->numel()) { + elementwise_mul(ctx, x, y, z); + } else { + default_elementwise_mul(ctx, x, y, z); + } } }; diff --git a/paddle/fluid/operators/elementwise_op_function.h b/paddle/fluid/operators/elementwise_op_function.h index 7223a972d2..f90dcdc156 100644 --- a/paddle/fluid/operators/elementwise_op_function.h +++ b/paddle/fluid/operators/elementwise_op_function.h @@ -80,6 +80,9 @@ inline framework::DDim trim_trailing_singular_dims( for (int i = 0; i < actual_dims_size; ++i) { trim_dims[i] = dims[i]; } + if (trim_dims.size() == 0) { + return framework::DDim(framework::make_dim()); + } framework::DDim actual_dims = framework::make_ddim(trim_dims); return actual_dims; } @@ -350,7 +353,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel( int j = blockIdx.x; int i = threadIdx.x; int tid = threadIdx.x; - T val(0); + T val = 0; do { int x_offset = i * w + j; @@ -418,7 +421,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel( int tid = threadIdx.x; int j = blockIdx.x; - T val(0); + T val = 0; int ttid = tid; while (true) { diff --git a/paddle/fluid/operators/elementwise_sub_op.cu b/paddle/fluid/operators/elementwise_sub_op.cu index ff3f6f8a2c..8709f686f9 100644 --- a/paddle/fluid/operators/elementwise_sub_op.cu +++ b/paddle/fluid/operators/elementwise_sub_op.cu @@ -14,25 +14,19 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/elementwise_sub_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( elementwise_sub, ops::ElementwiseSubKernel, ops::ElementwiseSubKernel, ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel, - ops::ElementwiseSubKernel); + ops::ElementwiseSubKernel); REGISTER_OP_CUDA_KERNEL( elementwise_sub_grad, ops::ElementwiseSubGradKernel, ops::ElementwiseSubGradKernel, ops::ElementwiseSubGradKernel, - ops::ElementwiseSubGradKernel, ops::ElementwiseSubGradKernel); diff --git a/paddle/fluid/operators/fc_op.cc b/paddle/fluid/operators/fc_op.cc index 099ca52c8e..fa4dec9cf1 100644 --- a/paddle/fluid/operators/fc_op.cc +++ b/paddle/fluid/operators/fc_op.cc @@ -15,8 +15,7 @@ limitations under the License. */ #include "paddle/fluid/operators/fc_op.h" #include #include "paddle/fluid/operators/math/blas.h" - -DECLARE_int32(paddle_num_threads); +#include "paddle/fluid/operators/math/fc_compute.h" namespace paddle { namespace operators { @@ -36,9 +35,14 @@ void FCOp::InferShape(framework::InferShapeContext* ctx) const { if (ctx->HasInput("Bias")) { auto bias_dims = ctx->GetInputDim("Bias"); - PADDLE_ENFORCE_EQ(bias_dims[0], 1, "The shape of Bias must be [1, dim]."); - PADDLE_ENFORCE_EQ(bias_dims[1], w_dims[1], - "The shape of Bias must be [1, dim]."); + if (bias_dims.size() == 2) { + PADDLE_ENFORCE_EQ(bias_dims[0], 1, "The shape of Bias must be [1, dim]."); + PADDLE_ENFORCE_EQ(bias_dims[1], w_dims[1], + "The shape of Bias must be [1, dim]."); + } else if (bias_dims.size() == 1) { + PADDLE_ENFORCE_EQ(bias_dims[0], w_dims[1], + "The shape of Bias must be [1, dim]."); + } } PADDLE_ENFORCE(in_dims.size() == 2 || in_dims.size() == 4, "Fully Connected input should be 2-D or 4-D tensor."); @@ -110,13 +114,8 @@ void FCOpMaker::Make() { AddComment(R"DOC( Fully Connected Operator. - The fully connected operation calculates the output based on the input, weights and bias attribute. + The fully connected operation calculates the output based on the input, weights and bias. The size of each dimension of the parameters checked in the infer-shape. - The matrix of bias is generated by the mkldnn framework, when the bias_attr is True. - Additional parametrs are use_mkldnn and bias_attr. - The input(X) size and output(Out) size may be diffrent. - - The fully connected layer only supports MKLDNN version )DOC"); } @@ -133,26 +132,15 @@ class FCOpKernel : public framework::OpKernel { auto in_dims = input->dims(); auto w_dims = w->dims(); - auto& dev_ctx = ctx.template device_context(); - auto blas = math::GetBlas(dev_ctx); const T* input_data = input->data(); const T* w_data = w->data(); T* output_data = output->mutable_data(ctx.GetPlace()); + auto blas = math::GetBlas(ctx); + math::FCCompute( + blas, in_dims[0], w_dims[1], w_dims[0], input_data, w_data, output_data, + bias ? bias->data() : NULL); - blas.GEMM(CblasNoTrans, CblasNoTrans, in_dims[0], w_dims[1], w_dims[0], - static_cast(1), input_data, w_data, static_cast(0), - output_data); - - if (bias) { - const T* bias_data = bias->data(); -#ifdef PADDLE_WITH_MKLML -#pragma omp parallel for if (FLAGS_paddle_num_threads > 1) -#endif - for (int bs = 0; bs < in_dims[0]; bs++) { - blas.AXPY(w_dims[1], static_cast(1), bias_data, - output_data + bs * w_dims[1]); - } - } + // TODO(TJ): fuse act } }; diff --git a/paddle/fluid/operators/fill_constant_op.cc b/paddle/fluid/operators/fill_constant_op.cc index 862249269e..130f18dde4 100644 --- a/paddle/fluid/operators/fill_constant_op.cc +++ b/paddle/fluid/operators/fill_constant_op.cc @@ -12,28 +12,48 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include "paddle/fluid/operators/fill_constant_op.h" -#include "paddle/fluid/platform/float16.h" +#include "paddle/fluid/framework/data_type.h" +#include "paddle/fluid/framework/op_registry.h" +#include "paddle/fluid/operators/math/math_function.h" +#include "paddle/fluid/platform/device_context.h" namespace paddle { namespace operators { -class FillConstantOp : public framework::OperatorWithKernel { +class FillConstantInferShape : public framework::InferShapeBase { public: - using framework::OperatorWithKernel::OperatorWithKernel; - - void InferShape(framework::InferShapeContext* ctx) const override { + void operator()(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of FillConstantOp should not be null."); - auto& shape = ctx->Attrs().Get>("shape"); + auto &shape = ctx->Attrs().Get>("shape"); ctx->SetOutputDim("Out", framework::make_ddim(shape)); } +}; + +class FillConstantOp : public framework::OperatorBase { + public: + using framework::OperatorBase::OperatorBase; + + private: + void RunImpl(const framework::Scope &scope, + const platform::Place &dev_place) const override { + auto data_type = + static_cast(Attr("dtype")); + auto value = Attr("value"); + auto force_cpu = Attr("force_cpu"); + auto &out = + *scope.FindVar(Output("Out"))->GetMutable(); + out.Resize(framework::make_ddim(Attr>("shape"))); + if (force_cpu) { + auto cpu = platform::CPUPlace(); + out.mutable_data(cpu, framework::ToTypeIndex(data_type)); + } else { + out.mutable_data(dev_place, framework::ToTypeIndex(data_type)); + } - framework::OpKernelType GetExpectedKernelType( - const framework::ExecutionContext& ctx) const override { - return framework::OpKernelType( - static_cast(ctx.Attr("dtype")), - ctx.device_context()); + platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); + auto &dev_ctx = *pool.Get(dev_place); + math::set_constant(dev_ctx, &out, value); } }; @@ -67,11 +87,6 @@ Fill up a variable with specified constant value. } // namespace paddle namespace ops = paddle::operators; -REGISTER_OPERATOR(fill_constant, ops::FillConstantOp, ops::FillConstantOpMaker, +REGISTER_OPERATOR(fill_constant, ops::FillConstantOp, + ops::FillConstantInferShape, ops::FillConstantOpMaker, paddle::framework::EmptyGradOpMaker); -REGISTER_OP_CPU_KERNEL( - fill_constant, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel, - ops::FillConstantOpKernel) diff --git a/paddle/fluid/operators/fill_constant_op.h b/paddle/fluid/operators/fill_constant_op.h deleted file mode 100644 index b2a2a7b2fa..0000000000 --- a/paddle/fluid/operators/fill_constant_op.h +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once -#include - -#include "paddle/fluid/framework/data_type.h" -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/math_function.h" - -namespace paddle { -namespace operators { - -template -class FillConstantOpKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& ctx) const override { - auto data_type = - static_cast(ctx.Attr("dtype")); - auto value = ctx.Attr("value"); - auto force_cpu = ctx.Attr("force_cpu"); - auto* out = ctx.Output("Out"); - out->Resize(framework::make_ddim(ctx.Attr>("shape"))); - if (force_cpu) { - auto cpu = platform::CPUPlace(); - out->mutable_data(cpu, framework::ToTypeIndex(data_type)); - } else { - out->mutable_data(ctx.GetPlace(), framework::ToTypeIndex(data_type)); - } - - math::set_constant(ctx.template device_context(), out, - value); - } -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/fill_op.cc b/paddle/fluid/operators/fill_op.cc index 352a17c927..925dc19061 100644 --- a/paddle/fluid/operators/fill_op.cc +++ b/paddle/fluid/operators/fill_op.cc @@ -16,7 +16,6 @@ limitations under the License. */ #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/detail/safe_ref.h" #include "paddle/fluid/platform/device_context.h" -#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -70,6 +69,7 @@ class FillOp : public framework::OperatorBase { framework::VisitDataType( dtype, FillOpVisitor(&tensor, Attr>("value"))); + if (!force_cpu && platform::is_gpu_place(place)) { // Copy tensor to out platform::DeviceContextPool &pool = diff --git a/paddle/fluid/operators/fusion_lstm_op.cc b/paddle/fluid/operators/fusion_lstm_op.cc new file mode 100644 index 0000000000..3888333ec5 --- /dev/null +++ b/paddle/fluid/operators/fusion_lstm_op.cc @@ -0,0 +1,354 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/fusion_lstm_op.h" +#include +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/detail/activation_functions.h" +#include "paddle/fluid/operators/math/fc_compute.h" +#include "paddle/fluid/operators/math/lstm_compute.h" +#include "paddle/fluid/operators/math/sequence2batch.h" + +namespace paddle { +namespace operators { + +void FusionLSTMOp::InferShape(framework::InferShapeContext* ctx) const { + PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("WeightX"), + "Input(WeightX) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("WeightH"), + "Input(WeightH) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasInput("Bias"), + "Input(Bias) of LSTM should not be null."); + + PADDLE_ENFORCE(ctx->HasOutput("XX"), + "Output(XX) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Hidden"), + "Output(Hidden) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Cell"), + "Output(Cell) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("BatchedGate"), + "Output(BatchedGate) of LSTM should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("BatchCellPreAct"), + "Output(BatchedGate) of LSTM should not be null."); + + auto x_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank must be 2."); + + if (ctx->HasInput("H0")) { + PADDLE_ENFORCE(ctx->HasInput("C0"), + "Input(Cell) and Input(Hidden) of LSTM should not " + "be null at the same time."); + auto h_dims = ctx->GetInputDim("H0"); + auto c_dims = ctx->GetInputDim("C0"); + PADDLE_ENFORCE(h_dims == c_dims, + "The dimension of Input(H0) and Input(C0) " + "should be the same."); + } + + auto wx_dims = ctx->GetInputDim("WeightX"); + PADDLE_ENFORCE_EQ(wx_dims.size(), 2, + "The rank of Input(WeightX) should be 2."); + PADDLE_ENFORCE_EQ(wx_dims[0], x_dims[1], + "The first dimension of Input(WeightX) " + "should be %d.", + x_dims[1]); + + int frame_size = wx_dims[1] / 4; + auto wh_dims = ctx->GetInputDim("WeightH"); + PADDLE_ENFORCE_EQ(wh_dims.size(), 2, + "The rank of Input(WeightH) should be 2."); + PADDLE_ENFORCE_EQ(wh_dims[0], frame_size, + "The first dimension of Input(WeightH) " + "should be %d.", + frame_size); + PADDLE_ENFORCE_EQ(wh_dims[1], 4 * frame_size, + "The second dimension of Input(WeightH) " + "should be 4 * %d.", + frame_size); + + auto b_dims = ctx->GetInputDim("Bias"); + PADDLE_ENFORCE_EQ(b_dims.size(), 2, "The rank of Input(Bias) should be 2."); + PADDLE_ENFORCE_EQ(b_dims[0], 1, + "The first dimension of Input(Bias) should be 1."); + + PADDLE_ENFORCE(!ctx->Attrs().Get("use_peepholes"), + "Do not support peephole yet."); + PADDLE_ENFORCE_EQ(b_dims[1], 4 * frame_size, + "The second dimension of Input(Bias) should be " + "4 * %d if disable peepholes connection", + frame_size); + + framework::DDim out_dims({x_dims[0], frame_size}); + ctx->SetOutputDim("Hidden", out_dims); + ctx->SetOutputDim("Cell", out_dims); + ctx->SetOutputDim("BatchedGate", {x_dims[0], wx_dims[1]}); + ctx->SetOutputDim("BatchCellPreAct", out_dims); + ctx->ShareLoD("X", "Hidden"); + ctx->ShareLoD("X", "Cell"); + + int xx_width = x_dims[1] > wx_dims[1] ? wx_dims[1] : x_dims[1]; + ctx->SetOutputDim("XX", {x_dims[0], xx_width}); + ctx->ShareLoD("X", "XX"); +} + +framework::OpKernelType FusionLSTMOp::GetExpectedKernelType( + const framework::ExecutionContext& ctx) const { + return framework::OpKernelType( + framework::ToDataType(ctx.Input("X")->type()), + ctx.device_context()); +} + +void FusionLSTMOpMaker::Make() { + AddInput("X", + "(LoDTensor) the input is a LodTensor, which support " + "variable-time length input sequence. The underlying tensor in " + "this LoDTensor is a matrix with shape (T X M), where T is the " + "total time steps in this mini-batch, M is the dim size of x."); + AddInput("WeightX", + "(Tensor) the learnable weights of X." + " - The shape is (M x 4D), where M is the dim size of x, D is the " + "hidden size. " + " - Weight = {W_cx, W_ix, W_fx, W_ox}"); + AddInput("WeightH", + "(Tensor) same as LSTMOp, the learnable hidden-hidden weights." + " - The shape is (D x 4D), where D is the hidden size. " + " - Weight = {W_ch, W_ih, W_fh, W_oh}"); + AddInput("Bias", + "(Tensor) the learnable weights. Almost same as LSTMOp" + "Note: we should add the fc bias into this (1x4D) in bias." + "input-hidden bias weight and peephole connections weight if " + "setting `use_peepholes` True. " + "1. `use_peepholes = False` " + " - The shape is (1 x 4D). " + " - Bias = {b_c, b_i, b_f, b_o}." + "2. `use_peepholes = True` " + " - The shape is (1 x 7D). " + " - Bias = {b_c, b_i, b_f, b_o, W_ic, W_fc, W_oc}."); + AddInput("H0", + "(Tensor, optional) (same as LSTMOp) the initial hidden state is an " + "optional " + "input. This is a tensor with shape (N x D), where N is the " + "batch size and D is the hidden size.") + .AsDispensable(); + AddInput("C0", + "(Tensor, optional) (same as LSTMOp) (the initial cell state is an " + "optional " + "input. This is a tensor with shape (N x D), where N is the " + "batch size. `H0` and `C0` can be NULL but only at the same time.") + .AsDispensable(); + AddOutput("Hidden", + "(LoDTensor) (same as LSTMOp) the hidden state of LSTM operator. " + "The shape is (T x D), and lod is the same with the `Input`."); + AddOutput("Cell", + "(LoDTensor) (same as LSTMOp) the cell state of LSTM operator. " + "The shape is (T x D), and lod is the same with the `Input`."); + AddOutput("XX", + "(LoDTensor) the result after X * WeightX (size is T x 4D)" + " or batched_X (size is T x M), this will be automatically chosen," + " where T is the total time steps in this mini-batch," + " D is the hidden size, M is the dim size of x input.") + .AsIntermediate(); + AddOutput("BatchedGate", "(LoDTensor) (same as LSTMOp).").AsIntermediate(); + AddOutput("BatchCellPreAct", "(LoDTensor) (same as LSTMOp).") + .AsIntermediate(); + AddAttr("use_peepholes", + "(bool, defalut: True) " + "whether to enable diagonal/peephole connections.") + .SetDefault(true); + AddAttr("is_reverse", + "(bool, defalut: False) " + "whether to compute reversed LSTM.") + .SetDefault(false); + AddAttr("gate_activation", + "(string, default: sigmoid)" + "The activation for input gate, forget gate and output " + "gate, `sigmoid` by default.") + .SetDefault("sigmoid") + .InEnum({"sigmoid", "tanh", "relu", "identity"}); + AddAttr("cell_activation", + "(string, default: tanh)" + "The activation for cell output, `tanh` by defalut.") + .SetDefault("tanh") + .InEnum({"sigmoid", "tanh", "relu", "identity"}); + AddAttr("candidate_activation", + "(string, default: tanh)" + "The activation for candidate hidden state, " + "`tanh` by default.") + .SetDefault("tanh") + .InEnum({"sigmoid", "tanh", "relu", "identity"}); + AddComment(R"DOC( +Fusion Long-Short Term Memory (LSTM) Operator. +This operator fuse the X into LSTM, more details can refer to LSTM op. +)DOC"); +} + +template +inline void ReorderInitState(const DeviceContext& ctx, + const framework::Tensor& src, + framework::Vector index_lod, + framework::Tensor* dst, bool indexed_src) { + math::CopyMatrixRowsFunctor row_shuffle; + dst->mutable_data(src.dims(), ctx.GetPlace()); + // TODO(TJ): check mem copy perf + row_shuffle(ctx, src, index_lod, dst, indexed_src); +} + +template +class FuisonLSTMKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& ctx) const override { + auto* x = ctx.Input("X"); + auto* wx = ctx.Input("WeightX"); + auto* wh = ctx.Input("WeightH"); + auto* bias = ctx.Input("Bias"); + auto* hidden_t0 = ctx.Input("H0"); + auto* cell_t0 = ctx.Input("C0"); + + auto* xx = ctx.Output("XX"); + auto* batched_gate = ctx.Output("BatchedGate"); + auto* hidden_out = ctx.Output("Hidden"); + auto* cell_out = ctx.Output("Cell"); + bool is_reverse = ctx.Attr("is_reverse"); + + T* xx_data = xx->mutable_data(ctx.GetPlace()); + T* batched_gate_data = batched_gate->mutable_data(ctx.GetPlace()); + hidden_out->mutable_data(ctx.GetPlace()); + cell_out->mutable_data(ctx.GetPlace()); + + const T* x_data = x->data(); + const T* wx_data = wx->data(); + auto x_dims = x->dims(); + auto wx_dims = wx->dims(); + + math::LoDTensor2BatchFunctor to_batch; + auto& dev_ctx = ctx.template device_context(); + auto blas = math::GetBlas(dev_ctx); + if (x_dims[1] > wx_dims[1]) { + math::FCCompute(blas, x_dims[0], wx_dims[1], x_dims[1], + x_data, wx_data, xx_data, + bias->data()); + to_batch(dev_ctx, *xx, batched_gate, true, is_reverse); + } else { + to_batch(dev_ctx, *x, xx, true, is_reverse); + batched_gate->set_lod(xx->lod()); + math::FCCompute(blas, x_dims[0], wx_dims[1], x_dims[1], + xx_data, wx_data, batched_gate_data, + bias->data()); + } + + int frame_size = static_cast(wx_dims[1] / 4); + framework::DDim out_dims({x_dims[0], frame_size}); + math::LstmMetaValue lstm_value; + // no peephole + lstm_value.check_ig = nullptr; + lstm_value.check_fg = nullptr; + lstm_value.check_og = nullptr; + lstm_value.prev_state_value = nullptr; + Tensor ordered_c0; + + framework::Vector order(batched_gate->lod()[2]); + + if (cell_t0) { + // Since the batch computing for LSTM reorders the input sequence + // according to their length. The initialized cell state also needs + // to reorder. + ReorderInitState(dev_ctx, *cell_t0, order, &ordered_c0, + true); + lstm_value.prev_state_value = ordered_c0.data(); + } + + // Use the local variable as here. + LoDTensor batch_hidden, batch_cell; + auto* batch_cell_pre_act = ctx.Output("BatchCellPreAct"); + batch_hidden.mutable_data(out_dims, ctx.GetPlace()); + batch_cell.mutable_data(out_dims, ctx.GetPlace()); + batch_cell_pre_act->mutable_data(out_dims, ctx.GetPlace()); + + auto batch_starts = batched_gate->lod()[0]; + size_t max_seq_len = batch_starts.size() - 1; + auto gate_act = math::detail::GetActivationType( + ctx.Attr("gate_activation")); + auto cell_act = math::detail::GetActivationType( + ctx.Attr("cell_activation")); + auto cand_act = math::detail::GetActivationType( + ctx.Attr("candidate_activation")); + + for (size_t n = 0; n < max_seq_len; n++) { + int bstart = static_cast(batch_starts[n]); + int bend = static_cast(batch_starts[n + 1]); + + Tensor gate_t = batched_gate->Slice(bstart, bend); + Tensor out_t = batch_hidden.Slice(bstart, bend); + Tensor cell_t = batch_cell.Slice(bstart, bend); + Tensor cell_pre_act_t = batch_cell_pre_act->Slice(bstart, bend); + + int cur_batch_size = bend - bstart; + + if (n > 0) { + int pre_h_start = static_cast(batch_starts[n - 1]); + int pre_h_end = pre_h_start + cur_batch_size; + auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end); + // TODO(TJ): use gemm directly + blas.MatMul(pre_hidden_t, false, *wh, false, static_cast(1.0), + &gate_t, static_cast(1.0)); + } else if (hidden_t0) { + // TODO(TJ): move h0 outside for + // If n == 0 and there is no initialized hidden state, that is to say + // the H0 is zeros, the calculation W_h * H0 will be skiped. + // If n == 0 and there is initialized hidden state, calculate W_h * H0. + + // Since the batch computing for LSTM reorders the input sequence + // according to their length. The initialized hidden state also needs + // to reorder. + Tensor ordered_h0; + ReorderInitState(dev_ctx, *hidden_t0, order, + &ordered_h0, true); + // TODO(TJ): use gemm directly + blas.MatMul(ordered_h0, false, *wh, false, static_cast(1.0), &gate_t, + static_cast(1.0)); + } + + lstm_value.gate_value = gate_t.data(); + lstm_value.output_value = out_t.data(); + lstm_value.state_value = cell_t.data(); + lstm_value.state_active_value = cell_pre_act_t.data(); + math::LstmUnitFunctor::compute( + dev_ctx, lstm_value, frame_size, cur_batch_size, gate_act, cell_act, + cand_act); + lstm_value.prev_state_value = lstm_value.state_value; + } + + math::Batch2LoDTensorFunctor to_seq; + batch_hidden.set_lod(batched_gate->lod()); + // restore the output hidden in LoDTensor from the batch hidden + to_seq(dev_ctx, batch_hidden, hidden_out); + + batch_cell.set_lod(batched_gate->lod()); + // restore the output cell state in LoDTensor from the batch cell + to_seq(dev_ctx, batch_cell, cell_out); + } +}; + +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(fusion_lstm, ops::FusionLSTMOp, ops::FusionLSTMOpMaker, + paddle::framework::DefaultGradOpDescMaker); + +REGISTER_OP_CPU_KERNEL( + fusion_lstm, + ops::FuisonLSTMKernel, + ops::FuisonLSTMKernel); diff --git a/paddle/fluid/operators/fusion_lstm_op.h b/paddle/fluid/operators/fusion_lstm_op.h new file mode 100644 index 0000000000..39dc09b4d1 --- /dev/null +++ b/paddle/fluid/operators/fusion_lstm_op.h @@ -0,0 +1,42 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + +http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once +// #include +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using LoDTensor = framework::LoDTensor; +using Tensor = framework::Tensor; + +class FusionLSTMOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override; + + protected: + framework::OpKernelType GetExpectedKernelType( + const framework::ExecutionContext& ctx) const override; +}; + +class FusionLSTMOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override; +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/gaussian_random_op.cu b/paddle/fluid/operators/gaussian_random_op.cu index b490723795..7784856417 100644 --- a/paddle/fluid/operators/gaussian_random_op.cu +++ b/paddle/fluid/operators/gaussian_random_op.cu @@ -15,7 +15,6 @@ limitations under the License. */ #include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" -#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -61,7 +60,6 @@ class GPUGaussianRandomKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(gaussian_random, paddle::operators::GPUGaussianRandomKernel, paddle::operators::GPUGaussianRandomKernel); diff --git a/paddle/fluid/operators/gru_op.cc b/paddle/fluid/operators/gru_op.cc index 5c74687882..087f903a8b 100644 --- a/paddle/fluid/operators/gru_op.cc +++ b/paddle/fluid/operators/gru_op.cc @@ -14,6 +14,11 @@ limitations under the License. */ #include "paddle/fluid/operators/gru_op.h" #include +#include "paddle/fluid/operators/math/blas.h" +#include "paddle/fluid/operators/math/detail/gru_cpu_kernel.h" +#include "paddle/fluid/operators/math/detail/gru_kernel.h" + +DECLARE_int32(paddle_num_threads); namespace paddle { namespace operators { @@ -211,6 +216,158 @@ class GRUGradOp : public framework::OperatorWithKernel { } }; +template +class GRUCPUKernel : public framework::OpKernel { + public: + void BatchCompute(const framework::ExecutionContext& context) const { + using DeviceContext = paddle::platform::CPUDeviceContext; + auto* input = context.Input("Input"); + auto* h0 = context.Input("H0"); + auto* weight = context.Input("Weight"); + const T* weight_data = weight->data(); + auto* bias = context.Input("Bias"); + auto* batch_gate = context.Output("BatchGate"); + batch_gate->mutable_data(context.GetPlace()); + auto* batch_reset_hidden_prev = + context.Output("BatchResetHiddenPrev"); + batch_reset_hidden_prev->mutable_data(context.GetPlace()); + auto* batch_hidden = context.Output("BatchHidden"); + batch_hidden->mutable_data(context.GetPlace()); + auto* hidden = context.Output("Hidden"); + hidden->mutable_data(context.GetPlace()); + + auto hidden_dims = hidden->dims(); + + bool is_reverse = context.Attr("is_reverse"); + math::LoDTensor2BatchFunctor to_batch; + auto& dev_ctx = context.template device_context(); + to_batch(dev_ctx, *input, batch_gate, true, is_reverse); + + if (bias) { + math::RowwiseAdd add_bias; + add_bias(dev_ctx, *batch_gate, *bias, batch_gate); + } + + int frame_size = hidden_dims[1]; + math::GRUMetaValue gru_value; + gru_value.gate_weight = const_cast(weight_data); + gru_value.state_weight = + const_cast(weight_data + 2 * frame_size * frame_size); + Tensor ordered_h0; + + framework::Vector order(batch_gate->lod()[2]); + + if (h0) { + // Since the batch computing for GRU reorders the input sequences + // according to their length. The initialized cell state also needs + // to reorder. + ReorderInitState( + context.template device_context(), *h0, order, + &ordered_h0, true); + gru_value.prev_out_value = ordered_h0.data(); + } else { + gru_value.prev_out_value = nullptr; + } + auto batch_starts = batch_gate->lod()[0]; + size_t seq_len = batch_starts.size() - 1; + auto active_node = math::detail::GetActivationType( + context.Attr("activation")); + auto active_gate = math::detail::GetActivationType( + context.Attr("gate_activation")); + +#ifdef PADDLE_WITH_MKLML + // use MKL packed to speedup GEMM + if (FLAGS_paddle_num_threads >= 4) { + auto blas = math::GetBlas(dev_ctx); + T* packed_gate = blas.GEMM_ALLOC(CblasBMatrix, 1 /*height of C*/, + frame_size * 2 /*width of weight*/, + frame_size /*height of height*/); + PADDLE_ENFORCE(packed_gate); + blas.GEMM_PACK(CblasBMatrix, CblasNoTrans, 1 /*cur bs?*/, frame_size * 2, + frame_size, T(1.0), gru_value.gate_weight, frame_size * 2, + packed_gate); + T* packed_state = blas.GEMM_ALLOC(CblasBMatrix, 1 /*height of C*/, + frame_size /*width of weight*/, + frame_size /*height of height*/); + PADDLE_ENFORCE(packed_state); + blas.GEMM_PACK(CblasBMatrix, CblasNoTrans, 1 /*cur bs?*/, frame_size, + frame_size, T(1.0), gru_value.state_weight, frame_size, + packed_state); + for (size_t n = 0; n < seq_len; n++) { + int bstart = static_cast(batch_starts[n]); + int bend = static_cast(batch_starts[n + 1]); + int cur_batch_size = bend - bstart; + + Tensor gate_t = batch_gate->Slice(bstart, bend); + Tensor reset_hidden_prev_t = + batch_reset_hidden_prev->Slice(bstart, bend); + Tensor hidden_t = batch_hidden->Slice(bstart, bend); + gru_value.output_value = hidden_t.data(); + gru_value.gate_value = gate_t.data(); + gru_value.reset_output_value = reset_hidden_prev_t.data(); + + if (gru_value.prev_out_value) { + blas.GEMM_COMPUTE( + CblasNoTrans, CblasPacked, cur_batch_size, frame_size * 2, + frame_size, gru_value.prev_out_value, frame_size, packed_gate, + frame_size * 2, T(1), gru_value.gate_value, frame_size * 3); + } + + math::detail::forward_reset_output( + math::detail::forward::gru_resetOutput(), gru_value, frame_size, + cur_batch_size, active_gate); + + if (gru_value.prev_out_value) { + blas.GEMM_COMPUTE( + CblasNoTrans, CblasPacked, cur_batch_size, frame_size, frame_size, + gru_value.reset_output_value, frame_size, packed_state, + frame_size, T(1), gru_value.gate_value + frame_size * 2, + frame_size * 3); + } + + math::detail::forward_final_output( + math::detail::forward::gru_finalOutput(), gru_value, frame_size, + cur_batch_size, active_node); + + gru_value.prev_out_value = gru_value.output_value; + } + + blas.GEMM_FREE(packed_gate); + blas.GEMM_FREE(packed_state); + } else { +#endif + for (size_t n = 0; n < seq_len; n++) { + int bstart = static_cast(batch_starts[n]); + int bend = static_cast(batch_starts[n + 1]); + int cur_batch_size = bend - bstart; + + Tensor gate_t = batch_gate->Slice(bstart, bend); + Tensor reset_hidden_prev_t = + batch_reset_hidden_prev->Slice(bstart, bend); + Tensor hidden_t = batch_hidden->Slice(bstart, bend); + gru_value.output_value = hidden_t.data(); + gru_value.gate_value = gate_t.data(); + gru_value.reset_output_value = reset_hidden_prev_t.data(); + + math::GRUUnitFunctor::compute( + dev_ctx, gru_value, frame_size, cur_batch_size, active_node, + active_gate); + + gru_value.prev_out_value = gru_value.output_value; + } +#ifdef PADDLE_WITH_MKLML + } +#endif + math::Batch2LoDTensorFunctor to_seq; + batch_hidden->set_lod(batch_gate->lod()); + to_seq(dev_ctx, *batch_hidden, hidden); + } + + void Compute(const framework::ExecutionContext& context) const override { + BatchCompute(context); + } +}; + } // namespace operators } // namespace paddle @@ -218,9 +375,8 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(gru, ops::GRUOp, ops::GRUOpMaker, paddle::framework::DefaultGradOpDescMaker); REGISTER_OPERATOR(gru_grad, ops::GRUGradOp); -REGISTER_OP_CPU_KERNEL( - gru, ops::GRUKernel, - ops::GRUKernel); +REGISTER_OP_CPU_KERNEL(gru, ops::GRUCPUKernel, + ops::GRUCPUKernel); REGISTER_OP_CPU_KERNEL( gru_grad, ops::GRUGradKernel, ops::GRUGradKernel); diff --git a/paddle/fluid/operators/gru_op.cu.cc b/paddle/fluid/operators/gru_op.cu.cc index baf455a840..55721c283d 100644 --- a/paddle/fluid/operators/gru_op.cu.cc +++ b/paddle/fluid/operators/gru_op.cu.cc @@ -14,6 +14,96 @@ limitations under the License. */ #include "paddle/fluid/operators/gru_op.h" +namespace paddle { +namespace operators { + +template +class GRUKernel : public framework::OpKernel { + public: + void BatchCompute(const framework::ExecutionContext& context) const { + auto* input = context.Input("Input"); + auto* h0 = context.Input("H0"); + auto* weight = context.Input("Weight"); + const T* weight_data = weight->data(); + auto* bias = context.Input("Bias"); + auto* batch_gate = context.Output("BatchGate"); + batch_gate->mutable_data(context.GetPlace()); + auto* batch_reset_hidden_prev = + context.Output("BatchResetHiddenPrev"); + batch_reset_hidden_prev->mutable_data(context.GetPlace()); + auto* batch_hidden = context.Output("BatchHidden"); + batch_hidden->mutable_data(context.GetPlace()); + auto* hidden = context.Output("Hidden"); + hidden->mutable_data(context.GetPlace()); + + auto hidden_dims = hidden->dims(); + + bool is_reverse = context.Attr("is_reverse"); + math::LoDTensor2BatchFunctor to_batch; + auto& dev_ctx = context.template device_context(); + to_batch(dev_ctx, *input, batch_gate, true, is_reverse); + + if (bias) { + math::RowwiseAdd add_bias; + add_bias(dev_ctx, *batch_gate, *bias, batch_gate); + } + + int frame_size = hidden_dims[1]; + math::GRUMetaValue gru_value; + gru_value.gate_weight = const_cast(weight_data); + gru_value.state_weight = + const_cast(weight_data + 2 * frame_size * frame_size); + Tensor ordered_h0; + + framework::Vector order(batch_gate->lod()[2]); + + if (h0) { + // Since the batch computing for GRU reorders the input sequences + // according to their length. The initialized cell state also needs + // to reorder. + ReorderInitState( + context.template device_context(), *h0, order, + &ordered_h0, true); + gru_value.prev_out_value = ordered_h0.data(); + } else { + gru_value.prev_out_value = nullptr; + } + auto batch_starts = batch_gate->lod()[0]; + size_t num_batch = batch_starts.size() - 1; + auto active_node = math::detail::GetActivationType( + context.Attr("activation")); + auto active_gate = math::detail::GetActivationType( + context.Attr("gate_activation")); + for (size_t n = 0; n < num_batch; n++) { + int bstart = static_cast(batch_starts[n]); + int bend = static_cast(batch_starts[n + 1]); + int cur_batch_size = bend - bstart; + + Tensor gate_t = batch_gate->Slice(bstart, bend); + Tensor reset_hidden_prev_t = batch_reset_hidden_prev->Slice(bstart, bend); + Tensor hidden_t = batch_hidden->Slice(bstart, bend); + gru_value.output_value = hidden_t.data(); + gru_value.gate_value = gate_t.data(); + gru_value.reset_output_value = reset_hidden_prev_t.data(); + math::GRUUnitFunctor::compute( + dev_ctx, gru_value, frame_size, cur_batch_size, active_node, + active_gate); + gru_value.prev_out_value = gru_value.output_value; + } + + math::Batch2LoDTensorFunctor to_seq; + batch_hidden->set_lod(batch_gate->lod()); + to_seq(dev_ctx, *batch_hidden, hidden); + } + + void Compute(const framework::ExecutionContext& context) const override { + BatchCompute(context); + } +}; + +} // namespace operators +} // namespace paddle + namespace ops = paddle::operators; REGISTER_OP_CUDA_KERNEL( gru, ops::GRUKernel, diff --git a/paddle/fluid/operators/gru_op.h b/paddle/fluid/operators/gru_op.h index 3b0d93e54b..0b551e8046 100644 --- a/paddle/fluid/operators/gru_op.h +++ b/paddle/fluid/operators/gru_op.h @@ -37,90 +37,6 @@ inline void ReorderInitState(const DeviceContext& ctx, row_shuffle(ctx, src, index_lod, dst, indexed_src); } -template -class GRUKernel : public framework::OpKernel { - public: - void BatchCompute(const framework::ExecutionContext& context) const { - auto* input = context.Input("Input"); - auto* h0 = context.Input("H0"); - auto* weight = context.Input("Weight"); - const T* weight_data = weight->data(); - auto* bias = context.Input("Bias"); - auto* batch_gate = context.Output("BatchGate"); - batch_gate->mutable_data(context.GetPlace()); - auto* batch_reset_hidden_prev = - context.Output("BatchResetHiddenPrev"); - batch_reset_hidden_prev->mutable_data(context.GetPlace()); - auto* batch_hidden = context.Output("BatchHidden"); - batch_hidden->mutable_data(context.GetPlace()); - auto* hidden = context.Output("Hidden"); - hidden->mutable_data(context.GetPlace()); - - auto hidden_dims = hidden->dims(); - - bool is_reverse = context.Attr("is_reverse"); - math::LoDTensor2BatchFunctor to_batch; - auto& dev_ctx = context.template device_context(); - to_batch(dev_ctx, *input, batch_gate, true, is_reverse); - - if (bias) { - math::RowwiseAdd add_bias; - add_bias(dev_ctx, *batch_gate, *bias, batch_gate); - } - - int frame_size = hidden_dims[1]; - math::GRUMetaValue gru_value; - gru_value.gate_weight = const_cast(weight_data); - gru_value.state_weight = - const_cast(weight_data + 2 * frame_size * frame_size); - Tensor ordered_h0; - - framework::Vector order(batch_gate->lod()[2]); - - if (h0) { - // Since the batch computing for GRU reorders the input sequences - // according to their length. The initialized cell state also needs - // to reorder. - ReorderInitState( - context.template device_context(), *h0, order, - &ordered_h0, true); - gru_value.prev_out_value = ordered_h0.data(); - } else { - gru_value.prev_out_value = nullptr; - } - auto batch_starts = batch_gate->lod()[0]; - size_t num_batch = batch_starts.size() - 1; - auto active_node = math::detail::GetActivationType( - context.Attr("activation")); - auto active_gate = math::detail::GetActivationType( - context.Attr("gate_activation")); - for (size_t n = 0; n < num_batch; n++) { - int bstart = static_cast(batch_starts[n]); - int bend = static_cast(batch_starts[n + 1]); - int cur_batch_size = bend - bstart; - - Tensor gate_t = batch_gate->Slice(bstart, bend); - Tensor reset_hidden_prev_t = batch_reset_hidden_prev->Slice(bstart, bend); - Tensor hidden_t = batch_hidden->Slice(bstart, bend); - gru_value.output_value = hidden_t.data(); - gru_value.gate_value = gate_t.data(); - gru_value.reset_output_value = reset_hidden_prev_t.data(); - math::GRUUnitFunctor::compute( - dev_ctx, gru_value, frame_size, cur_batch_size, active_node, - active_gate); - gru_value.prev_out_value = gru_value.output_value; - } - - math::Batch2LoDTensorFunctor to_seq; - batch_hidden->set_lod(batch_gate->lod()); - to_seq(dev_ctx, *batch_hidden, hidden); - } - - void Compute(const framework::ExecutionContext& context) const override { - BatchCompute(context); - } -}; - template class GRUGradKernel : public framework::OpKernel { public: diff --git a/paddle/fluid/operators/load_op.cc b/paddle/fluid/operators/load_op.cc index 27e26cb1b5..51219504ff 100644 --- a/paddle/fluid/operators/load_op.cc +++ b/paddle/fluid/operators/load_op.cc @@ -92,6 +92,7 @@ class LoadOp : public framework::OperatorBase { platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); auto &dev_ctx = *pool.Get(place); framework::DeserializeFromStream(fin, selectedRows, dev_ctx); + selectedRows->SyncIndex(); } }; diff --git a/paddle/fluid/operators/lookup_sparse_table_op.cc b/paddle/fluid/operators/lookup_sparse_table_op.cc index 2ce11e712f..de3f0990e1 100644 --- a/paddle/fluid/operators/lookup_sparse_table_op.cc +++ b/paddle/fluid/operators/lookup_sparse_table_op.cc @@ -17,7 +17,6 @@ limitations under the License. */ #include "paddle/fluid/framework/data_type.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/math/math_function.h" -#include "paddle/fluid/platform/device_context.h" namespace paddle { namespace operators { @@ -46,10 +45,6 @@ class LookupSparseTableOp : public framework::OperatorBase { auto out_var = scope.FindVar(Output("Out")); auto w_var = scope.FindVar(Input("W")); auto ids_var = scope.FindVar(Input("Ids")); - unsigned int seed = static_cast(Attr("seed")); - float min = Attr("min"); - float max = Attr("max"); - bool auto_grown_table = Attr("auto_grown_table"); PADDLE_ENFORCE(out_var->IsType(), "The type of Out var should be LodTensor."); @@ -60,46 +55,17 @@ class LookupSparseTableOp : public framework::OperatorBase { auto &ids_t = ids_var->Get(); auto out_t = out_var->GetMutable(); auto w_t = w_var->GetMutable(); - std::vector keys; - keys.resize(ids_t.numel()); - for (int64_t i = 0; i < ids_t.numel(); ++i) { - keys[i] = ids_t.data()[i]; - } // TODO(Yancey1989): support CUDA Place for the sparse table platform::CPUPlace cpu; auto out_shape = w_t->value().dims(); - out_shape[0] = keys.size(); + out_shape[0] = ids_t.numel(); out_t->Resize(out_shape); out_t->mutable_data(cpu, w_t->value().type()); PADDLE_ENFORCE_EQ(framework::ToDataType(w_t->value().type()), framework::proto::VarType::FP32, "The sparse table only support FP32"); - auto non_keys_pair = w_t->Get(keys, out_t); - if (!auto_grown_table) { - PADDLE_ENFORCE_EQ(non_keys_pair.size(), static_cast(0), - "there is some keys does exists in the sparse table."); - } - auto value_shape = w_t->value().dims(); - value_shape[0] = 1; - for (const auto &it : non_keys_pair) { - const auto key = it.first; - const auto index = it.second; - framework::Tensor value; - value.Resize(value_shape); - auto data = value.mutable_data(cpu); - - std::minstd_rand engine; - engine.seed(seed); - std::uniform_real_distribution dist(min, max); - int64_t size = value.numel(); - for (int64_t i = 0; i < size; ++i) { - data[i] = dist(engine); - } - w_t->Set(key, value); - memory::Copy(cpu, out_t->mutable_data(cpu) + index * value.numel(), - cpu, value.data(), value.numel() * sizeof(float)); - } + w_t->Get(ids_t, out_t, true); } }; @@ -121,21 +87,6 @@ class LookupSparseTableOpMaker : public framework::OpProtoAndCheckerMaker { "Otherwise the given value indicates padding the output " "with zeros whenever lookup encounters it in Ids.") .SetDefault(kNoPadding); - AddAttr("min", - "(float, default -1.0) " - "Minimum value of uniform random") - .SetDefault(-1.0f); - AddAttr("max", - "(float, default 1.0) " - "Maximum value of uniform random") - .SetDefault(1.0f); - AddAttr("seed", - "(int, default 0) " - "Random seed used for generating samples. " - "0 means use a seed generated by the system." - "Note that if seed is not 0, this operator will always " - "generate the same random numbers every time.") - .SetDefault(0); AddAttr("auto_grown_table", "(bool default false)" "Whether create new value if for nonexistent key.") diff --git a/paddle/fluid/operators/math/blas.h b/paddle/fluid/operators/math/blas.h index 70f88f24f6..8dcf7c99f3 100644 --- a/paddle/fluid/operators/math/blas.h +++ b/paddle/fluid/operators/math/blas.h @@ -90,6 +90,25 @@ class Blas { void GEMM(bool transA, bool transB, int M, int N, int K, T alpha, const T* A, int lda, const T* B, int ldb, T beta, T* C, int ldc) const; +#ifdef PADDLE_WITH_MKLML + template + T* GEMM_ALLOC(const CBLAS_IDENTIFIER id, const int M, const int N, + const int K) const; + + template + void GEMM_PACK(const CBLAS_IDENTIFIER id, const CBLAS_TRANSPOSE trans, int M, + int N, int K, const T alpha, const T* src, const int ld, + T* dst) const; + + template + void GEMM_COMPUTE(int transA, int transB, int M, int N, int K, const T* A, + const int lda, const T* B, const int ldb, T beta, T* C, + const int ldc) const; + + template + void GEMM_FREE(T* data) const; +#endif + template void MatMul(const framework::Tensor& mat_a, bool trans_a, const framework::Tensor& mat_b, bool trans_b, T alpha, @@ -115,6 +134,9 @@ class Blas { template void VADD(int n, const T* x, const T* y, T* z) const; + template + void VMUL(int n, const T* x, const T* y, T* z) const; + template void VCOPY(int n, const T* x, T* y) const; @@ -146,6 +168,28 @@ class BlasT : private Blas { Base()->template GEMM(args...); } +#ifdef PADDLE_WITH_MKLML + template + T* GEMM_ALLOC(ARGS... args) const { + return Base()->template GEMM_ALLOC(args...); + } + + template + void GEMM_PACK(ARGS... args) const { + Base()->template GEMM_PACK(args...); + } + + template + void GEMM_COMPUTE(ARGS... args) const { + Base()->template GEMM_COMPUTE(args...); + } + + template + void GEMM_FREE(ARGS... args) const { + Base()->template GEMM_FREE(args...); + } +#endif + template void MatMul(ARGS... args) const { Base()->template MatMul(args...); @@ -161,6 +205,11 @@ class BlasT : private Blas { Base()->template VADD(args...); } + template + void VMUL(ARGS... args) const { + Base()->template VMUL(args...); + } + template void VCOPY(ARGS... args) const { Base()->template VCOPY(args...); diff --git a/paddle/fluid/operators/math/blas_impl.h b/paddle/fluid/operators/math/blas_impl.h index a0802ef90c..dc77b6d793 100644 --- a/paddle/fluid/operators/math/blas_impl.h +++ b/paddle/fluid/operators/math/blas_impl.h @@ -31,6 +31,26 @@ struct CBlas { platform::dynload::cblas_sgemm(args...); } + template + static float *GEMM_ALLOC(ARGS... args) { + return platform::dynload::cblas_sgemm_alloc(args...); + } + + template + static void GEMM_PACK(ARGS... args) { + platform::dynload::cblas_sgemm_pack(args...); + } + + template + static void GEMM_COMPUTE(ARGS... args) { + platform::dynload::cblas_sgemm_compute(args...); + } + + template + static void GEMM_FREE(ARGS... args) { + platform::dynload::cblas_sgemm_free(args...); + } + #ifdef PADDLE_WITH_LIBXSMM template static void SMM_GEMM(ARGS... args) { @@ -62,6 +82,11 @@ struct CBlas { static void VADD(ARGS... args) { platform::dynload::vsAdd(args...); } + + template + static void VMUL(ARGS... args) { + platform::dynload::vsMul(args...); + } }; template <> @@ -71,6 +96,26 @@ struct CBlas { platform::dynload::cblas_dgemm(args...); } + template + static double *GEMM_ALLOC(ARGS... args) { + return platform::dynload::cblas_dgemm_alloc(args...); + } + + template + static void GEMM_PACK(ARGS... args) { + platform::dynload::cblas_dgemm_pack(args...); + } + + template + static void GEMM_COMPUTE(ARGS... args) { + platform::dynload::cblas_dgemm_compute(args...); + } + + template + static void GEMM_FREE(ARGS... args) { + platform::dynload::cblas_dgemm_free(args...); + } + #ifdef PADDLE_WITH_LIBXSMM template static void SMM_GEMM(ARGS... args) { @@ -102,6 +147,11 @@ struct CBlas { static void VADD(ARGS... args) { platform::dynload::vdAdd(args...); } + + template + static void VMUL(ARGS... args) { + platform::dynload::vdMul(args...); + } }; #else @@ -159,6 +209,7 @@ struct CBlas { static void SMM_GEMM(...) { PADDLE_THROW("float16 SMM_GEMM not supported on CPU"); } + static void VMUL(...) { PADDLE_THROW("float16 VMUL not supported on CPU"); } #ifdef PADDLE_WITH_MKLML static void GEMM_BATCH(...) { PADDLE_THROW("float16 GEMM_BATCH not supported on CPU"); @@ -224,6 +275,41 @@ inline void GEMM_WARP(CBLAS_ORDER order, CBLAS_TRANSPOSE transA, beta, C, ldc); } +#ifdef PADDLE_WITH_MKLML +template <> +template +T *Blas::GEMM_ALLOC(const CBLAS_IDENTIFIER id, + const int M, const int N, + const int K) const { + return CBlas::GEMM_ALLOC(id, M, N, K); +} + +template <> +template +void Blas::GEMM_PACK(const CBLAS_IDENTIFIER id, + const CBLAS_TRANSPOSE trans, + int M, int N, int K, + const T alpha, const T *src, + const int ld, T *dst) const { + CBlas::GEMM_PACK(CblasRowMajor, id, trans, M, N, K, alpha, src, ld, dst); +} + +template <> +template +void Blas::GEMM_COMPUTE( + int transA, int transB, int M, int N, int K, const T *A, const int lda, + const T *B, const int ldb, T beta, T *C, const int ldc) const { + CBlas::GEMM_COMPUTE(CblasRowMajor, transA, transB, M, N, K, A, lda, B, ldb, + beta, C, ldc); +} + +template <> +template +void Blas::GEMM_FREE(T *data) const { + CBlas::GEMM_FREE(data); +} +#endif + template <> template void Blas::GEMM(CBLAS_TRANSPOSE transA, @@ -299,6 +385,20 @@ void Blas::VADD(int n, const T *x, const T *y, #endif } +template <> +template +void Blas::VMUL(int n, const T *x, const T *y, + T *z) const { +#ifdef PADDLE_WITH_MKLML + CBlas::VMUL(n, x, y, z); +#else + // try to find if openblas support vmul + for (int i = 0; i < n; ++i) { + z[i] = x[i] * y[i]; + } +#endif +} + template <> template void Blas::GEMV(bool trans_a, int M, int N, T alpha, diff --git a/paddle/fluid/operators/math/cross_entropy.cu b/paddle/fluid/operators/math/cross_entropy.cu index 58b85abf82..0de58d5fdd 100644 --- a/paddle/fluid/operators/math/cross_entropy.cu +++ b/paddle/fluid/operators/math/cross_entropy.cu @@ -15,25 +15,11 @@ limitations under the License. */ #include "paddle/fluid/operators/math/cross_entropy.h" #include "paddle/fluid/platform/cuda_device_function.h" #include "paddle/fluid/platform/cuda_primitives.h" -#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { namespace math { -template -HOSTDEVICE T log(const T& val) { - return std::log(val); -} - -template <> -HOSTDEVICE platform::float16 log(const platform::float16& val) { - // strage bug, hlog is not exists. - return static_cast(0); - // half tmp = static_cast(val); - // return static_cast(hlog(tmp)); -} - namespace { template __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label, @@ -49,12 +35,12 @@ template __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label, const int class_num) { int tid = threadIdx.x; - T val(0); + T val = 0; int idx = blockIdx.x * class_num + tid; int end = blockIdx.x * class_num + class_num; for (; idx < end; idx += blockDim.x) { - val += math::TolerableValue()(log(X[idx])) * label[idx]; + val += math::TolerableValue()(std::log(X[idx])) * label[idx]; } val = paddle::platform::reduceSum(val, tid, blockDim.x); @@ -98,8 +84,6 @@ class CrossEntropyFunctor { template class CrossEntropyFunctor; template class CrossEntropyFunctor; -template class CrossEntropyFunctor; } // namespace math } // namespace operators } // namespace paddle diff --git a/paddle/fluid/operators/math/cross_entropy.h b/paddle/fluid/operators/math/cross_entropy.h index 2e4e4781c2..adc5b3fe47 100644 --- a/paddle/fluid/operators/math/cross_entropy.h +++ b/paddle/fluid/operators/math/cross_entropy.h @@ -13,10 +13,8 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include #include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/tensor.h" -#include "paddle/fluid/platform/float16.h" #include "paddle/fluid/platform/hostdevice.h" namespace paddle { @@ -35,21 +33,6 @@ struct TolerableValue { } }; -// float16 value clip behave different. -using paddle::platform::float16; -using paddle::platform::isfinite; -template <> -struct TolerableValue { - HOSTDEVICE float16 operator()(const float16& x) const { - if (isfinite(x)) - return x; - else if (x > static_cast(0)) - return std::numeric_limits::max(); - else - return std::numeric_limits::min(); - } -}; - template class CrossEntropyFunctor { public: diff --git a/paddle/fluid/operators/math/fc_compute.h b/paddle/fluid/operators/math/fc_compute.h new file mode 100644 index 0000000000..8600fa9e2c --- /dev/null +++ b/paddle/fluid/operators/math/fc_compute.h @@ -0,0 +1,43 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#pragma once + +#include "paddle/fluid/operators/math/blas.h" + +DECLARE_int32(paddle_num_threads); + +namespace paddle { +namespace operators { +namespace math { + +template +inline void FCCompute(const BlasT& blas, const int M, + const int N, const int K, const T* X, const T* W, T* Y, + const T* B = NULL) { + blas.GEMM(CblasNoTrans, CblasNoTrans, M, N, K, static_cast(1), X, W, + static_cast(0), Y); + if (B) { +#ifdef PADDLE_WITH_MKLML +#pragma omp parallel for if (FLAGS_paddle_num_threads > 1) +#endif + for (int i = 0; i < M; i++) { + blas.AXPY(N, static_cast(1), B, Y + i * N); + } + } +} + +} // namespace math +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/math/selected_rows_functor.cu b/paddle/fluid/operators/math/selected_rows_functor.cu index 00dbfc11a2..a92762c7fe 100644 --- a/paddle/fluid/operators/math/selected_rows_functor.cu +++ b/paddle/fluid/operators/math/selected_rows_functor.cu @@ -18,7 +18,6 @@ limitations under the License. */ #include "paddle/fluid/operators/math/math_function.h" #include "paddle/fluid/operators/math/selected_rows_functor.h" #include "paddle/fluid/platform/cuda_primitives.h" -#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { @@ -77,7 +76,6 @@ struct SelectedRowsAdd { template struct SelectedRowsAdd; template struct SelectedRowsAdd; -template struct SelectedRowsAdd; namespace { template @@ -122,7 +120,7 @@ struct SelectedRowsAddTensor { auto* out_data = output->data(); SetConstant functor; - functor(context, output, static_cast(0)); + functor(context, output, 0.0); const int block_size = 256; dim3 threads(block_size, 1); @@ -140,8 +138,6 @@ struct SelectedRowsAddTensor { template struct SelectedRowsAddTensor; template struct SelectedRowsAddTensor; -template struct SelectedRowsAddTensor; template struct SelectedRowsAddTo { @@ -181,8 +177,6 @@ template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; template struct SelectedRowsAddTo; -template struct SelectedRowsAddTo; namespace { template @@ -235,8 +229,6 @@ template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; template struct SelectedRowsAddToTensor; -template struct SelectedRowsAddToTensor; namespace scatter { @@ -284,7 +276,7 @@ struct MergeAdd { context.GetPlace()); math::SetConstant constant_functor; - constant_functor(context, out.mutable_value(), static_cast(0)); + constant_functor(context, out.mutable_value(), 0.0); auto* out_data = out.mutable_value()->data(); auto* input_data = input.value().data(); @@ -308,7 +300,6 @@ template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; template struct MergeAdd; -template struct MergeAdd; template __global__ void UpdateToTensorKernel(const T* selected_rows, diff --git a/paddle/fluid/operators/math/softmax.cu b/paddle/fluid/operators/math/softmax.cu index 785c4baecb..3effe77625 100644 --- a/paddle/fluid/operators/math/softmax.cu +++ b/paddle/fluid/operators/math/softmax.cu @@ -94,15 +94,12 @@ void SoftmaxGradCUDNNFunctor::operator()( template class SoftmaxCUDNNFunctor; template class SoftmaxCUDNNFunctor; template class SoftmaxCUDNNFunctor; -template class SoftmaxGradCUDNNFunctor; template class SoftmaxGradCUDNNFunctor; template class SoftmaxGradCUDNNFunctor; template class SoftmaxFunctor; template class SoftmaxFunctor; template class SoftmaxFunctor; -template class SoftmaxGradFunctor; template class SoftmaxGradFunctor; template class SoftmaxGradFunctor; diff --git a/paddle/fluid/operators/mean_op.cu b/paddle/fluid/operators/mean_op.cu index 07aa23754f..91e0ab28ef 100644 --- a/paddle/fluid/operators/mean_op.cu +++ b/paddle/fluid/operators/mean_op.cu @@ -12,16 +12,14 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#define EIGEN_USE_GPU + #include "paddle/fluid/operators/mean_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( mean, ops::MeanKernel, - ops::MeanKernel, - ops::MeanKernel); + ops::MeanKernel); REGISTER_OP_CUDA_KERNEL( mean_grad, ops::MeanGradKernel, - ops::MeanGradKernel, - ops::MeanGradKernel); + ops::MeanGradKernel); diff --git a/paddle/fluid/operators/mean_op.h b/paddle/fluid/operators/mean_op.h index a41d50ae0b..362e9f9ae8 100644 --- a/paddle/fluid/operators/mean_op.h +++ b/paddle/fluid/operators/mean_op.h @@ -55,7 +55,7 @@ class MeanGradKernel : public framework::OpKernel { IG->mutable_data(context.GetPlace()); T ig_size = static_cast(IG->numel()); - Eigen::DSizes bcast(static_cast(ig_size)); + Eigen::DSizes bcast(ig_size); EigenVector::Flatten(*IG).device( *context.template device_context().eigen_device()) = diff --git a/paddle/fluid/operators/mul_op.cc b/paddle/fluid/operators/mul_op.cc index 51993398bd..2a8e4af516 100644 --- a/paddle/fluid/operators/mul_op.cc +++ b/paddle/fluid/operators/mul_op.cc @@ -54,9 +54,9 @@ class MulOp : public framework::OperatorWithKernel { auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims); auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims); - PADDLE_ENFORCE_EQ( - x_mat_dims[1], y_mat_dims[0], - "First matrix's width must be equal with second matrix's height."); + PADDLE_ENFORCE_EQ(x_mat_dims[1], y_mat_dims[0], + "First matrix's width must be equal with second matrix's " + "height. %s, %s"); std::vector output_dims; output_dims.reserve( static_cast(x_num_col_dims + y_dims.size() - y_num_col_dims)); diff --git a/paddle/fluid/operators/mul_op.cu.cc b/paddle/fluid/operators/mul_op.cu.cc index 6c5a83c6a5..81f3e42bf4 100644 --- a/paddle/fluid/operators/mul_op.cu.cc +++ b/paddle/fluid/operators/mul_op.cu.cc @@ -20,7 +20,6 @@ namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel, ops::MulKernel, ops::MulKernel); -REGISTER_OP_CUDA_KERNEL( - mul_grad, ops::MulGradKernel, - ops::MulGradKernel, - ops::MulGradKernel); +REGISTER_OP_CUDA_KERNEL(mul_grad, + ops::MulGradKernel, + ops::MulGradKernel); diff --git a/paddle/fluid/operators/nccl/CMakeLists.txt b/paddle/fluid/operators/nccl/CMakeLists.txt index ce0ddd89bf..cdcba80357 100644 --- a/paddle/fluid/operators/nccl/CMakeLists.txt +++ b/paddle/fluid/operators/nccl/CMakeLists.txt @@ -1,3 +1,3 @@ -if(WITH_GPU) +if(WITH_GPU AND NOT WIN32) nv_library(nccl_common SRCS nccl_gpu_common.cc DEPS device_context operator ) endif() diff --git a/paddle/fluid/operators/pool_cudnn_op.cu.cc b/paddle/fluid/operators/pool_cudnn_op.cu.cc index 9fdbee818a..31f083565f 100644 --- a/paddle/fluid/operators/pool_cudnn_op.cu.cc +++ b/paddle/fluid/operators/pool_cudnn_op.cu.cc @@ -174,8 +174,7 @@ REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel); REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, - ops::PoolCUDNNGradOpKernel, - ops::PoolCUDNNGradOpKernel); + ops::PoolCUDNNGradOpKernel); REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel, @@ -183,5 +182,4 @@ REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace, ops::PoolCUDNNOpKernel); REGISTER_OP_KERNEL(pool3d_grad, CUDNN, plat::CUDAPlace, ops::PoolCUDNNGradOpKernel, - ops::PoolCUDNNGradOpKernel, - ops::PoolCUDNNGradOpKernel); + ops::PoolCUDNNGradOpKernel); diff --git a/paddle/fluid/operators/recv_op.cc b/paddle/fluid/operators/recv_op.cc index 4a6ce938a5..a1f368e869 100644 --- a/paddle/fluid/operators/recv_op.cc +++ b/paddle/fluid/operators/recv_op.cc @@ -57,6 +57,8 @@ class RecvOp : public framework::OperatorBase { class RecvOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() { + AddInput("X", "(Any) Dummy inputs, used for control dependency") + .AsDuplicable(); AddOutput("Out", "(Tensor) Variables to get from server.").AsDuplicable(); AddComment(R"DOC( Recv operator diff --git a/paddle/fluid/operators/sampling_id_op.cc b/paddle/fluid/operators/sampling_id_op.cc new file mode 100644 index 0000000000..724463c95c --- /dev/null +++ b/paddle/fluid/operators/sampling_id_op.cc @@ -0,0 +1,76 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/sampling_id_op.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +class SamplingIdOp : public framework::OperatorWithKernel { + public: + using framework::OperatorWithKernel::OperatorWithKernel; + + void InferShape(framework::InferShapeContext* ctx) const override { + PADDLE_ENFORCE(ctx->HasInput("X"), + "Input(X) of SamplingIdOp should not be null."); + PADDLE_ENFORCE(ctx->HasOutput("Out"), + "Output(Out) of SamplingIdOp should not be null."); + PADDLE_ENFORCE( + ctx->Attrs().Get("min") < ctx->Attrs().Get("max"), + "min must less then max"); + + auto input_dims = ctx->GetInputDim("X"); + PADDLE_ENFORCE(input_dims.size() == 2, + "Input(X, Filter) should be 2-D tensor."); + + framework::DDim dims = input_dims; + ctx->SetOutputDim("Out", dims); + ctx->ShareLoD("X", "Out"); + } +}; + +class SamplingIdOpMaker : public framework::OpProtoAndCheckerMaker { + public: + void Make() override { + AddInput("X", + "The input tensor of softmax. " + "2-D with shape [batch_size, input_feature_dimensions]."); + AddOutput("Out", "SamplingId data tensor."); + AddComment(R"DOC( +SamplingId Operator. +A layer for sampling id from multinomial distribution from the + input. Sampling one id for one sample.)DOC"); + AddAttr("min", "Minimum value of random. [default 0.0].") + .SetDefault(0.0f); + AddAttr("max", "Maximun value of random. [default 1.0].") + .SetDefault(1.0f); + AddAttr("seed", + "Random seed used for the random number engine. " + "0 means use a seed generated by the system." + "Note that if seed is not 0, this operator will always " + "generate the same random numbers every time. [default 0].") + .SetDefault(0); + } +}; +} // namespace operators +} // namespace paddle + +namespace ops = paddle::operators; +REGISTER_OPERATOR(sampling_id, ops::SamplingIdOp, ops::SamplingIdOpMaker, + paddle::framework::EmptyGradOpMaker); + +REGISTER_OP_CPU_KERNEL(sampling_id, paddle::operators::SamplingIdKernel, + paddle::operators::SamplingIdKernel); diff --git a/paddle/fluid/operators/sampling_id_op.cu b/paddle/fluid/operators/sampling_id_op.cu new file mode 100644 index 0000000000..a4f0470314 --- /dev/null +++ b/paddle/fluid/operators/sampling_id_op.cu @@ -0,0 +1,19 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#include "paddle/fluid/operators/sampling_id_op.h" + +namespace ops = paddle::operators; +REGISTER_OP_CUDA_KERNEL(sampling_id, paddle::operators::SamplingIdKernel, + paddle::operators::SamplingIdKernel); diff --git a/paddle/fluid/operators/sampling_id_op.h b/paddle/fluid/operators/sampling_id_op.h new file mode 100644 index 0000000000..f730a9746d --- /dev/null +++ b/paddle/fluid/operators/sampling_id_op.h @@ -0,0 +1,80 @@ +/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. */ + +#pragma once + +#include +#include +#include +#include +#include +#include + +#include "paddle/fluid/framework/op_registry.h" + +namespace paddle { +namespace operators { + +using Tensor = framework::Tensor; + +template +class SamplingIdKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + const Tensor* input = context.Input("X"); + const int batch_size = static_cast(input->dims()[0]); + const int width = static_cast(input->dims()[1]); + + PADDLE_ENFORCE_GE(batch_size, 0, + "batch_size(dims[0]) must be nonnegative."); + PADDLE_ENFORCE_GE(width, 0, "width(dims[1]) must be nonnegative."); + + std::vector ins_vector; + framework::TensorToVector(*input, context.device_context(), &ins_vector); + + unsigned int seed = static_cast(context.Attr("seed")); + std::minstd_rand engine; + if (seed == 0) { + seed = std::random_device()(); + } + engine.seed(seed); + std::uniform_real_distribution dist( + static_cast(context.Attr("min")), + static_cast(context.Attr("max"))); + + std::vector ids(batch_size); + for (size_t i = 0; i < batch_size; ++i) { + T r = dist(engine); + int idx = width - 1; + for (int j = 0; j < width; ++j) { + if ((r -= ins_vector[i * width + j]) < 0) { + idx = j; + break; + } + } + ids[i] = ins_vector[i * width + idx]; + } + + std::vector out_dim; + out_dim.push_back(static_cast(batch_size)); + + Tensor* output = context.Output("Out"); + output->Resize(framework::make_ddim(out_dim)); + output->mutable_data(context.GetPlace()); + framework::TensorFromVector(ids, context.device_context(), output); + } +}; + +} // namespace operators +} // namespace paddle diff --git a/paddle/fluid/operators/save_op.cc b/paddle/fluid/operators/save_op.cc index 201a51130d..85de37416b 100644 --- a/paddle/fluid/operators/save_op.cc +++ b/paddle/fluid/operators/save_op.cc @@ -142,6 +142,8 @@ class SaveOp : public framework::OperatorBase { std::string filename = lt_var->data(); VLOG(4) << "SaveSelectedRows get File name: " << filename; + MkDirRecursively(DirName(filename).c_str()); + auto &selectedRows = var->Get(); // get device context from pool diff --git a/paddle/fluid/operators/scale_op.cu b/paddle/fluid/operators/scale_op.cu index d266867046..04c802da12 100644 --- a/paddle/fluid/operators/scale_op.cu +++ b/paddle/fluid/operators/scale_op.cu @@ -13,15 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #include "paddle/fluid/operators/scale_op.h" -#include "paddle/fluid/platform/float16.h" -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( scale, paddle::operators::ScaleKernel, paddle::operators::ScaleKernel, paddle::operators::ScaleKernel, paddle::operators::ScaleKernel, - paddle::operators::ScaleKernel); + int64_t>); diff --git a/paddle/fluid/operators/scatter_op.cc b/paddle/fluid/operators/scatter_op.cc index bf5e0d8644..c32d2603cf 100644 --- a/paddle/fluid/operators/scatter_op.cc +++ b/paddle/fluid/operators/scatter_op.cc @@ -81,8 +81,8 @@ class ScatterOpMaker : public framework::OpProtoAndCheckerMaker { void Make() override { AddInput("X", "The source input of scatter op"); AddInput("Ids", "The index input of scatter op where X will be updated"); - AddInput("Updates", "The updated value of updates op"); - AddOutput("Out", "The output of add op"); + AddInput("Updates", "The updated value of scatter op"); + AddOutput("Out", "The output of scatter op"); AddComment(R"DOC( Scatter Operator. @@ -90,7 +90,7 @@ This operator obtains output by updating the input on selected indices on the fi $$ Out = X \\ -Out[Ids] = X[Ids] + Updates +Out[Ids] = Updates $$ )DOC"); diff --git a/paddle/fluid/operators/scatter_op.h b/paddle/fluid/operators/scatter_op.h index d29947b55e..2eefbba972 100644 --- a/paddle/fluid/operators/scatter_op.h +++ b/paddle/fluid/operators/scatter_op.h @@ -34,9 +34,9 @@ class ScatterOpKernel : public framework::OpKernel { auto *Updates = ctx.Input("Updates"); auto *Out = ctx.Output("Out"); - // In place output: Out = X, Out[Ids] += Updates - Out->ShareDataWith(*X); - // Apply ScatterUpdate: Out[index] += Updates[:] + // In place output: Out = X, Out[Ids] = Updates + framework::TensorCopySync(*X, ctx.GetPlace(), Out); + // Apply ScatterUpdate: Out[index] = Updates[:] ScatterAssign(ctx.device_context(), *Updates, *Ids, Out); } }; @@ -53,9 +53,9 @@ class ScatterGradientOpKernel : public framework::OpKernel { auto *dOut = ctx.Input(framework::GradVarName("Out")); // In place gradient: dX = dO - dX->ShareDataWith(*dOut); + framework::TensorCopySync(*dOut, ctx.GetPlace(), dX); dUpdates->mutable_data(ctx.GetPlace()); - // Gradient by Gather: dUpdates += dO[Ids] + // Gradient by Gather: dUpdates = dO[Ids] CPUGather(ctx.device_context(), *dOut, *Ids, dUpdates); } }; diff --git a/paddle/fluid/operators/send_barrier_op.cc b/paddle/fluid/operators/send_barrier_op.cc index 1866a86048..14b07649c4 100644 --- a/paddle/fluid/operators/send_barrier_op.cc +++ b/paddle/fluid/operators/send_barrier_op.cc @@ -37,22 +37,19 @@ class SendBarrierOp : public framework::OperatorBase { void RunImpl(const framework::Scope& scope, const platform::Place& place) const override { std::vector eps = Attr>("endpoints"); - bool sync_mode = Attr("sync_mode"); distributed::RPCClient* rpc_client = distributed::RPCClient::GetInstance(); - VLOG(3) << "SendBarrierOp sync_mode:" << sync_mode; + VLOG(3) << "SendBarrierOp sync"; // need to wait before sending send_barrier message PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); - if (sync_mode) { - for (auto& ep : eps) { - VLOG(3) << "send barrier, ep: " << ep; - rpc_client->AsyncSendBatchBarrier(ep); - } - PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); + for (auto& ep : eps) { + VLOG(3) << "send barrier, ep: " << ep; + rpc_client->AsyncSendBatchBarrier(ep); } + PADDLE_ENFORCE(rpc_client->Wait(), "internal error in RPCClient"); } }; @@ -70,7 +67,6 @@ the Parameter Server would knew all variables have been sent. "(string vector, default 127.0.0.1:6164)" "Server endpoints to send variables to.") .SetDefault({"127.0.0.1:6164"}); - AddAttr("sync_mode", "work in sync_mode or not").SetDefault(true); } }; diff --git a/paddle/fluid/operators/send_op.cc b/paddle/fluid/operators/send_op.cc index 3cd42f2d05..82a70e4bf1 100644 --- a/paddle/fluid/operators/send_op.cc +++ b/paddle/fluid/operators/send_op.cc @@ -66,6 +66,8 @@ class SendOpMaker : public framework::OpProtoAndCheckerMaker { void Make() { AddInput("X", "(Tensor, SelectedRows) Input variables to be sent") .AsDuplicable(); + AddOutput("Out", "(Any) Dummy outputs, used for control dependency") + .AsDuplicable(); AddComment(R"DOC( Send operator diff --git a/paddle/fluid/operators/sgd_op.h b/paddle/fluid/operators/sgd_op.h index 2685ce217e..d8b0165b2a 100644 --- a/paddle/fluid/operators/sgd_op.h +++ b/paddle/fluid/operators/sgd_op.h @@ -111,7 +111,7 @@ class SGDOpKernel : public framework::OpKernel { for (size_t i = 0; i < grad.rows().size(); i++) { PADDLE_ENFORCE(grad.rows()[i] < grad.height(), "Input rows index should less than height"); - int64_t id_index = param.Index(grad.rows()[i]); + int64_t id_index = param_out->AutoGrownIndex(grad.rows()[i], false); PADDLE_ENFORCE_GE(id_index, static_cast(0), "id should be in the table"); for (int64_t j = 0; j < grad_row_width; j++) { diff --git a/paddle/fluid/operators/softmax_cudnn_op.cu.cc b/paddle/fluid/operators/softmax_cudnn_op.cu.cc index c2d45c3d2e..2bdb23e999 100644 --- a/paddle/fluid/operators/softmax_cudnn_op.cu.cc +++ b/paddle/fluid/operators/softmax_cudnn_op.cu.cc @@ -78,5 +78,4 @@ REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace, ops::SoftmaxCUDNNKernel, ops::SoftmaxCUDNNKernel); REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace, - ops::SoftmaxGradCUDNNKernel, - ops::SoftmaxGradCUDNNKernel); + ops::SoftmaxGradCUDNNKernel); diff --git a/paddle/fluid/operators/softmax_op.cu.cc b/paddle/fluid/operators/softmax_op.cu.cc index 19359b7eef..5fb4f011d9 100644 --- a/paddle/fluid/operators/softmax_op.cu.cc +++ b/paddle/fluid/operators/softmax_op.cu.cc @@ -23,5 +23,4 @@ REGISTER_OP_CUDA_KERNEL( ops::SoftmaxKernel); REGISTER_OP_CUDA_KERNEL( softmax_grad, ops::SoftmaxGradKernel, - ops::SoftmaxGradKernel, - ops::SoftmaxGradKernel); + ops::SoftmaxGradKernel); diff --git a/paddle/fluid/operators/squeeze_op.cc b/paddle/fluid/operators/squeeze_op.cc index 6c507baf3a..8a683116b8 100644 --- a/paddle/fluid/operators/squeeze_op.cc +++ b/paddle/fluid/operators/squeeze_op.cc @@ -23,9 +23,9 @@ class SqueezeOpInferShape : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of SqueezeOp should not be null."); + "Input(X) of Squeeze operator should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of SqueezeOp should not be null."); + "Output(Out) of Squeeze operator should not be null."); const auto &x_dims = ctx->GetInputDim("X"); // Check input tensor dims (<6) Eigen limit. @@ -107,7 +107,6 @@ class SqueezeOp : public framework::OperatorBase { framework::AttributeMap attrs; attrs["shape"] = framework::vectorize2int(out_dims); - attrs["inplace"] = Attr("inplace"); // Invoke Reshape Op auto reshape_op = framework::OpRegistry::CreateOp( "reshape", {{"X", {Input("X")}}, {"Shape", {}}}, @@ -125,12 +124,6 @@ class SqueezeOpMaker : public framework::OpProtoAndCheckerMaker { "(std::vector). List of integers," " indicating the dimensions to squeeze.") .SetDefault({}); - AddAttr("inplace", - "(default: false) Squeeze the source tensor's shape without " - "memory copy. When Attr(inplace) is set true, the output " - "tensor shares memory with Input(X), otherwise, a new output " - "tensor is created, and its data are copied from Input(x).") - .SetDefault(false); AddComment(R"DOC( Squeeze Operator. @@ -180,7 +173,6 @@ class SqueezeGradOp : public framework::OperatorBase { auto x_dims = scope.FindVar(Input("X"))->Get().dims(); framework::AttributeMap attrs; attrs["shape"] = framework::vectorize2int(x_dims); - attrs["inplace"] = Attr("inplace"); auto reshape_op = framework::OpRegistry::CreateOp( "reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}}, diff --git a/paddle/fluid/operators/sum_mkldnn_op.cc b/paddle/fluid/operators/sum_mkldnn_op.cc index d2035777ee..f9a16ef35e 100644 --- a/paddle/fluid/operators/sum_mkldnn_op.cc +++ b/paddle/fluid/operators/sum_mkldnn_op.cc @@ -34,15 +34,15 @@ namespace paddle { namespace operators { -using paddle::framework::Tensor; -using paddle::platform::MKLDNNDeviceContext; -using paddle::platform::CPUDeviceContext; using framework::DataLayout; using mkldnn::memory; using mkldnn::primitive; +using mkldnn::reorder; using mkldnn::stream; using mkldnn::sum; -using mkldnn::reorder; +using paddle::framework::Tensor; +using paddle::platform::CPUDeviceContext; +using paddle::platform::MKLDNNDeviceContext; using platform::to_void_cast; template @@ -175,18 +175,35 @@ class SumMKLDNNOpKernel : public paddle::framework::OpKernel { auto& sel_row = get_selected_row(i); first_dim += sel_row.rows().size(); } - auto in_dim = - framework::vectorize(get_selected_row(N - 1).value().dims()); + + std::vector in_dim; + for (int i = 0; i < N; i++) { + auto& sel_row = get_selected_row(i); + if (sel_row.rows().size() > 0) { + in_dim = framework::vectorize(sel_row.value().dims()); + break; + } + } + + if (in_dim.empty()) { + VLOG(3) << "WARNING: all the inputs are empty"; + in_dim = framework::vectorize(get_selected_row(N - 1).value().dims()); + } else { + in_dim[0] = static_cast(first_dim); + } + in_dim[0] = static_cast(first_dim); out_value->Resize(framework::make_ddim(in_dim)); + out_value->mutable_data(ctx.GetPlace()); + // if all the input sparse vars are empty, no need to // merge these vars. if (first_dim == 0UL) { return; } - out_value->mutable_data(ctx.GetPlace()); + math::SelectedRowsAddTo functor; int64_t offset = 0; for (int i = 0; i < N; i++) { diff --git a/paddle/fluid/operators/sum_op.cu b/paddle/fluid/operators/sum_op.cu index db4c2d6c11..89bcd1bbc8 100644 --- a/paddle/fluid/operators/sum_op.cu +++ b/paddle/fluid/operators/sum_op.cu @@ -11,13 +11,10 @@ limitations under the License. */ #define EIGEN_USE_GPU #include "paddle/fluid/operators/sum_op.h" -#include "paddle/fluid/platform/float16.h" namespace ops = paddle::operators; -namespace plat = paddle::platform; REGISTER_OP_CUDA_KERNEL( sum, ops::SumKernel, ops::SumKernel, ops::SumKernel, - ops::SumKernel, - ops::SumKernel); + ops::SumKernel); diff --git a/paddle/fluid/operators/sum_op.h b/paddle/fluid/operators/sum_op.h index dda6772796..6dffe527c1 100644 --- a/paddle/fluid/operators/sum_op.h +++ b/paddle/fluid/operators/sum_op.h @@ -46,7 +46,7 @@ class SumKernel : public framework::OpKernel { if (!in_place) { math::SetConstant constant_functor; constant_functor(context.template device_context(), out, - static_cast(0)); + 0.0); } math::SelectedRowsAddToTensor functor; @@ -105,18 +105,30 @@ class SumKernel : public framework::OpKernel { auto &sel_row = get_selected_row(i); first_dim += sel_row.rows().size(); } - auto in_dim = - framework::vectorize(get_selected_row(N - 1).value().dims()); - in_dim[0] = static_cast(first_dim); + + std::vector in_dim; + for (int i = 0; i < N; i++) { + auto &sel_row = get_selected_row(i); + if (sel_row.rows().size() > 0) { + in_dim = framework::vectorize(sel_row.value().dims()); + break; + } + } + if (in_dim.empty()) { + VLOG(3) << "WARNING: all the inputs are empty"; + in_dim = framework::vectorize(get_selected_row(N - 1).value().dims()); + } else { + in_dim[0] = static_cast(first_dim); + } out_value->Resize(framework::make_ddim(in_dim)); + out_value->mutable_data(context.GetPlace()); // if all the input sparse vars are empty, no need to // merge these vars. if (first_dim == 0UL) { return; } - out_value->mutable_data(context.GetPlace()); math::SelectedRowsAddTo functor; diff --git a/paddle/fluid/operators/tensorrt_engine_op.cc b/paddle/fluid/operators/tensorrt_engine_op.cc index ee3078876c..1048d30171 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.cc +++ b/paddle/fluid/operators/tensorrt_engine_op.cc @@ -17,112 +17,16 @@ #include #include -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" -#include "paddle/fluid/inference/tensorrt/engine.h" -#include "paddle/fluid/inference/utils/singleton.h" #include "paddle/fluid/operators/tensorrt_engine_op.h" namespace paddle { DEFINE_int32(tensorrt_engine_batch_size, 1, "the batch_size of TensorRT"); +DEFINE_int32(tensorrt_max_batch_size, 1, "TensorRT maximum batch size"); +DEFINE_int32(tensorrt_workspace_size, 16 << 20, "TensorRT workspace size"); namespace operators { -using inference::Singleton; -using inference::tensorrt::TRT_EngineManager; - -using FluidDT = framework::proto::VarType_Type; -using TRT_DT = nvinfer1::DataType; - -namespace { - -TRT_DT FluidDataType2TRT(FluidDT type) { - switch (type) { - case FluidDT::VarType_Type_FP32: - return TRT_DT::kFLOAT; - case FluidDT::VarType_Type_INT32: - return TRT_DT::kINT32; - default: - return TRT_DT::kINT32; - } - PADDLE_THROW("unkown type"); - return TRT_DT::kINT32; -} - -nvinfer1::Dims Vec2TRT_Dims(const std::vector &shape) { - PADDLE_ENFORCE_GT(shape.size(), 1UL, - "TensorRT' tensor input requires at least 2 dimensions"); - PADDLE_ENFORCE_LE(shape.size(), 4UL, - "TensorRT' tensor input requires at most 4 dimensions"); - PADDLE_ENFORCE_EQ(shape.size(), 4UL); - return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]); -} - -} // namespace - -template -void TensorRTEngineKernel::Prepare( - const framework::ExecutionContext &context) const { - VLOG(4) << "Prepare engine"; - // Get the ProgramDesc and pass to convert. - framework::proto::BlockDesc block_desc; - block_desc.ParseFromString(context.Attr("subgraph")); - int max_batch = context.Attr("max_batch"); - auto max_workspace = context.Attr("max_workspace"); - auto params = context.Attr>("parameters"); - std::unordered_set parameters; - for (const auto ¶m : params) { - parameters.insert(param); - } - - std::vector output_maps = - context.Attr>("output_name_mapping"); - - // TODO(Superjomn) replace this with a different stream - auto *engine = Singleton::Global().Create( - max_batch, max_workspace, nullptr /*engine hold its own stream*/, - context.Attr("engine_uniq_key")); - engine->InitNetwork(); - - framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); - VLOG(4) << "parsed var size " << block.AllVars().size(); - // Add inputs - VLOG(4) << "declare inputs"; - for (auto &input : context.Inputs("Xs")) { - if (parameters.count(input)) continue; - VLOG(4) << "declare input " << input; - auto *var = block.FindVar(input); - // TensorRT engine need to create parameters. The parameter's description - // should be set in - PADDLE_ENFORCE(var, "no variable called %s", input); - PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, - "TensorRT engine only takes LoDTensor as input"); - auto shape = var->GetShape(); - // For the special batch_size placeholder -1, drop it and pass the real - // shape of data. - // TODO(Superjomn) fix this with batch broadcast, or it can't handle - // variational batch size. - if (shape[0] == -1) { - shape[0] = FLAGS_tensorrt_engine_batch_size; - } - engine->DeclareInput( - input, FluidDataType2TRT( - var->Proto()->type().lod_tensor().tensor().data_type()), - Vec2TRT_Dims(shape)); - } - - inference::Singleton::Global().ConvertBlock( - block_desc, parameters, context.scope(), engine); - - // Add outputs - for (auto &output : output_maps) { - engine->DeclareOutput(output); - } - - engine->FreezeNetwork(); -} - class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { public: void Make() override { @@ -130,8 +34,6 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("Ys", "A list of outputs").AsDuplicable(); AddAttr("subgraph", "the subgraph."); AddAttr("engine_uniq_key", "unique key for the TRT engine."); - AddAttr("max_batch", "the maximum batch size."); - AddAttr("max_workspace", "the maximum batch size."); AddComment("TensorRT engine operator."); } }; @@ -150,11 +52,4 @@ namespace ops = paddle::operators; REGISTER_OPERATOR(tensorrt_engine, ops::TensorRTEngineOp, ops::TensorRTEngineOpMaker, ops::TensorRTEngineOpMaker); -REGISTER_OP_CPU_KERNEL( - tensorrt_engine, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel, - ops::TensorRTEngineKernel); - #endif // PADDLE_WITH_CUDA diff --git a/paddle/fluid/operators/tensorrt_engine_op.cu.cc b/paddle/fluid/operators/tensorrt_engine_op.cu.cc new file mode 100644 index 0000000000..e1ddfde6d5 --- /dev/null +++ b/paddle/fluid/operators/tensorrt_engine_op.cu.cc @@ -0,0 +1,24 @@ +/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. */ + +#include "paddle/fluid/operators/tensorrt_engine_op.h" + +namespace ops = paddle::operators; + +REGISTER_OP_CUDA_KERNEL( + tensorrt_engine, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel, + ops::TensorRTEngineKernel); diff --git a/paddle/fluid/operators/tensorrt_engine_op.h b/paddle/fluid/operators/tensorrt_engine_op.h index 2cbe1213a2..bc556ab364 100644 --- a/paddle/fluid/operators/tensorrt_engine_op.h +++ b/paddle/fluid/operators/tensorrt_engine_op.h @@ -19,16 +19,51 @@ #include #include +#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" #include "paddle/fluid/inference/analysis/helper.h" +#include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/engine.h" namespace paddle { DECLARE_int32(tensorrt_engine_batch_size); +DECLARE_int32(tensorrt_max_batch_size); +DECLARE_int32(tensorrt_workspace_size); namespace operators { +using FluidDT = framework::proto::VarType_Type; +using TRT_DT = nvinfer1::DataType; + +namespace { + +TRT_DT FluidDataType2TRT(FluidDT type) { + switch (type) { + case FluidDT::VarType_Type_FP32: + return TRT_DT::kFLOAT; + case FluidDT::VarType_Type_INT32: + return TRT_DT::kINT32; + default: + return TRT_DT::kINT32; + } + PADDLE_THROW("unkown type"); + return TRT_DT::kINT32; +} + +nvinfer1::Dims Vec2TRT_Dims(const std::vector& shape) { + PADDLE_ENFORCE_GT(shape.size(), 1UL, + "TensorRT' tensor input requires at least 2 dimensions"); + PADDLE_ENFORCE_LE(shape.size(), 4UL, + "TensorRT' tensor input requires at most 4 dimensions"); + PADDLE_ENFORCE(shape.size() == 4UL || shape.size() == 2UL); + if (shape.size() == 4UL) + return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]); + return nvinfer1::DimsCHW(shape[1], 1, 1); +} + +} // namespace + using inference::Singleton; using inference::tensorrt::TRT_EngineManager; @@ -47,7 +82,7 @@ class TensorRTEngineOp : public framework::OperatorWithKernel { .FindVar(input0) ->GetMutable() ->type()), - platform::CPUPlace()); + ctx.GetPlace()); return kt; } }; @@ -64,7 +99,7 @@ class TensorRTEngineKernel : public framework::OpKernel { auto input_names = context.op().Inputs("Xs"); PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs"); PADDLE_ENFORCE_LE(FLAGS_tensorrt_engine_batch_size, - context.Attr("max_batch")); + FLAGS_tensorrt_max_batch_size); std::vector output_maps = context.Attr>("output_name_mapping"); @@ -94,12 +129,19 @@ class TensorRTEngineKernel : public framework::OpKernel { // Convert output tensor from engine to fluid int output_index = 0; + VLOG(4) << "TensorRT Engine Op Outputs:"; for (const auto& y : context.Outputs("Ys")) { + VLOG(4) << y; // convert output and copy to fluid. nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]); auto dims = trt_t->getDimensions(); // Use the output ITensor's dims to reshape the Fluid Tensor. - std::vector ddim(dims.d, dims.d + dims.nbDims); + // The ITensor doesn't contain the batch size dim. + std::vector ddim; + ddim.push_back(FLAGS_tensorrt_engine_batch_size); + for (int i = 0; i < dims.nbDims; i++) { + ddim.push_back(dims.d[i]); + } auto* fluid_v = context.scope().FindVar(y); PADDLE_ENFORCE_NOT_NULL(fluid_v, "no output variable called %s", y); @@ -113,9 +155,11 @@ class TensorRTEngineKernel : public framework::OpKernel { // TODO(Superjomn) change this float to dtype size. auto size = inference::analysis::AccuDims(dims.d, dims.nbDims) * FLAGS_tensorrt_engine_batch_size; - engine->GetOutputInCPU(output_maps[output_index], - fluid_t->mutable_data(platform::CPUPlace()), - size * sizeof(float)); + engine->GetOutputInGPU( + output_maps[output_index], + fluid_t->mutable_data(platform::CUDAPlace( + boost::get(context.GetPlace()).device)), + size * sizeof(float)); //} else { // engine->GetOutputInGPU( // y, fluid_t->mutable_data(platform::CUDAPlace()), @@ -128,8 +172,67 @@ class TensorRTEngineKernel : public framework::OpKernel { } protected: - // Build the engine. - void Prepare(const framework::ExecutionContext& context) const; + void Prepare(const framework::ExecutionContext& context) const { + VLOG(4) << "Prepare engine"; + // Get the ProgramDesc and pass to convert. + framework::proto::BlockDesc block_desc; + block_desc.ParseFromString(context.Attr("subgraph")); + int max_batch = FLAGS_tensorrt_max_batch_size; + auto max_workspace = FLAGS_tensorrt_workspace_size; + auto params = context.Attr>("parameters"); + std::unordered_set parameters; + for (const auto& param : params) { + parameters.insert(param); + } + + std::vector output_maps = + context.Attr>("output_name_mapping"); + + // TODO(Superjomn) replace this with a different stream + auto* engine = Singleton::Global().Create( + max_batch, max_workspace, nullptr /*engine hold its own stream*/, + context.Attr("engine_uniq_key"), + boost::get(context.GetPlace()).device); + + engine->InitNetwork(); + + framework::BlockDesc block(nullptr /*programdesc*/, &block_desc); + VLOG(4) << "parsed var size " << block.AllVars().size(); + // Add inputs + VLOG(4) << "declare inputs"; + for (auto& input : context.Inputs("Xs")) { + if (parameters.count(input)) continue; + VLOG(4) << "declare input " << input; + auto* var = block.FindVar(input); + // TensorRT engine need to create parameters. The parameter's description + // should be set in + PADDLE_ENFORCE(var, "no variable called %s", input); + PADDLE_ENFORCE_EQ(var->GetType(), FluidDT::VarType_Type_LOD_TENSOR, + "TensorRT engine only takes LoDTensor as input"); + auto shape = var->GetShape(); + // For the special batch_size placeholder -1, drop it and pass the real + // shape of data. + // TODO(Superjomn) fix this with batch broadcast, or it can't handle + // variational batch size. + if (shape[0] == -1) { + shape[0] = FLAGS_tensorrt_engine_batch_size; + } + engine->DeclareInput( + input, FluidDataType2TRT( + var->Proto()->type().lod_tensor().tensor().data_type()), + Vec2TRT_Dims(shape)); + } + + inference::Singleton::Global() + .ConvertBlock(block_desc, parameters, context.scope(), engine); + + // Add outputs + for (auto& output : output_maps) { + engine->DeclareOutput(output); + } + + engine->FreezeNetwork(); + } }; } // namespace operators diff --git a/paddle/fluid/operators/tensorrt_engine_op_test.cc b/paddle/fluid/operators/tensorrt_engine_op_test.cc index 37657fa0b0..27c1d29762 100644 --- a/paddle/fluid/operators/tensorrt_engine_op_test.cc +++ b/paddle/fluid/operators/tensorrt_engine_op_test.cc @@ -12,6 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ +#include "paddle/fluid/operators/tensorrt_engine_op.h" #include #include "paddle/fluid/framework/block_desc.h" #include "paddle/fluid/framework/lod_tensor.h" @@ -23,20 +24,20 @@ limitations under the License. */ #include "paddle/fluid/inference/tensorrt/convert/op_converter.h" #include "paddle/fluid/inference/tensorrt/convert/ut_helper.h" -USE_CPU_ONLY_OP(tensorrt_engine); +USE_CUDA_ONLY_OP(tensorrt_engine); namespace paddle { namespace operators { namespace { -void CreateCPUTensor(framework::Scope* scope, const std::string& name, - const std::vector& shape) { +void CreateCUDATensor(framework::Scope* scope, const std::string& name, + const std::vector& shape) { auto* var = scope->Var(name); auto* tensor = var->GetMutable(); auto dims = framework::make_ddim(shape); tensor->Resize(dims); - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); inference::tensorrt::RandomizeTensor(tensor, place, ctx); } @@ -57,6 +58,8 @@ void AddTensorToBlockDesc(framework::proto::BlockDesc* block, using inference::analysis::SetAttr; TEST(TensorRTEngineOp, manual) { + FLAGS_tensorrt_engine_batch_size = 2; + FLAGS_tensorrt_max_batch_size = 2; framework::ProgramDesc program; auto* block_ = program.Proto()->add_blocks(); block_->set_idx(0); @@ -98,8 +101,6 @@ TEST(TensorRTEngineOp, manual) { engine_op_desc.SetOutput("Ys", std::vector({"z0"})); SetAttr(engine_op_desc.Proto(), "subgraph", block_->SerializeAsString()); - SetAttr(engine_op_desc.Proto(), "max_batch", 100); - SetAttr(engine_op_desc.Proto(), "max_workspace", 1 << 10); SetAttr(engine_op_desc.Proto(), "engine_uniq_key", "a_engine"); SetAttr>(engine_op_desc.Proto(), "parameters", std::vector({})); @@ -112,15 +113,15 @@ TEST(TensorRTEngineOp, manual) { LOG(INFO) << "engine_op " << engine_op.get(); framework::Scope scope; - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); // Prepare variables. - CreateCPUTensor(&scope, "x", std::vector({2, 4})); - CreateCPUTensor(&scope, "y", std::vector({4, 6})); - CreateCPUTensor(&scope, "z", std::vector({2, 6})); + CreateCUDATensor(&scope, "x", std::vector({2, 4})); + CreateCUDATensor(&scope, "y", std::vector({4, 6})); + CreateCUDATensor(&scope, "z", std::vector({2, 6})); - CreateCPUTensor(&scope, "y0", std::vector({6, 8})); - CreateCPUTensor(&scope, "z0", std::vector({2, 8})); + CreateCUDATensor(&scope, "y0", std::vector({6, 8})); + CreateCUDATensor(&scope, "z0", std::vector({2, 8})); // Execute them. LOG(INFO) << "engine_op run"; @@ -128,10 +129,12 @@ TEST(TensorRTEngineOp, manual) { } void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { + FLAGS_tensorrt_engine_batch_size = batch_size; + FLAGS_tensorrt_max_batch_size = batch_size; framework::ProgramDesc program; framework::Scope scope; - platform::CPUPlace place; - platform::CPUDeviceContext ctx(place); + platform::CUDAPlace place; + platform::CUDADeviceContext ctx(place); auto* block_ = program.Proto()->add_blocks(); block_->set_idx(0); @@ -165,10 +168,10 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) { // Prepare variables. if (!x_created) { - CreateCPUTensor(&scope, x_name, std::vector(x_shape)); + CreateCUDATensor(&scope, x_name, std::vector(x_shape)); } - CreateCPUTensor(&scope, y_name, std::vector(y_shape)); - CreateCPUTensor(&scope, z_name, std::vector(z_shape)); + CreateCUDATensor(&scope, y_name, std::vector(y_shape)); + CreateCUDATensor(&scope, z_name, std::vector(z_shape)); // It is wired, need to copy manually. *block_->add_ops() = *fc->Proto(); diff --git a/paddle/fluid/operators/top_k_op.cu b/paddle/fluid/operators/top_k_op.cu index 5fc0784f66..9da8551eb2 100644 --- a/paddle/fluid/operators/top_k_op.cu +++ b/paddle/fluid/operators/top_k_op.cu @@ -11,19 +11,16 @@ distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/operators/top_k_op.h" #include "paddle/fluid/platform/assert.h" #include "paddle/fluid/platform/cuda_device_function.h" -#include "paddle/fluid/platform/float16.h" namespace paddle { namespace operators { using Tensor = framework::Tensor; -using paddle::platform::float16; template struct Pair { @@ -35,11 +32,6 @@ struct Pair { id = id; } - __device__ __forceinline__ void clear() { - v = -INFINITY; - id = -1; - } - __device__ __forceinline__ void operator=(const Pair& in) { v = in.v; id = in.id; @@ -61,12 +53,6 @@ struct Pair { int64_t id; }; -template <> -__device__ __forceinline__ void Pair::clear() { - v = platform::raw_uint16_to_float16(0x400); - id = -1; -} - template __device__ __forceinline__ void AddTo(Pair topk[], const Pair& p, int beam_size) { @@ -164,7 +150,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair topk[], int* beam, if (k < MaxLength - (*beam)) { topk[k] = topk[k + *beam]; } else { - topk[k].clear(); + topk[k].set(-INFINITY, -1); } } if (!(*is_empty)) { @@ -174,7 +160,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair topk[], int* beam, } *max = topk[MaxLength - 1]; - if ((*max).v == static_cast(-1)) *is_empty = true; + if ((*max).v == -1) *is_empty = true; *beam = 0; } } @@ -195,7 +181,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair topk[], int* beam, if (k < MaxLength - *beam) { topk[k] = topk[k + *beam]; } else { - topk[k].set(std::numeric_limits::min(), -1); + topk[k].set(-INFINITY, -1); } } if (!(*is_empty)) { @@ -287,7 +273,7 @@ __global__ void KeMatrixTopK(T* output, int output_stride, int64_t* indices, bool firststep = true; for (int k = 0; k < MaxLength; k++) { - topk[k].clear(); + topk[k].set(-INFINITY, -1); } while (k) { ThreadGetTopK(topk, &beam, k, @@ -339,7 +325,5 @@ class TopkOpCUDAKernel : public framework::OpKernel { } // namespace operators } // namespace paddle -REGISTER_OP_CUDA_KERNEL( - top_k, paddle::operators::TopkOpCUDAKernel, - paddle::operators::TopkOpCUDAKernel, - paddle::operators::TopkOpCUDAKernel); +REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel, + paddle::operators::TopkOpCUDAKernel); diff --git a/paddle/fluid/operators/uniform_random_op.cc b/paddle/fluid/operators/uniform_random_op.cc index edd1baa4ac..5248767c2e 100644 --- a/paddle/fluid/operators/uniform_random_op.cc +++ b/paddle/fluid/operators/uniform_random_op.cc @@ -30,8 +30,10 @@ class CPUUniformRandomKernel : public framework::OpKernel { tensor = out_var->GetMutable(); } else if (out_var->IsType()) { auto shape = ctx.Attr>("shape"); - tensor = out_var->GetMutable()->mutable_value(); + auto* selected_rows = out_var->GetMutable(); + tensor = selected_rows->mutable_value(); tensor->Resize(framework::make_ddim(shape)); + selected_rows->mutable_rows()->reserve(shape[0]); } else { PADDLE_THROW( "uniform_random_op's output only" diff --git a/paddle/fluid/operators/uniform_random_op.cu b/paddle/fluid/operators/uniform_random_op.cu index 2b8039a0c1..e1c7323a30 100644 --- a/paddle/fluid/operators/uniform_random_op.cu +++ b/paddle/fluid/operators/uniform_random_op.cu @@ -11,14 +11,10 @@ distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. */ -#include #include #include -#include "paddle/fluid/framework/eigen.h" #include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/operator.h" -#include "paddle/fluid/platform/float16.h" -#include "paddle/fluid/platform/transform.h" namespace paddle { namespace operators { @@ -40,11 +36,6 @@ struct UniformGenerator { } }; -template -struct CastFunctor { - HOSTDEVICE V operator()(const T& a) { return static_cast(a); } -}; - // It seems that Eigen::Tensor::random in GPU will SEGFAULT. // Use std::random and thrust::random(thrust is a std library in CUDA) to // implement uniform random. @@ -75,50 +66,18 @@ class GPUUniformRandomKernel : public framework::OpKernel { T max = static_cast(context.Attr("max")); thrust::counting_iterator index_sequence_begin(0); int64_t size = tensor->numel(); - if (out_var->IsType() && - std::type_index(typeid(T)) == - std::type_index(typeid(platform::float16))) { - framework::Tensor master_copy_tensor; - master_copy_tensor.Resize(tensor->dims()); - float* master_copy_tensor_data = - master_copy_tensor.mutable_data(context.GetPlace()); - thrust::transform(index_sequence_begin, index_sequence_begin + size, - thrust::device_ptr(master_copy_tensor_data), - UniformGenerator(static_cast(min), - static_cast(max), seed)); - platform::Transform trans; - auto* in_begin = master_copy_tensor.data(); - auto* in_end = in_begin + master_copy_tensor.numel(); - auto* out_begin = tensor->mutable_data(context.GetPlace()); - trans(context.template device_context(), - in_begin, in_end, out_begin, CastFunctor()); - } else { - thrust::transform(index_sequence_begin, index_sequence_begin + size, - thrust::device_ptr(data), - UniformGenerator(min, max, seed)); - } - if (VLOG_IS_ON(5)) { - framework::Tensor cpu_tensor; - framework::TensorCopySync(*tensor, platform::CPUPlace(), &cpu_tensor); - auto& dev_ctx = - *platform::DeviceContextPool::Instance().Get(context.GetPlace()); - dev_ctx.Wait(); - auto x = framework::EigenVector::Flatten(cpu_tensor); - VLOG(5) << "The Uniform output " << x; - } + thrust::transform(index_sequence_begin, index_sequence_begin + size, + thrust::device_ptr(data), + UniformGenerator(min, max, seed)); } }; } // namespace operators } // namespace paddle -namespace plat = paddle::platform; -REGISTER_OP_CUDA_KERNEL( - uniform_random, paddle::operators::GPUUniformRandomKernel, - paddle::operators::GPUUniformRandomKernel, - paddle::operators::GPUUniformRandomKernel); -REGISTER_OP_CUDA_KERNEL( - uniform_random_batch_size_like, - paddle::operators::GPUUniformRandomKernel, - paddle::operators::GPUUniformRandomKernel, - paddle::operators::GPUUniformRandomKernel); +REGISTER_OP_CUDA_KERNEL(uniform_random, + paddle::operators::GPUUniformRandomKernel, + paddle::operators::GPUUniformRandomKernel); +REGISTER_OP_CUDA_KERNEL(uniform_random_batch_size_like, + paddle::operators::GPUUniformRandomKernel, + paddle::operators::GPUUniformRandomKernel); diff --git a/paddle/fluid/operators/unsqueeze_op.cc b/paddle/fluid/operators/unsqueeze_op.cc index f2a15fdf57..0fc8d54f64 100644 --- a/paddle/fluid/operators/unsqueeze_op.cc +++ b/paddle/fluid/operators/unsqueeze_op.cc @@ -23,9 +23,9 @@ class UnsqueezeOpInferShape : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext *ctx) const override { PADDLE_ENFORCE(ctx->HasInput("X"), - "Input(X) of UnsqueezeOp should not be null."); + "Input(X) of Unsqueeze operator should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), - "Output(Out) of UnsqueezeOp should not be null."); + "Output(Out) of Unsqueeze operator should not be null."); const auto &axes = ctx->Attrs().Get>("axes"); const auto &x_dims = ctx->GetInputDim("X"); @@ -95,7 +95,6 @@ class UnsqueezeOp : public framework::OperatorBase { framework::AttributeMap attrs; attrs["shape"] = framework::vectorize2int(out_dims); - attrs["inplace"] = Attr("inplace"); // Invoke Reshape op. auto reshape_op = framework::OpRegistry::CreateOp( "reshape", {{"X", {Input("X")}}, {"Shape", {}}}, @@ -126,13 +125,6 @@ class UnsqueezeOpMaker : public framework::OpProtoAndCheckerMaker { " within [1, 6] dimensions (Eigen limit)."); } }); - AddAttr( - "inplace", - "(default: false) Unsqueeze the source tensor's shape without " - "memory copy. When Attr(inplace) is set true, the output " - "tensor shares memory with Input(X), otherwise, a new output " - "tensor is created, and its data are copied from Input(x).") - .SetDefault(false); AddComment(R"DOC( Unsqueeze Operator. @@ -168,7 +160,6 @@ class UnsqueezeGradOp : public framework::OperatorBase { framework::AttributeMap attrs; attrs["shape"] = framework::vectorize2int(x_dims); - attrs["inplace"] = Attr("inplace"); auto reshape_op = framework::OpRegistry::CreateOp( "reshape", {{"X", {dout_name}}, {"Shape", {}}}, {{"Out", {dx_name}}}, diff --git a/paddle/fluid/operators/while_op.cc b/paddle/fluid/operators/while_op.cc index 733157ea05..48e37796e1 100644 --- a/paddle/fluid/operators/while_op.cc +++ b/paddle/fluid/operators/while_op.cc @@ -57,12 +57,12 @@ class WhileOp : public framework::OperatorBase { PADDLE_ENFORCE(platform::is_cpu_place(cond.place()), "Condition of while op must in CPU memory."); + + auto ctx = executor.Prepare(*program, block->ID()); while (cond.data()[0]) { auto ¤t_scope = scope.NewScope(); step_scopes->push_back(¤t_scope); - - executor.Run(*program, ¤t_scope, block->ID(), - false /*create_local_scope*/); + executor.RunPreparedContext(ctx.get(), ¤t_scope, false); } } }; @@ -109,6 +109,7 @@ class WhileGradOp : public framework::OperatorBase { framework::Executor executor(dev_place); auto *block = Attr(kStepBlock); auto *program = block->Program(); + auto ctx = executor.Prepare(*program, block->ID()); auto *step_scopes = scope.FindVar(Input(kStepScopes))->GetMutable(); @@ -161,8 +162,7 @@ class WhileGradOp : public framework::OperatorBase { } } } - - executor.Run(*program, *cur_scope_iter, block->ID(), false); + executor.RunPreparedContext(ctx.get(), *cur_scope_iter, false); auto &pg_names = Outputs(kXGRAD); auto &p_names = Inputs(kX); diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h index 23457ff5fe..9f504d14a8 100644 --- a/paddle/fluid/platform/cuda_device_function.h +++ b/paddle/fluid/platform/cuda_device_function.h @@ -36,7 +36,7 @@ __forceinline__ __device__ T CudaShuffleDownSync(unsigned mask, T val, #if CUDA_VERSION < 9000 return __shfl_down(val, delta, width); #else - return __shfl_down_sync(mask, val, delta, width); + return __shfl_down_sync(mask, val, static_cast(delta), width); #endif } @@ -46,9 +46,16 @@ template <> __forceinline__ __device__ float16 CudaShuffleDownSync(unsigned mask, float16 val, int delta, int width) { - half tmp = static_cast(val); - __shfl_down(tmp, static_cast(delta), width); - return float16(tmp); + return float16( + __shfl_down(static_cast(val), static_cast(delta), width)); +} +#else +template <> +__forceinline__ __device__ float16 CudaShuffleDownSync(unsigned mask, + float16 val, int delta, + int width) { + return float16(__shfl_down_sync(mask, static_cast(val), + static_cast(delta), width)); } #endif diff --git a/paddle/fluid/platform/cuda_helper_test.cu b/paddle/fluid/platform/cuda_helper_test.cu index ca5ca1caeb..ee45afab93 100644 --- a/paddle/fluid/platform/cuda_helper_test.cu +++ b/paddle/fluid/platform/cuda_helper_test.cu @@ -13,6 +13,7 @@ // limitations under the License. #include +#include #include #include @@ -123,7 +124,7 @@ void TestUnalign(size_t num, const int shift_bit) { cudaMemcpy(out, d_in2, array_size, cudaMemcpyDeviceToHost); cudaDeviceSynchronize(); for (size_t i = 0; i < num / 2; ++i) { - // NOTE(dzhwinter): the float16 add has small underflow/overflow + // NOTE(dzhwinter): the float16 add has small truncate error. // so we use EXPECT_NEAR to check the result. EXPECT_NEAR(static_cast(out[i]), static_cast(AddFunctor()(r_in1[i], r_in2[i])), @@ -151,3 +152,83 @@ TEST(CudaAtomic, float16Unalign) { TestUnalign(static_cast(1024), /*shift_bit*/ 3); TestUnalign(static_cast(1024 * 1024), /*shift_bit*/ 3); } + +// https://devblogs.nvidia.com/faster-parallel-reductions-kepler/ +template +static __forceinline__ __device__ T WarpReduceSum(T val) { + unsigned mask = 0u; + CREATE_SHFL_MASK(mask, true); + for (int offset = warpSize / 2; offset > 0; offset /= 2) { + val += paddle::platform::CudaShuffleDownSync(mask, val, offset); + } + return val; +} + +template +__forceinline__ __device__ T BlockReduce(T val) { + static __shared__ T shared[32]; // Shared mem for 32 partial sums + int lane = threadIdx.x % warpSize; + int wid = threadIdx.x / warpSize; + + val = WarpReduceSum(val); // Each warp performs partial reduction + + if (lane == 0) shared[wid] = val; // Write reduced value to shared memory + + __syncthreads(); // Wait for all partial reductions + + // read from shared memory only if that warp existed + val = + (threadIdx.x < blockDim.x / warpSize) ? shared[lane] : static_cast(0); + + if (wid == 0) val = WarpReduceSum(val); // Final reduce within first warp + + return val; +} + +template +__global__ void DeviceReduceSum(T* in, T* out, size_t N) { + T sum(0); + for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + sum += in[i]; + } + sum = BlockReduce(sum); + __syncthreads(); + if (threadIdx.x == 0) out[blockIdx.x] = sum; +} + +template +void TestReduce(size_t num, float atol = 0.01) { + T* in1; + T *d_in1, *d_in2; + size_t size = sizeof(T) * num; + cudaMalloc(reinterpret_cast(&d_in1), size); + cudaMalloc(reinterpret_cast(&d_in2), sizeof(T)); + in1 = reinterpret_cast(malloc(size)); + std::minstd_rand engine; + std::uniform_real_distribution dist(0.0, 1.0); + for (size_t i = 0; i < num; ++i) { + in1[i] = static_cast(dist(engine)); + } + auto out = std::accumulate(in1, in1 + num, static_cast(0)); + cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); + cudaDeviceSynchronize(); + DeviceReduceSum<<<1, PADDLE_CUDA_NUM_THREADS>>>(d_in1, d_in2, num); + cudaMemcpy(in1, d_in2, sizeof(T), cudaMemcpyDeviceToHost); + cudaDeviceSynchronize(); + // NOTE(dzhwinter): the float16 add has small underflow/overflow + // so we use EXPECT_NEAR to check the result. + EXPECT_NEAR(static_cast(in1[0]), static_cast(out), atol); + free(in1); + cudaFree(d_in1); + cudaFree(d_in2); +} + +TEST(CudaShuffleSync, float16) { + TestReduce(10); + TestReduce(1000); + + // float16 will overflow or accumulate truncate errors in big size. + TestReduce(10); + TestReduce(100, /*atol error*/ 1.0); +} diff --git a/paddle/fluid/platform/dynload/CMakeLists.txt b/paddle/fluid/platform/dynload/CMakeLists.txt index 9da787a407..07159d4a12 100644 --- a/paddle/fluid/platform/dynload/CMakeLists.txt +++ b/paddle/fluid/platform/dynload/CMakeLists.txt @@ -3,7 +3,7 @@ cc_library(dynamic_loader SRCS dynamic_loader.cc DEPS glog gflags enforce) list(APPEND CUDA_SRCS cublas.cc cudnn.cc curand.cc) # There is no macOS version of NCCL. -if (NOT APPLE) +if (NOT APPLE AND NOT WIN32) list(APPEND CUDA_SRCS nccl.cc) endif() diff --git a/paddle/fluid/platform/dynload/cublas.h b/paddle/fluid/platform/dynload/cublas.h index 25bcda7eed..c7c533bd42 100644 --- a/paddle/fluid/platform/dynload/cublas.h +++ b/paddle/fluid/platform/dynload/cublas.h @@ -17,10 +17,10 @@ #include #include #include -#include #include // NOLINT #include #include "paddle/fluid/platform/dynload/dynamic_loader.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace platform { diff --git a/paddle/fluid/platform/dynload/cudnn.h b/paddle/fluid/platform/dynload/cudnn.h index 77e46fa768..0103e7a3ac 100644 --- a/paddle/fluid/platform/dynload/cudnn.h +++ b/paddle/fluid/platform/dynload/cudnn.h @@ -15,9 +15,9 @@ limitations under the License. */ #pragma once #include -#include #include // NOLINT #include "paddle/fluid/platform/dynload/dynamic_loader.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace platform { diff --git a/paddle/fluid/platform/dynload/cupti.h b/paddle/fluid/platform/dynload/cupti.h index e8f4a82ef1..b946f46e82 100644 --- a/paddle/fluid/platform/dynload/cupti.h +++ b/paddle/fluid/platform/dynload/cupti.h @@ -17,10 +17,10 @@ limitations under the License. */ #include #include -#include #include // NOLINT #include "paddle/fluid/platform/dynload/dynamic_loader.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace platform { diff --git a/paddle/fluid/platform/dynload/curand.h b/paddle/fluid/platform/dynload/curand.h index 5b9e0820e0..2daf1b4215 100644 --- a/paddle/fluid/platform/dynload/curand.h +++ b/paddle/fluid/platform/dynload/curand.h @@ -14,9 +14,9 @@ limitations under the License. */ #pragma once #include -#include #include // NOLINT +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/platform/dynload/dynamic_loader.h" diff --git a/paddle/fluid/platform/dynload/mklml.h b/paddle/fluid/platform/dynload/mklml.h index 17acefe8cd..15ad4a3b40 100644 --- a/paddle/fluid/platform/dynload/mklml.h +++ b/paddle/fluid/platform/dynload/mklml.h @@ -14,10 +14,10 @@ limitations under the License. */ #pragma once -#include #include #include // NOLINT #include "paddle/fluid/platform/dynload/dynamic_loader.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace platform { @@ -49,17 +49,27 @@ extern void* mklml_dso_handle; #define MKLML_ROUTINE_EACH(__macro) \ __macro(cblas_sgemm); \ - __macro(cblas_saxpy); \ - __macro(cblas_scopy); \ - __macro(cblas_sgemv); \ - __macro(cblas_sgemm_batch); \ __macro(cblas_dgemm); \ + __macro(cblas_saxpy); \ __macro(cblas_daxpy); \ + __macro(cblas_scopy); \ __macro(cblas_dcopy); \ + __macro(cblas_sgemv); \ __macro(cblas_dgemv); \ + __macro(cblas_sgemm_alloc); \ + __macro(cblas_dgemm_alloc); \ + __macro(cblas_sgemm_pack); \ + __macro(cblas_dgemm_pack); \ + __macro(cblas_sgemm_compute); \ + __macro(cblas_dgemm_compute); \ + __macro(cblas_sgemm_free); \ + __macro(cblas_dgemm_free); \ + __macro(cblas_sgemm_batch); \ __macro(cblas_dgemm_batch); \ __macro(vsAdd); \ __macro(vdAdd); \ + __macro(vsMul); \ + __macro(vdMul); \ __macro(MKL_Set_Num_Threads) MKLML_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_MKLML_WRAP); diff --git a/paddle/fluid/platform/dynload/nccl.h b/paddle/fluid/platform/dynload/nccl.h index 575516f818..331ca9908e 100644 --- a/paddle/fluid/platform/dynload/nccl.h +++ b/paddle/fluid/platform/dynload/nccl.h @@ -13,12 +13,11 @@ See the License for the specific language governing permissions and limitations under the License. */ #pragma once -#include #include #include // NOLINT - #include "paddle/fluid/platform/dynload/dynamic_loader.h" +#include "paddle/fluid/platform/port.h" namespace paddle { namespace platform { diff --git a/paddle/fluid/platform/dynload/warpctc.h b/paddle/fluid/platform/dynload/warpctc.h index d157c1fda7..18ed9956f1 100644 --- a/paddle/fluid/platform/dynload/warpctc.h +++ b/paddle/fluid/platform/dynload/warpctc.h @@ -14,10 +14,9 @@ limitations under the License. */ #pragma once -#include #include // NOLINT - #include "paddle/fluid/platform/dynload/dynamic_loader.h" +#include "paddle/fluid/platform/port.h" #include "warpctc/include/ctc.h" namespace paddle { diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h index 566485cd3c..a76ba75f9e 100644 --- a/paddle/fluid/platform/enforce.h +++ b/paddle/fluid/platform/enforce.h @@ -14,9 +14,6 @@ limitations under the License. */ #pragma once -#include // for dladdr -#include // for backtrace - #ifdef __GNUC__ #include // for __cxa_demangle #endif // __GNUC__ @@ -37,6 +34,7 @@ limitations under the License. */ #include "glog/logging.h" #include "paddle/fluid/platform/macros.h" +#include "paddle/fluid/platform/port.h" #include "paddle/fluid/string/printf.h" #include "paddle/fluid/string/to_string.h" @@ -44,7 +42,7 @@ limitations under the License. */ #include "paddle/fluid/platform/dynload/cublas.h" #include "paddle/fluid/platform/dynload/cudnn.h" #include "paddle/fluid/platform/dynload/curand.h" -#ifndef __APPLE__ +#if !defined(__APPLE__) and !defined(_WIN32) #include "paddle/fluid/platform/dynload/nccl.h" #endif // __APPLE__ #endif // PADDLE_WITH_CUDA @@ -75,7 +73,7 @@ struct EnforceNotMet : public std::exception { sout << string::Sprintf("%s at [%s:%d]", exp.what(), f, l) << std::endl; sout << "PaddlePaddle Call Stacks: " << std::endl; - +#if !defined(_WIN32) void* call_stack[TRACE_STACK_LIMIT]; auto size = backtrace(call_stack, TRACE_STACK_LIMIT); auto symbols = backtrace_symbols(call_stack, size); @@ -95,6 +93,9 @@ struct EnforceNotMet : public std::exception { } } free(symbols); +#else + sout << "Windows not support stack backtrace yet."; +#endif err_str_ = sout.str(); } } @@ -205,7 +206,7 @@ inline typename std::enable_if::type throw_on_error( #endif } -#ifndef __APPLE__ +#if !defined(__APPLE__) and !defined(_WIN32) template inline typename std::enable_if::type throw_on_error( ncclResult_t stat, const Args&... args) { @@ -221,7 +222,7 @@ inline typename std::enable_if::type throw_on_error( #endif } } -#endif // __APPLE__ +#endif // __APPLE__ and windows #endif // PADDLE_WITH_CUDA template @@ -263,7 +264,8 @@ inline void throw_on_error(T e) { * PADDLE_ENFORCE_EQ(a, b); * * will raise an expression described as follows: - * "enforce a == b failed, 1 != 2" with detailed stack information. + * "Enforce failed. Expected input a == b, but received a(1) != b(2)." + * with detailed stack information. * * extra messages is also supported, for example: * PADDLE_ENFORCE(a, b, "some simple enforce failed between %d numbers", 2) @@ -292,9 +294,10 @@ inline void throw_on_error(T e) { #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \ do { \ if (UNLIKELY(!((__VAL0)__CMP(__VAL1)))) { \ - PADDLE_THROW("enforce %s " #__CMP " %s failed, %s " #__INV_CMP \ - " %s\n%s", \ - #__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \ + PADDLE_THROW("Enforce failed. Expected %s " #__CMP \ + " %s, but received %s:%s " #__INV_CMP " %s:%s.\n%s", \ + #__VAL0, #__VAL1, #__VAL0, \ + paddle::string::to_string(__VAL0), #__VAL1, \ paddle::string::to_string(__VAL1), \ paddle::string::Sprintf("" __VA_ARGS__)); \ } \ diff --git a/paddle/fluid/platform/enforce_test.cc b/paddle/fluid/platform/enforce_test.cc index 0e8684581a..d521829655 100644 --- a/paddle/fluid/platform/enforce_test.cc +++ b/paddle/fluid/platform/enforce_test.cc @@ -54,7 +54,9 @@ TEST(ENFORCE_EQ, NO_EXTRA_MSG_FAIL) { PADDLE_ENFORCE_EQ(a, 1 + 3); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - HasPrefix(StringPiece(error.what()), "enforce a == 1 + 3 failed, 2 != 4"); + HasPrefix( + StringPiece(error.what()), + "Enforce failed. Expected a == 1 + 3, but received a:2 != 1 + 3:4."); } EXPECT_TRUE(caught_exception); } @@ -67,7 +69,8 @@ TEST(ENFORCE_EQ, EXTRA_MSG_FAIL) { } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; HasPrefix(StringPiece(error.what()), - "enforce a == 1 + 3 failed, 2 != 4\ntheir size not match"); + "Enforce failed. Expected a == 1 + 3, but received a:2 != 1 + " + "3:4.\ntheir size not match"); } EXPECT_TRUE(caught_exception); } @@ -84,8 +87,9 @@ TEST(ENFORCE_NE, FAIL) { PADDLE_ENFORCE_NE(1.0, 1UL); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE(HasPrefix(StringPiece(error.what()), - "enforce 1.0 != 1UL failed, 1 == 1")) + EXPECT_TRUE(HasPrefix( + StringPiece(error.what()), + "Enforce failed. Expected 1.0 != 1UL, but received 1.0:1 == 1UL:1.")) << error.what() << " does not have expected prefix"; } EXPECT_TRUE(caught_exception); @@ -98,8 +102,9 @@ TEST(ENFORCE_GT, FAIL) { PADDLE_ENFORCE_GT(1, 2UL); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE( - HasPrefix(StringPiece(error.what()), "enforce 1 > 2UL failed, 1 <= 2")); + EXPECT_TRUE(HasPrefix( + StringPiece(error.what()), + "Enforce failed. Expected 1 > 2UL, but received 1:1 <= 2UL:2.")); } EXPECT_TRUE(caught_exception); } @@ -116,8 +121,9 @@ TEST(ENFORCE_GE, FAIL) { PADDLE_ENFORCE_GE(1, 2UL); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE( - HasPrefix(StringPiece(error.what()), "enforce 1 >= 2UL failed, 1 < 2")); + EXPECT_TRUE(HasPrefix( + StringPiece(error.what()), + "Enforce failed. Expected 1 >= 2UL, but received 1:1 < 2UL:2.")); } EXPECT_TRUE(caught_exception); } @@ -135,8 +141,9 @@ TEST(ENFORCE_LE, FAIL) { PADDLE_ENFORCE_GT(1, 2UL); } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; - EXPECT_TRUE( - HasPrefix(StringPiece(error.what()), "enforce 1 > 2UL failed, 1 <= 2")); + EXPECT_TRUE(HasPrefix( + StringPiece(error.what()), + "Enforce failed. Expected 1 > 2UL, but received 1:1 <= 2UL:2.")); } EXPECT_TRUE(caught_exception); } @@ -153,7 +160,8 @@ TEST(ENFORCE_LT, FAIL) { } catch (paddle::platform::EnforceNotMet error) { caught_exception = true; EXPECT_TRUE(HasPrefix(StringPiece(error.what()), - "enforce 1UL < 0.12 failed, 1 >= 0.12")); + "Enforce failed. Expected 1UL < 0.12, but " + "received 1UL:1 >= 0.12:0.12.")); } EXPECT_TRUE(caught_exception); } diff --git a/paddle/fluid/platform/gpu_info.cc b/paddle/fluid/platform/gpu_info.cc index 4cee93f3a4..126636d879 100644 --- a/paddle/fluid/platform/gpu_info.cc +++ b/paddle/fluid/platform/gpu_info.cc @@ -116,7 +116,8 @@ size_t GpuMaxChunkSize() { size_t allocating = static_cast(FLAGS_fraction_of_gpu_memory_to_use * (total - reserving)); - PADDLE_ENFORCE_LE(allocating, available); + PADDLE_ENFORCE_LE(allocating, available, + "Insufficient GPU memory to allocation."); return allocating; } diff --git a/paddle/fluid/platform/mkldnn_helper.h b/paddle/fluid/platform/mkldnn_helper.h index 10a3ad256b..f6e9a52b27 100644 --- a/paddle/fluid/platform/mkldnn_helper.h +++ b/paddle/fluid/platform/mkldnn_helper.h @@ -125,6 +125,11 @@ class MKLDNNHandler { return this->AcquireMemory(md, ptr, "@user_weights_mem_p"); } + std::shared_ptr AcquireBiasMemory( + const mkldnn::memory::desc& md, void* ptr) { + return this->AcquireMemory(md, ptr, "@user_bias_mem_p"); + } + std::shared_ptr AcquireDstMemory( const mkldnn::memory::desc& md, void* ptr) { return this->AcquireMemory(md, ptr, "@user_dst_mem_p"); diff --git a/paddle/fluid/platform/port.h b/paddle/fluid/platform/port.h new file mode 100644 index 0000000000..a0a2d29500 --- /dev/null +++ b/paddle/fluid/platform/port.h @@ -0,0 +1,37 @@ +// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +#pragma once + +#include +#include + +#if !defined(_WIN32) +#include // for dladdr +#include // for backtrace +#else +#include +#include + +static void* dlsym(void* handle, const char* symbol_name) { + FARPROC found_symbol; + found_symbol = GetProcAddress((HMODULE)handle, symbol_name); + + if (found_symbol == NULL) { + throw std::runtime_error(std::string(symbol_name) + " not found."); + } + return reinterpret_cast(found_symbol); +} + +#endif diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt index 89ca4f7812..d6a14b3305 100644 --- a/paddle/fluid/pybind/CMakeLists.txt +++ b/paddle/fluid/pybind/CMakeLists.txt @@ -1,19 +1,22 @@ +set(PYBIND_DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method + ) +if(NOT WIN32) +list(APPEND PYBIND_DEPS parallel_executor) +endif() if(WITH_PYTHON) if(WITH_AMD_GPU) hip_library(paddle_pybind SHARED SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc - DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method - parallel_executor + DEPS ${PYBIND_DEPS} ${GLOB_OP_LIB}) else() cc_library(paddle_pybind SHARED SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc - DEPS pybind python proto_desc memory executor prune profiler feed_fetch_method - parallel_executor + DEPS ${PYBIND_DEPS} ${GLOB_OP_LIB}) - if(NOT APPLE AND NOT ANDROID) + if(NOT APPLE AND NOT ANDROID AND NOT WIN32) target_link_libraries(paddle_pybind rt) - endif(NOT APPLE AND NOT ANDROID) + endif(NOT APPLE AND NOT ANDROID AND NOT WIN32) endif(WITH_AMD_GPU) cc_test(tensor_py_test SRCS tensor_py_test.cc DEPS python) diff --git a/paddle/fluid/pybind/protobuf.cc b/paddle/fluid/pybind/protobuf.cc index be623703c2..c2137ec6d7 100644 --- a/paddle/fluid/pybind/protobuf.cc +++ b/paddle/fluid/pybind/protobuf.cc @@ -205,12 +205,7 @@ void BindBlockDesc(pybind11::module *m) { void BindVarDsec(pybind11::module *m) { pybind11::class_ var_desc(*m, "VarDesc", ""); var_desc - .def("name", - [](pd::VarDesc &self) { - pybind11::bytes name = self.Name(); - return name; - }, - pybind11::return_value_policy::reference) + .def("name", &pd::VarDesc::Name, pybind11::return_value_policy::reference) .def("set_name", &pd::VarDesc::SetName) .def("set_shape", &pd::VarDesc::SetShape) .def("set_shapes", &pd::VarDesc::SetShapes) diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc index 7127bb38f6..6773465923 100644 --- a/paddle/fluid/pybind/pybind.cc +++ b/paddle/fluid/pybind/pybind.cc @@ -54,6 +54,8 @@ limitations under the License. */ #include "paddle/fluid/platform/gpu_info.h" #endif +#include "pybind11/stl.h" + // disable auto conversion to list in Python PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray); @@ -247,6 +249,7 @@ PYBIND11_PLUGIN(core) { self.set_rows(new_rows); #endif }) + .def("sync_index", [](SelectedRows &instance) { instance.SyncIndex(); }) .def("rows", [](SelectedRows &self) { auto rows = self.rows(); std::vector new_rows; @@ -593,8 +596,8 @@ All parameter, weight, gradient are variables in Paddle. // -- python binds for parallel executor. py::class_ pe(m, "ParallelExecutor"); - py::class_(pe, "ExecutionStrategy") - .def(py::init()) + py::class_ exec_strategy(pe, "ExecutionStrategy"); + exec_strategy.def(py::init()) .def_property( "num_threads", [](const ExecutionStrategy &self) { return self.num_threads_; }, @@ -621,6 +624,16 @@ All parameter, weight, gradient are variables in Paddle. [](ExecutionStrategy &self, size_t num_iteration_per_drop_scope) { self.num_iteration_per_drop_scope_ = num_iteration_per_drop_scope; }); + exec_strategy.def_property( + "use_experimental_executor", + [](const ExecutionStrategy &self) { + return self.type_ == ExecutionStrategy::kExperimental; + }, + [](ExecutionStrategy &self, bool experimental) { + self.type_ = experimental ? ExecutionStrategy::kExperimental + : ExecutionStrategy::kDefault; + }); + py::class_ build_strategy(pe, "BuildStrategy"); py::enum_(build_strategy, "ReduceStrategy") diff --git a/paddle/scripts/submit_local.sh.in b/paddle/scripts/submit_local.sh.in index 1283de9d95..622a2d5104 100755 --- a/paddle/scripts/submit_local.sh.in +++ b/paddle/scripts/submit_local.sh.in @@ -54,7 +54,7 @@ function cpu_config() { if [ $platform == "Linux" ]; then ht=`lscpu |grep "per core"|awk -F':' '{print $2}'|xargs` elif [ $platform == "Darwin" ]; then - if [`sysctl -n hw.physicalcpu` -eq `sysctl -n hw.logicalcpu`]; then + if [ `sysctl -n hw.physicalcpu` -eq `sysctl -n hw.logicalcpu` ]; then # HT is OFF ht=1 fi diff --git a/python/CMakeLists.txt b/python/CMakeLists.txt index 2590081150..9cdcb87df5 100644 --- a/python/CMakeLists.txt +++ b/python/CMakeLists.txt @@ -97,10 +97,11 @@ if(APPLE) if(NOT INSTALL_NAME_TOOL_EXECUTABLE) message(FATAL_ERROR "install_name_tool not found, please check.\n") endif() -else(APPLE) +endif() +if(LINUX) find_program(PATCHELF_EXECUTABLE patchelf) if(NOT PATCHELF_EXECUTABLE) message(FATAL_ERROR "patchelf not found, please install it.\n" "For Ubuntu, the command is: apt-get install -y patchelf.") endif() -endif(APPLE) +endif(LINUX) diff --git a/python/paddle/__init__.py b/python/paddle/__init__.py index 241a07a352..53746afdb2 100644 --- a/python/paddle/__init__.py +++ b/python/paddle/__init__.py @@ -24,4 +24,5 @@ except ImportError: import paddle.reader import paddle.dataset import paddle.batch +import paddle.compat batch = batch.batch diff --git a/python/paddle/compat.py b/python/paddle/compat.py new file mode 100644 index 0000000000..50726b6fa1 --- /dev/null +++ b/python/paddle/compat.py @@ -0,0 +1,237 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import six +import math + +__all__ = [ + 'long_type', + 'to_text', + 'to_bytes', + 'round', + 'floor_division', + 'get_exception_message', +] + +if six.PY2: + int_type = int + long_type = long +else: + int_type = int + long_type = int + + +# str and bytes related functions +def to_text(obj, encoding='utf-8', inplace=False): + """ + All string in PaddlePaddle should be represented as a literal string. + This function will convert object to a literal string without any encoding. + Especially, if the object type is a list or set container, we will iterate + all items in the object and convert them to literal string. + + In Python3: + Decode the bytes type object to str type with specific encoding + + In Python2: + Decode the str type object to unicode type with specific encoding + + Args: + obj(unicode|str|bytes|list|set) : The object to be decoded. + encoding(str) : The encoding format to decode a string + inplace(bool) : If we change the original object or we create a new one + + Returns: + Decoded result of obj + """ + if obj is None: + return obj + + if isinstance(obj, list): + if inplace: + for i in six.moves.xrange(len(obj)): + obj[i] = _to_text(obj[i], encoding) + return obj + else: + return [_to_text(item, encoding) for item in obj] + elif isinstance(obj, set): + if inplace: + for item in obj: + obj.remove(item) + obj.add(_to_text(item, encoding)) + return obj + else: + return set([_to_text(item, encoding) for item in obj]) + else: + return _to_text(obj, encoding) + + +def _to_text(obj, encoding): + """ + In Python3: + Decode the bytes type object to str type with specific encoding + + In Python2: + Decode the str type object to unicode type with specific encoding, + or we just return the unicode string of object + + Args: + obj(unicode|str|bytes) : The object to be decoded. + encoding(str) : The encoding format + + Returns: + decoded result of obj + """ + if obj is None: + return obj + + if isinstance(obj, six.binary_type): + return obj.decode(encoding) + elif isinstance(obj, six.text_type): + return obj + else: + return six.u(obj) + + +def to_bytes(obj, encoding='utf-8', inplace=False): + """ + All string in PaddlePaddle should be represented as a literal string. + This function will convert object to a bytes with specific encoding. + Especially, if the object type is a list or set container, we will iterate + all items in the object and convert them to bytes. + + In Python3: + Encode the str type object to bytes type with specific encoding + + In Python2: + Encode the unicode type object to str type with specific encoding, + or we just return the 8-bit string of object + + Args: + obj(unicode|str|bytes|list|set) : The object to be encoded. + encoding(str) : The encoding format to encode a string + inplace(bool) : If we change the original object or we create a new one + + Returns: + Decoded result of obj + """ + if obj is None: + return obj + + if isinstance(obj, list): + if inplace: + for i in six.moves.xrange(len(obj)): + obj[i] = _to_bytes(obj[i], encoding) + return obj + else: + return [_to_bytes(item, encoding) for item in obj] + elif isinstance(obj, set): + if inplace: + for item in obj: + obj.remove(item) + obj.add(_to_bytes(item, encoding)) + return obj + else: + return set([_to_bytes(item, encoding) for item in obj]) + else: + return _to_bytes(obj, encoding) + + +def _to_bytes(obj, encoding): + """ + In Python3: + Encode the str type object to bytes type with specific encoding + + In Python2: + Encode the unicode type object to str type with specific encoding, + or we just return the 8-bit string of object + + Args: + obj(unicode|str|bytes) : The object to be encoded. + encoding(str) : The encoding format + + Returns: + encoded result of obj + """ + if obj is None: + return obj + + assert encoding is not None + if isinstance(obj, six.text_type): + return obj.encode(encoding) + elif isinstance(obj, six.binary_type): + return obj + else: + return six.b(obj) + + +# math related functions +def round(x, d=0): + """ + Compatible round which act the same behaviour in Python3. + + Args: + x(float) : The number to round halfway. + + Returns: + round result of x + """ + if six.PY3: + # The official walkaround of round in Python3 is incorrect + # we implement accroding this answer: https://www.techforgeek.info/round_python.html + if x > 0.0: + p = 10**d + return float(math.floor((x * p) + math.copysign(0.5, x))) / p + elif x < 0.0: + p = 10**d + return float(math.ceil((x * p) + math.copysign(0.5, x))) / p + else: + return math.copysign(0.0, x) + else: + import __builtin__ + return __builtin__.round(x, d) + + +def floor_division(x, y): + """ + Compatible division which act the same behaviour in Python3 and Python2, + whose result will be a int value of floor(x / y) in Python3 and value of + (x / y) in Python2. + + Args: + x(int|float) : The number to divide. + y(int|float) : The number to be divided + + Returns: + division result of x // y + """ + return x // y + + +# exception related functions +def get_exception_message(exc): + """ + Get the error message of a specific exception + + Args: + exec(Exception) : The exception to get error message. + + Returns: + the error message of exec + """ + assert exc is not None + + if six.PY2: + return exc.message + else: + return str(exc) diff --git a/python/paddle/dataset/cifar.py b/python/paddle/dataset/cifar.py index f6b4ff8fbd..b83fa78c4c 100644 --- a/python/paddle/dataset/cifar.py +++ b/python/paddle/dataset/cifar.py @@ -28,11 +28,13 @@ images per class. """ +from __future__ import print_function + import itertools import numpy import paddle.dataset.common import tarfile -from six.moves import zip +import six from six.moves import cPickle as pickle __all__ = ['train100', 'test100', 'train10', 'test10', 'convert'] @@ -46,10 +48,11 @@ CIFAR100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85' def reader_creator(filename, sub_name, cycle=False): def read_batch(batch): - data = batch['data'] - labels = batch.get('labels', batch.get('fine_labels', None)) + data = batch[six.b('data')] + labels = batch.get( + six.b('labels'), batch.get(six.b('fine_labels'), None)) assert labels is not None - for sample, label in zip(data, labels): + for sample, label in six.moves.zip(data, labels): yield (sample / 255.0).astype(numpy.float32), int(label) def reader(): @@ -59,7 +62,11 @@ def reader_creator(filename, sub_name, cycle=False): while True: for name in names: - batch = pickle.load(f.extractfile(name)) + if six.PY2: + batch = pickle.load(f.extractfile(name)) + else: + batch = pickle.load( + f.extractfile(name), encoding='bytes') for item in read_batch(batch): yield item if not cycle: diff --git a/python/paddle/dataset/common.py b/python/paddle/dataset/common.py index 6195cc50df..1d7ff582c8 100644 --- a/python/paddle/dataset/common.py +++ b/python/paddle/dataset/common.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import requests import hashlib import os @@ -85,10 +87,10 @@ def download(url, module_name, md5sum, save_name=None): total_length = r.headers.get('content-length') if total_length is None: - with open(filename, 'w') as f: + with open(filename, 'wb') as f: shutil.copyfileobj(r.raw, f) else: - with open(filename, 'w') as f: + with open(filename, 'wb') as f: dl = 0 total_length = int(total_length) for data in r.iter_content(chunk_size=4096): diff --git a/python/paddle/dataset/conll05.py b/python/paddle/dataset/conll05.py index a97c95d067..55cfd92721 100644 --- a/python/paddle/dataset/conll05.py +++ b/python/paddle/dataset/conll05.py @@ -20,15 +20,18 @@ dataset. And a pre-trained word vector model based on Wikipedia corpus is used to initialize SRL model. """ +from __future__ import print_function + import tarfile import gzip import itertools import paddle.dataset.common -from six.moves import zip +import paddle.compat as cpt +from six.moves import zip, range __all__ = ['test, get_dict', 'get_embedding', 'convert'] -DATA_URL = 'http://www.cs.upc.edu/~srlconll/conll05st-tests.tar.gz' +DATA_URL = 'http://paddlemodels.bj.bcebos.com/conll05st/conll05st-tests.tar.gz' DATA_MD5 = '387719152ae52d60422c016e92a742fc' WORDDICT_URL = 'http://paddlemodels.bj.bcebos.com/conll05st%2FwordDict.txt' WORDDICT_MD5 = 'ea7fb7d4c75cc6254716f0177a506baa' @@ -89,8 +92,8 @@ def corpus_reader(data_path, words_name, props_name): labels = [] one_seg = [] for word, label in zip(words_file, props_file): - word = word.strip() - label = label.strip().split() + word = cpt.to_text(word.strip()) + label = cpt.to_text(label.strip().split()) if len(label) == 0: # end of sentence for i in range(len(one_seg[0])): diff --git a/python/paddle/dataset/flowers.py b/python/paddle/dataset/flowers.py index 914dae348b..aa73bbaf70 100644 --- a/python/paddle/dataset/flowers.py +++ b/python/paddle/dataset/flowers.py @@ -28,6 +28,9 @@ Graphics and Image Processing (2008) http://www.robots.ox.ac.uk/~vgg/publications/papers/nilsback08.{pdf,ps.gz}. """ + +from __future__ import print_function + import itertools import functools from .common import download @@ -116,7 +119,7 @@ def reader_creator(data_file, for file in open(file_list): file = file.strip() batch = None - with open(file, 'r') as f: + with open(file, 'rb') as f: batch = pickle.load(f) data = batch['data'] labels = batch['label'] diff --git a/python/paddle/dataset/image.py b/python/paddle/dataset/image.py index 3b3d89c93c..1cd50bd180 100644 --- a/python/paddle/dataset/image.py +++ b/python/paddle/dataset/image.py @@ -29,10 +29,18 @@ the image layout as follows. formats can be used for training. Noted that, the format should be keep consistent between the training and inference peroid. """ + +from __future__ import print_function + import numpy as np try: import cv2 except ImportError: + import sys + sys.stderr.write( + '''Warning with paddle image module: opencv-python should be imported, + or paddle image module could NOT work; please install opencv-python first.''' + ) cv2 = None import os import tarfile @@ -56,7 +64,7 @@ def batch_images_from_tar(data_file, :type data_file: string :param dataset_name: 'train','test' or 'valid' :type dataset_name: string - :param img2label: a dic with image file name as key + :param img2label: a dic with image file name as key and image's label as value :type img2label: dic :param num_per_batch: image number per batch file @@ -88,7 +96,7 @@ def batch_images_from_tar(data_file, output['data'] = data pickle.dump( output, - open('%s/batch_%d' % (out_path, file_id), 'w'), + open('%s/batch_%d' % (out_path, file_id), 'wb'), protocol=pickle.HIGHEST_PROTOCOL) file_id += 1 data = [] @@ -99,7 +107,7 @@ def batch_images_from_tar(data_file, output['data'] = data pickle.dump( output, - open('%s/batch_%d' % (out_path, file_id), 'w'), + open('%s/batch_%d' % (out_path, file_id), 'wb'), protocol=pickle.HIGHEST_PROTOCOL) with open(meta_file, 'a') as meta: @@ -113,7 +121,7 @@ def load_image_bytes(bytes, is_color=True): Load an color or gray image from bytes array. Example usage: - + .. code-block:: python with open('cat.jpg') as f: @@ -126,6 +134,8 @@ def load_image_bytes(bytes, is_color=True): load and return a gray image. :type is_color: bool """ + assert cv2 is not None + flag = 1 if is_color else 0 file_bytes = np.asarray(bytearray(bytes), dtype=np.uint8) img = cv2.imdecode(file_bytes, flag) @@ -137,7 +147,7 @@ def load_image(file, is_color=True): Load an color or gray image from the file path. Example usage: - + .. code-block:: python im = load_image('cat.jpg') @@ -149,6 +159,8 @@ def load_image(file, is_color=True): load and return a gray image. :type is_color: bool """ + assert cv2 is not None + # cv2.IMAGE_COLOR for OpenCV3 # cv2.CV_LOAD_IMAGE_COLOR for older OpenCV Version # cv2.IMAGE_GRAYSCALE for OpenCV3 @@ -161,27 +173,29 @@ def load_image(file, is_color=True): def resize_short(im, size): - """ + """ Resize an image so that the length of shorter edge is size. Example usage: - + .. code-block:: python im = load_image('cat.jpg') im = resize_short(im, 256) - + :param im: the input image with HWC layout. :type im: ndarray :param size: the shorter edge size of image after resizing. :type size: int """ + assert cv2 is not None + h, w = im.shape[:2] h_new, w_new = size, size if h > w: - h_new = size * h / w + h_new = size * h // w else: - w_new = size * w / h + w_new = size * w // h im = cv2.resize(im, (h_new, w_new), interpolation=cv2.INTER_CUBIC) return im @@ -193,17 +207,17 @@ def to_chw(im, order=(2, 0, 1)): according the order (2,0,1). Example usage: - + .. code-block:: python im = load_image('cat.jpg') im = resize_short(im, 256) im = to_chw(im) - + :param im: the input image with HWC layout. :type im: ndarray :param order: the transposed order. - :type order: tuple|list + :type order: tuple|list """ assert len(im.shape) == len(order) im = im.transpose(order) @@ -215,11 +229,11 @@ def center_crop(im, size, is_color=True): Crop the center of image with size. Example usage: - + .. code-block:: python im = center_crop(im, 224) - + :param im: the input image with HWC layout. :type im: ndarray :param size: the cropping size. @@ -228,8 +242,8 @@ def center_crop(im, size, is_color=True): :type is_color: bool """ h, w = im.shape[:2] - h_start = (h - size) / 2 - w_start = (w - size) / 2 + h_start = (h - size) // 2 + w_start = (w - size) // 2 h_end, w_end = h_start + size, w_start + size if is_color: im = im[h_start:h_end, w_start:w_end, :] @@ -243,11 +257,11 @@ def random_crop(im, size, is_color=True): Randomly crop input image with size. Example usage: - + .. code-block:: python im = random_crop(im, 224) - + :param im: the input image with HWC layout. :type im: ndarray :param size: the cropping size. @@ -272,11 +286,11 @@ def left_right_flip(im, is_color=True): Return the flipped image. Example usage: - + .. code-block:: python im = left_right_flip(im) - + :param im: input image with HWC layout or HW layout for gray image :type im: ndarray :param is_color: whether input image is color or not @@ -299,7 +313,7 @@ def simple_transform(im, resizing, croping and flipping. Example usage: - + .. code-block:: python im = simple_transform(im, 256, 224, True) @@ -314,7 +328,7 @@ def simple_transform(im, :type is_train: bool :param is_color: whether the image is color or not. :type is_color: bool - :param mean: the mean values, which can be element-wise mean values or + :param mean: the mean values, which can be element-wise mean values or mean values per channel. :type mean: numpy array | list """ @@ -332,7 +346,7 @@ def simple_transform(im, im = im.astype('float32') if mean is not None: mean = np.array(mean, dtype=np.float32) - # mean value, may be one value per channel + # mean value, may be one value per channel if mean.ndim == 1 and is_color: mean = mean[:, np.newaxis, np.newaxis] elif mean.ndim == 1: @@ -357,7 +371,7 @@ def load_and_transform(filename, for the transform operations. Example usage: - + .. code-block:: python im = load_and_transform('cat.jpg', 256, 224, True) @@ -372,7 +386,7 @@ def load_and_transform(filename, :type is_train: bool :param is_color: whether the image is color or not. :type is_color: bool - :param mean: the mean values, which can be element-wise mean values or + :param mean: the mean values, which can be element-wise mean values or mean values per channel. :type mean: numpy array | list """ diff --git a/python/paddle/dataset/imdb.py b/python/paddle/dataset/imdb.py index e7fe4e0b7e..fd92523a94 100644 --- a/python/paddle/dataset/imdb.py +++ b/python/paddle/dataset/imdb.py @@ -20,11 +20,14 @@ of 25,000 highly polar movie reviews for training, and 25,000 for testing. Besides, this module also provides API for building dictionary. """ +from __future__ import print_function + import paddle.dataset.common import collections import tarfile import re import string +import six __all__ = ['build_dict', 'train', 'test', 'convert'] @@ -42,13 +45,14 @@ def tokenize(pattern): # sequential access of member files, other than # tarfile.extractfile, which does random access and might # destroy hard disks. - tf = next(tarf) + tf = tarf.next() while tf != None: if bool(pattern.match(tf.name)): # newline and punctuations removal and ad-hoc tokenization. - yield tarf.extractfile(tf).read().rstrip("\n\r").translate( - None, string.punctuation).lower().split() - tf = next(tarf) + yield tarf.extractfile(tf).read().rstrip(six.b( + "\n\r")).translate( + None, six.b(string.punctuation)).lower().split() + tf = tarf.next() def build_dict(pattern, cutoff): @@ -62,11 +66,11 @@ def build_dict(pattern, cutoff): word_freq[word] += 1 # Not sure if we should prune less-frequent words here. - word_freq = [x for x in list(word_freq.items()) if x[1] > cutoff] + word_freq = [x for x in six.iteritems(word_freq) if x[1] > cutoff] dictionary = sorted(word_freq, key=lambda x: (-x[1], x[0])) words, _ = list(zip(*dictionary)) - word_idx = dict(list(zip(words, list(range(len(words)))))) + word_idx = dict(list(zip(words, six.moves.range(len(words))))) word_idx[''] = len(words) return word_idx diff --git a/python/paddle/dataset/imikolov.py b/python/paddle/dataset/imikolov.py index bc007c9d3c..8eecb75231 100644 --- a/python/paddle/dataset/imikolov.py +++ b/python/paddle/dataset/imikolov.py @@ -14,13 +14,17 @@ """ imikolov's simple dataset. -This module will download dataset from +This module will download dataset from http://www.fit.vutbr.cz/~imikolov/rnnlm/ and parse training set and test set into paddle reader creators. """ + +from __future__ import print_function + import paddle.dataset.common import collections import tarfile +import six __all__ = ['train', 'test', 'build_dict', 'convert'] @@ -64,11 +68,13 @@ def build_dict(min_word_freq=50): # remove for now, since we will set it as last index del word_freq[''] - word_freq = [x for x in list(word_freq.items()) if x[1] > min_word_freq] + word_freq = [ + x for x in six.iteritems(word_freq) if x[1] > min_word_freq + ] word_freq_sorted = sorted(word_freq, key=lambda x: (-x[1], x[0])) words, _ = list(zip(*word_freq_sorted)) - word_idx = dict(list(zip(words, list(range(len(words)))))) + word_idx = dict(list(zip(words, six.moves.range(len(words))))) word_idx[''] = len(words) return word_idx @@ -89,7 +95,7 @@ def reader_creator(filename, word_idx, n, data_type): l = [''] + l.strip().split() + [''] if len(l) >= n: l = [word_idx.get(w, UNK) for w in l] - for i in range(n, len(l) + 1): + for i in six.moves.range(n, len(l) + 1): yield tuple(l[i - n:i]) elif DataType.SEQ == data_type: l = l.strip().split() diff --git a/python/paddle/dataset/mnist.py b/python/paddle/dataset/mnist.py index ffa9008c80..38addd0cfd 100644 --- a/python/paddle/dataset/mnist.py +++ b/python/paddle/dataset/mnist.py @@ -17,10 +17,15 @@ MNIST dataset. This module will download dataset from http://yann.lecun.com/exdb/mnist/ and parse training set and test set into paddle reader creators. """ + +from __future__ import print_function + import paddle.dataset.common import subprocess import numpy import platform +import tempfile +from six.moves import range __all__ = ['train', 'test', 'convert'] URL_PREFIX = 'http://yann.lecun.com/exdb/mnist/' @@ -45,23 +50,28 @@ def reader_creator(image_filename, label_filename, buffer_size): # According to http://stackoverflow.com/a/38061619/724872, we # cannot use standard package gzip here. - m = subprocess.Popen([zcat_cmd, image_filename], stdout=subprocess.PIPE) - m.stdout.read(16) # skip some magic bytes + tmp_image_file = tempfile.TemporaryFile(prefix='paddle_dataset') + m = subprocess.Popen( + [zcat_cmd, image_filename], stdout=tmp_image_file).communicate() + tmp_image_file.seek(16) # skip some magic bytes - l = subprocess.Popen([zcat_cmd, label_filename], stdout=subprocess.PIPE) - l.stdout.read(8) # skip some magic bytes + # Python3 will not take stdout as file + tmp_label_file = tempfile.TemporaryFile(prefix='paddle_dataset') + l = subprocess.Popen( + [zcat_cmd, label_filename], stdout=tmp_label_file).communicate() + tmp_label_file.seek(8) # skip some magic bytes try: # reader could be break. while True: labels = numpy.fromfile( - l.stdout, 'ubyte', count=buffer_size).astype("int") + tmp_label_file, 'ubyte', count=buffer_size).astype("int") if labels.size != buffer_size: break # numpy.fromfile returns empty slice after EOF. images = numpy.fromfile( - m.stdout, 'ubyte', count=buffer_size * 28 * 28).reshape( - (buffer_size, 28 * 28)).astype('float32') + tmp_image_file, 'ubyte', count=buffer_size * 28 * + 28).reshape((buffer_size, 28 * 28)).astype('float32') images = images / 255.0 * 2.0 - 1.0 diff --git a/python/paddle/dataset/movielens.py b/python/paddle/dataset/movielens.py index 056ec21786..c98e0019f7 100644 --- a/python/paddle/dataset/movielens.py +++ b/python/paddle/dataset/movielens.py @@ -22,11 +22,15 @@ set and test set into paddle reader creators. """ +from __future__ import print_function + import zipfile import paddle.dataset.common import re import random import functools +import six +import paddle.compat as cpt __all__ = [ 'train', 'test', 'get_movie_title_dict', 'max_movie_id', 'max_user_id', @@ -112,6 +116,7 @@ def __initialize_meta_info__(): categories_set = set() with package.open('ml-1m/movies.dat') as movie_file: for i, line in enumerate(movie_file): + line = cpt.to_text(line, encoding='latin') movie_id, title, categories = line.strip().split('::') categories = categories.split('|') for c in categories: @@ -136,6 +141,7 @@ def __initialize_meta_info__(): USER_INFO = dict() with package.open('ml-1m/users.dat') as user_file: for line in user_file: + line = cpt.to_text(line, encoding='latin') uid, gender, age, job, _ = line.strip().split("::") USER_INFO[int(uid)] = UserInfo( index=uid, gender=gender, age=age, job_id=job) @@ -148,6 +154,7 @@ def __reader__(rand_seed=0, test_ratio=0.1, is_test=False): with zipfile.ZipFile(file=fn) as package: with package.open('ml-1m/ratings.dat') as rating: for line in rating: + line = cpt.to_text(line, encoding='latin') if (rand.random() < test_ratio) == is_test: uid, mov_id, rating, _ = line.strip().split("::") uid = int(uid) @@ -187,7 +194,7 @@ def max_movie_id(): Get the maximum value of movie id. """ __initialize_meta_info__() - return reduce(__max_index_info__, list(MOVIE_INFO.values())).index + return six.moves.reduce(__max_index_info__, list(MOVIE_INFO.values())).index def max_user_id(): @@ -195,7 +202,7 @@ def max_user_id(): Get the maximum value of user id. """ __initialize_meta_info__() - return reduce(__max_index_info__, list(USER_INFO.values())).index + return six.moves.reduce(__max_index_info__, list(USER_INFO.values())).index def __max_job_id_impl__(a, b): @@ -210,7 +217,8 @@ def max_job_id(): Get the maximum value of job id. """ __initialize_meta_info__() - return reduce(__max_job_id_impl__, list(USER_INFO.values())).job_id + return six.moves.reduce(__max_job_id_impl__, + list(USER_INFO.values())).job_id def movie_categories(): diff --git a/python/paddle/dataset/mq2007.py b/python/paddle/dataset/mq2007.py index cc4d088316..d5740f30c8 100644 --- a/python/paddle/dataset/mq2007.py +++ b/python/paddle/dataset/mq2007.py @@ -23,6 +23,8 @@ http://research.microsoft.com/en-us/um/beijing/projects/letor/LETOR4.0/Data/MQ20 """ +from __future__ import print_function + import os import functools import rarfile diff --git a/python/paddle/dataset/sentiment.py b/python/paddle/dataset/sentiment.py index 953ada057b..22d867beea 100644 --- a/python/paddle/dataset/sentiment.py +++ b/python/paddle/dataset/sentiment.py @@ -20,6 +20,9 @@ The script fetch and preprocess movie_reviews data set that provided by NLTK TODO(yuyang18): Complete dataset. """ +from __future__ import print_function + +import six import collections from itertools import chain @@ -64,7 +67,7 @@ def get_word_dict(): for field in movie_reviews.fileids(category): for words in movie_reviews.words(field): word_freq_dict[words] += 1 - words_sort_list = list(word_freq_dict.items()) + words_sort_list = six.iteritems(word_freq_dict) words_sort_list.sort(cmp=lambda a, b: b[1] - a[1]) for index, word in enumerate(words_sort_list): words_freq_sorted.append((word[0], index)) diff --git a/python/paddle/dataset/tests/cifar_test.py b/python/paddle/dataset/tests/cifar_test.py index 839125b09d..8e514f0fd9 100644 --- a/python/paddle/dataset/tests/cifar_test.py +++ b/python/paddle/dataset/tests/cifar_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.cifar import unittest diff --git a/python/paddle/dataset/tests/common_test.py b/python/paddle/dataset/tests/common_test.py index 777cd06a19..0ce7d83f37 100644 --- a/python/paddle/dataset/tests/common_test.py +++ b/python/paddle/dataset/tests/common_test.py @@ -12,10 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.common import unittest import tempfile import glob +from six.moves import range class TestCommon(unittest.TestCase): diff --git a/python/paddle/dataset/tests/flowers_test.py b/python/paddle/dataset/tests/flowers_test.py index 06260fd796..06a0a7761c 100644 --- a/python/paddle/dataset/tests/flowers_test.py +++ b/python/paddle/dataset/tests/flowers_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.flowers import unittest diff --git a/python/paddle/dataset/tests/imdb_test.py b/python/paddle/dataset/tests/imdb_test.py index 539da04944..415947e347 100644 --- a/python/paddle/dataset/tests/imdb_test.py +++ b/python/paddle/dataset/tests/imdb_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.imdb import unittest import re diff --git a/python/paddle/dataset/tests/imikolov_test.py b/python/paddle/dataset/tests/imikolov_test.py index 50f50d947d..1f78a5dd4d 100644 --- a/python/paddle/dataset/tests/imikolov_test.py +++ b/python/paddle/dataset/tests/imikolov_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.imikolov import unittest diff --git a/python/paddle/dataset/tests/mnist_test.py b/python/paddle/dataset/tests/mnist_test.py index 8ada19d3f2..fbb5d92649 100644 --- a/python/paddle/dataset/tests/mnist_test.py +++ b/python/paddle/dataset/tests/mnist_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.mnist import unittest diff --git a/python/paddle/dataset/tests/mq2007_test.py b/python/paddle/dataset/tests/mq2007_test.py index fba388724a..ee0897e88f 100644 --- a/python/paddle/dataset/tests/mq2007_test.py +++ b/python/paddle/dataset/tests/mq2007_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.mq2007 import unittest diff --git a/python/paddle/dataset/tests/test_image.py b/python/paddle/dataset/tests/test_image.py index 8bd56607ae..32d2eb17ae 100644 --- a/python/paddle/dataset/tests/test_image.py +++ b/python/paddle/dataset/tests/test_image.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/dataset/tests/test_sentiment.py b/python/paddle/dataset/tests/test_sentiment.py index 37326517f7..bb9830132e 100644 --- a/python/paddle/dataset/tests/test_sentiment.py +++ b/python/paddle/dataset/tests/test_sentiment.py @@ -15,6 +15,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import nltk import paddle.dataset.sentiment as st diff --git a/python/paddle/dataset/tests/voc2012_test.py b/python/paddle/dataset/tests/voc2012_test.py index 0d285461a8..cddeb91cab 100644 --- a/python/paddle/dataset/tests/voc2012_test.py +++ b/python/paddle/dataset/tests/voc2012_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.voc2012 import unittest diff --git a/python/paddle/dataset/tests/wmt16_test.py b/python/paddle/dataset/tests/wmt16_test.py index 8b949d8bf5..be121bb101 100644 --- a/python/paddle/dataset/tests/wmt16_test.py +++ b/python/paddle/dataset/tests/wmt16_test.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.wmt16 import unittest diff --git a/python/paddle/dataset/uci_housing.py b/python/paddle/dataset/uci_housing.py index 410ca7af0d..f87fdcc4f0 100644 --- a/python/paddle/dataset/uci_housing.py +++ b/python/paddle/dataset/uci_housing.py @@ -19,9 +19,10 @@ https://archive.ics.uci.edu/ml/machine-learning-databases/housing/ and parse training set and test set into paddle reader creators. """ -import os +from __future__ import print_function import numpy as np +import six import tempfile import tarfile import os @@ -70,11 +71,11 @@ def load_data(filename, feature_num=14, ratio=0.8): return data = np.fromfile(filename, sep=' ') - data = data.reshape(data.shape[0] / feature_num, feature_num) + data = data.reshape(data.shape[0] // feature_num, feature_num) maximums, minimums, avgs = data.max(axis=0), data.min(axis=0), data.sum( axis=0) / data.shape[0] feature_range(maximums[:-1], minimums[:-1]) - for i in range(feature_num - 1): + for i in six.moves.range(feature_num - 1): data[:, i] = (data[:, i] - avgs[i]) / (maximums[i] - minimums[i]) offset = int(data.shape[0] * ratio) UCI_TRAIN_DATA = data[:offset] @@ -137,7 +138,7 @@ def predict_reader(): It returns just one tuple data to do inference. :return: one tuple data - :rtype: tuple + :rtype: tuple """ global UCI_TEST_DATA load_data(paddle.dataset.common.download(URL, 'uci_housing', MD5)) diff --git a/python/paddle/dataset/voc2012.py b/python/paddle/dataset/voc2012.py index 9c945574db..5068893765 100644 --- a/python/paddle/dataset/voc2012.py +++ b/python/paddle/dataset/voc2012.py @@ -19,6 +19,8 @@ to training/test sets has been maintained. The total number of images with segmentation has been increased from 7,062 to 9,993. """ +from __future__ import print_function + import tarfile import io import numpy as np diff --git a/python/paddle/dataset/wmt14.py b/python/paddle/dataset/wmt14.py index 7504474591..f8c1a33574 100644 --- a/python/paddle/dataset/wmt14.py +++ b/python/paddle/dataset/wmt14.py @@ -19,10 +19,15 @@ http://paddlepaddle.cdn.bcebos.com/demo/wmt_shrinked_data/wmt14.tgz and parse training set and test set into paddle reader creators. """ + +from __future__ import print_function + +import six import tarfile import gzip import paddle.dataset.common +import paddle.compat as cpt __all__ = [ 'train', @@ -53,7 +58,7 @@ def __read_to_dict(tar_file, dict_size): out_dict = dict() for line_count, line in enumerate(fd): if line_count < size: - out_dict[line.strip()] = line_count + out_dict[cpt.to_text(line.strip())] = line_count else: break return out_dict @@ -84,7 +89,7 @@ def reader_creator(tar_file, file_name, dict_size): ] for name in names: for line in f.extractfile(name): - line_split = line.strip().split('\t') + line_split = line.strip().split(six.b('\t')) if len(line_split) != 2: continue src_seq = line_split[0] # one source sequence @@ -153,8 +158,8 @@ def get_dict(dict_size, reverse=True): tar_file = paddle.dataset.common.download(URL_TRAIN, 'wmt14', MD5_TRAIN) src_dict, trg_dict = __read_to_dict(tar_file, dict_size) if reverse: - src_dict = {v: k for k, v in list(src_dict.items())} - trg_dict = {v: k for k, v in list(trg_dict.items())} + src_dict = {v: k for k, v in six.iteritems(src_dict)} + trg_dict = {v: k for k, v in six.iteritems(trg_dict)} return src_dict, trg_dict diff --git a/python/paddle/dataset/wmt16.py b/python/paddle/dataset/wmt16.py index 4e3c466c38..f30dcd518e 100644 --- a/python/paddle/dataset/wmt16.py +++ b/python/paddle/dataset/wmt16.py @@ -28,12 +28,16 @@ Multi30K: Multilingual English-German Image Descriptions. } """ +from __future__ import print_function + import os +import six import tarfile import gzip from collections import defaultdict import paddle.dataset.common +import paddle.compat as cpt __all__ = [ "train", @@ -60,7 +64,7 @@ def __build_dict(tar_file, dict_size, save_path, lang): word_dict = defaultdict(int) with tarfile.open(tar_file, mode="r") as f: for line in f.extractfile("wmt16/train"): - line_split = line.strip().split("\t") + line_split = line.strip().split(six.b("\t")) if len(line_split) != 2: continue sen = line_split[0] if lang == "en" else line_split[1] for w in sen.split(): @@ -70,8 +74,7 @@ def __build_dict(tar_file, dict_size, save_path, lang): fout.write("%s\n%s\n%s\n" % (START_MARK, END_MARK, UNK_MARK)) for idx, word in enumerate( sorted( - iter(list(word_dict.items())), - key=lambda x: x[1], + six.iteritems(word_dict), key=lambda x: x[1], reverse=True)): if idx + 3 == dict_size: break fout.write("%s\n" % (word[0])) @@ -81,16 +84,16 @@ def __load_dict(tar_file, dict_size, lang, reverse=False): dict_path = os.path.join(paddle.dataset.common.DATA_HOME, "wmt16/%s_%d.dict" % (lang, dict_size)) if not os.path.exists(dict_path) or ( - len(open(dict_path, "r").readlines()) != dict_size): + len(open(dict_path, "rb").readlines()) != dict_size): __build_dict(tar_file, dict_size, dict_path, lang) word_dict = {} - with open(dict_path, "r") as fdict: + with open(dict_path, "rb") as fdict: for idx, line in enumerate(fdict): if reverse: - word_dict[idx] = line.strip() + word_dict[idx] = cpt.to_text(line.strip()) else: - word_dict[line.strip()] = idx + word_dict[cpt.to_text(line.strip())] = idx return word_dict @@ -120,7 +123,7 @@ def reader_creator(tar_file, file_name, src_dict_size, trg_dict_size, src_lang): with tarfile.open(tar_file, mode="r") as f: for line in f.extractfile(file_name): - line_split = line.strip().split("\t") + line_split = line.strip().split(six.b("\t")) if len(line_split) != 2: continue src_words = line_split[src_col].split() diff --git a/python/paddle/fluid/average.py b/python/paddle/fluid/average.py index 358e24df31..42cd3b3642 100644 --- a/python/paddle/fluid/average.py +++ b/python/paddle/fluid/average.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import warnings """ diff --git a/python/paddle/fluid/backward.py b/python/paddle/fluid/backward.py index fd6a76dd0c..a415cdbeaa 100644 --- a/python/paddle/fluid/backward.py +++ b/python/paddle/fluid/backward.py @@ -12,11 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from paddle.fluid import framework as framework from . import core import collections import copy import six +from .. import compat as cpt from . import unique_name __all__ = ['append_backward'] @@ -45,13 +48,13 @@ def _create_op_desc_(op_type, inputs, outputs, attrs): """ op_desc = core.OpDesc() op_desc.set_type(op_type) - for para, args in list(inputs.items()): + for para, args in six.iteritems(inputs): op_desc.set_input( para, list( map(lambda arg: arg.decode() if isinstance(arg, six.binary_type) else arg, args))) - for para, args in list(outputs.items()): + for para, args in six.iteritems(outputs): op_desc.set_output( para, list( @@ -63,7 +66,7 @@ def _create_op_desc_(op_type, inputs, outputs, attrs): if op_role_attr_name not in attrs: attrs[ op_role_attr_name] = core.op_proto_and_checker_maker.OpRole.Backward - for name, val in list(attrs.items()): + for name, val in six.iteritems(attrs): if isinstance(val, framework.Block): op_desc.set_block_attr(name, val.desc) else: @@ -75,10 +78,10 @@ def _infer_var_data_type_(grad_var_name, block): """ Infer the data type of given grad variable """ - grad_var = block.desc.find_var(grad_var_name.encode("ascii")) - fwd_name = _strip_grad_suffix_(grad_var_name.encode("ascii")) - if block.desc.has_var_recursive(fwd_name): - fwd_var = block.desc.find_var_recursive(fwd_name.encode("ascii")) + grad_var = block.desc.find_var(cpt.to_bytes(grad_var_name)) + fwd_name = _strip_grad_suffix_(grad_var_name) + if block.desc.has_var_recursive(cpt.to_bytes(fwd_name)): + fwd_var = block.desc.find_var_recursive(cpt.to_bytes(fwd_name)) grad_var.set_dtype(fwd_var.dtype()) else: grad_var.set_dtype(core.VarDesc.VarType.FP32) @@ -102,8 +105,10 @@ def _some_in_set_(cands, s): """ if len(cands) == 0: return False - for c in cands: - if c in s: + literal_set = cpt.to_text(s) + literal_cands = cpt.to_text(cands) + for c in literal_cands: + if c in literal_set: return True return False @@ -114,9 +119,8 @@ def _strip_grad_suffix_(name): e.g. x@GRAD ==> x y@GRAD@RENAME@1 ==> y """ - if isinstance(name, six.text_type): - name = name.encode() - pos = name.find(six.b(core.grad_var_suffix())) + name = cpt.to_text(name) + pos = name.find(core.grad_var_suffix()) return name[:pos] if pos != -1 else name @@ -125,9 +129,7 @@ def _append_grad_suffix_(name): Append grad suffix to the given variable name e.g. x ==> x@GRAD """ - if isinstance(name, six.text_type): - name = name.encode() - return name + six.b(core.grad_var_suffix()) + return cpt.to_text(name) + core.grad_var_suffix() def _addup_repetitive_outputs_(op_descs): @@ -187,7 +189,7 @@ def _addup_repetitive_outputs_(op_descs): op_desc.set_output(param_name, arg_names) renamed_vars[var_name].append(new_name) - for var_name, inputs in list(renamed_vars.items()): + for var_name, inputs in six.iteritems(renamed_vars): if len(inputs) > 1: pending_sum_ops.append( (_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]}, @@ -243,7 +245,7 @@ from .proto import framework_pb2 def serialize_op_decs(op_desc): protostr = op_desc.serialize_to_string() - proto = framework_pb2.OpDesc.FromString(str(protostr)) + proto = framework_pb2.OpDesc.FromString(six.binary_type(protostr)) return proto.__str__() @@ -364,7 +366,7 @@ def _append_backward_ops_(block, # Getting op's corresponding grad_op grad_op_desc, op_grad_to_var = core.get_grad_op_desc( - op.desc, no_grad_dict[block.idx], grad_sub_block_list) + op.desc, cpt.to_text(no_grad_dict[block.idx]), grad_sub_block_list) grad_op_descs.extend(grad_op_desc) grad_to_var.update(op_grad_to_var) @@ -411,11 +413,10 @@ def _append_backward_vars_(block, start_op_idx, grad_to_var, grad_info_map): new_vars = set() # create new gradient variables for grad_var_name in op_desc.output_arg_names(): - grad_var_name = grad_var_name.encode("ascii") - if block.desc.has_var_recursive( - grad_var_name) or grad_var_name == core.empty_var_name(): + if block.desc.has_var_recursive(cpt.to_bytes( + grad_var_name)) or grad_var_name == core.empty_var_name(): continue - block.desc.var(grad_var_name) + block.desc.var(cpt.to_bytes(grad_var_name)) new_vars.add(grad_var_name) if grad_var_name not in grad_to_var: continue @@ -445,7 +446,7 @@ def _rename_grad_(block, start_op_idx, grad_to_var, target_grad_map): op_desc.rename_output(name, new_name) var_map[name] = new_name - for g, ng in list(var_map.items()): + for g, ng in six.iteritems(var_map): if g in grad_to_var: grad_to_var[ng] = grad_to_var[g] grad_to_var.pop(g) @@ -595,11 +596,12 @@ def append_backward(loss, parameter_list=None, no_grad_set=None, parameters = parameter_list else: params = program.global_block().all_parameters() + program.global_block().iter_parameters() parameters = [param.name for param in params] params_and_grads = [] for param in parameters: - if param not in grad_info_map: + if cpt.to_text(param) not in grad_info_map: continue grad_info = grad_info_map[param] grad_block = grad_info[1] diff --git a/python/paddle/fluid/clip.py b/python/paddle/fluid/clip.py index 4b0a792f78..ba7ba3b5e9 100644 --- a/python/paddle/fluid/clip.py +++ b/python/paddle/fluid/clip.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import copy import six diff --git a/python/paddle/fluid/concurrency.py b/python/paddle/fluid/concurrency.py index 676a52a917..b4a06f23a6 100644 --- a/python/paddle/fluid/concurrency.py +++ b/python/paddle/fluid/concurrency.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from .layers.control_flow import BlockGuard, equal from .framework import Operator from .layer_helper import LayerHelper, unique_name diff --git a/python/paddle/fluid/contrib/__init__.py b/python/paddle/fluid/contrib/__init__.py index 58f2da1c3b..5607f11932 100644 --- a/python/paddle/fluid/contrib/__init__.py +++ b/python/paddle/fluid/contrib/__init__.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import decoder from .decoder import * from . import memory_usage_calc diff --git a/python/paddle/fluid/contrib/decoder/__init__.py b/python/paddle/fluid/contrib/decoder/__init__.py index 6343c1543d..9f973fd3c9 100644 --- a/python/paddle/fluid/contrib/decoder/__init__.py +++ b/python/paddle/fluid/contrib/decoder/__init__.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import beam_search_decoder from .beam_search_decoder import * diff --git a/python/paddle/fluid/contrib/decoder/beam_search_decoder.py b/python/paddle/fluid/contrib/decoder/beam_search_decoder.py index d268a948f7..f2b7ac8375 100644 --- a/python/paddle/fluid/contrib/decoder/beam_search_decoder.py +++ b/python/paddle/fluid/contrib/decoder/beam_search_decoder.py @@ -20,6 +20,8 @@ without using the low level API such as while ops. This API is still under active development and may change drastically. """ +from __future__ import print_function + import contextlib import numpy as np import six diff --git a/python/paddle/fluid/contrib/memory_usage_calc.py b/python/paddle/fluid/contrib/memory_usage_calc.py index 5da846edb6..09721e430b 100644 --- a/python/paddle/fluid/contrib/memory_usage_calc.py +++ b/python/paddle/fluid/contrib/memory_usage_calc.py @@ -14,12 +14,16 @@ """ This module privides a memory usage calculate function for user. The purpose of this API is to allow users to estimate memory usage of -a program under a special batch size, then user can set appropriate -batch size to fully utilize a GPU. +a program under a special batch size, then user can set appropriate +batch size to fully utilize a GPU. This API is still under active development and may change drastically. """ +from __future__ import print_function + +import six + from .. import core from ..framework import Program, Variable @@ -45,15 +49,15 @@ def memory_usage(program, batch_size): Args: program(Program): The current Program. - batch_size(int): The current input data batch_size. - + batch_size(int): The current input data batch_size. + Returns: min_total_memory(float): the estimate memory usage lower bound. max_total_memory(float): the estimate memory usage upper bound. unit_str(string): the unit of estimate usage result. - + Examples: - + >>> import paddle.fluid as fluid >>> lower_usage, upper_usage, unit = fluid.contrib.memory_usage( fluid.default_main_program(), batch_size=10) @@ -72,7 +76,7 @@ def memory_usage(program, batch_size): # Get the var_name list of first block and calculate total_memory = 0.0 - for var in program.global_block().vars.itervalues(): + for var in six.itervalues(program.global_block().vars): data_count = 1 for x in var.shape: if x == -1: @@ -81,10 +85,10 @@ def memory_usage(program, batch_size): data_count *= x var_memory = data_count * dtype_to_size[var.dtype] if DEBUG: - print "%s memory usage: %d" % (var.name, var_memory) + print("%s memory usage: %d" % (var.name, var_memory)) total_memory += var_memory if DEBUG: - print "total memory usage: %.2f" % (total_memory) + print("total memory usage: %.2f" % (total_memory)) # Convert appropriate unit unit_str = "B" diff --git a/python/paddle/fluid/data_feeder.py b/python/paddle/fluid/data_feeder.py index 9452cf0e2a..631bbfe1fe 100644 --- a/python/paddle/fluid/data_feeder.py +++ b/python/paddle/fluid/data_feeder.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import core import numpy import os diff --git a/python/paddle/fluid/debugger.py b/python/paddle/fluid/debugger.py index b7a92cf044..63060a77d1 100644 --- a/python/paddle/fluid/debugger.py +++ b/python/paddle/fluid/debugger.py @@ -12,7 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import sys +import six import re from .graphviz import GraphPreviewGenerator from .proto import framework_pb2 @@ -225,7 +228,7 @@ def draw_block_graphviz(block, highlights=None, path="./temp.dot"): graph = GraphPreviewGenerator("some graph") # collect parameters and args protostr = block.desc.serialize_to_string() - desc = framework_pb2.BlockDesc.FromString(str(protostr)) + desc = framework_pb2.BlockDesc.FromString(six.binary_type(protostr)) def need_highlight(name): if highlights is None: return False diff --git a/python/paddle/fluid/default_scope_funcs.py b/python/paddle/fluid/default_scope_funcs.py index f8faf69425..a5b2c84dfe 100644 --- a/python/paddle/fluid/default_scope_funcs.py +++ b/python/paddle/fluid/default_scope_funcs.py @@ -26,6 +26,8 @@ A `scoped_function` will take a `function` as input. That function will be invoked in a new local scope. """ +from __future__ import print_function + import paddle.fluid.core import threading diff --git a/python/paddle/fluid/evaluator.py b/python/paddle/fluid/evaluator.py index c0671cce9a..7a82038ff7 100644 --- a/python/paddle/fluid/evaluator.py +++ b/python/paddle/fluid/evaluator.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import warnings import numpy as np diff --git a/python/paddle/fluid/executor.py b/python/paddle/fluid/executor.py index e24b9faae2..288951cd7c 100644 --- a/python/paddle/fluid/executor.py +++ b/python/paddle/fluid/executor.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import contextlib import six @@ -320,8 +322,9 @@ class Executor(object): # append fetch_operators if not has_fetch_operators(global_block, fetch_list, fetch_var_name): for i, var in enumerate(fetch_list): - assert isinstance(var, Variable) or isinstance(var, str), ( - "Wrong type for fetch_list[%s]: %s" % (i, type(var))) + assert isinstance(var, Variable) or isinstance( + var, six.string_types), ( + "Wrong type for fetch_list[%s]: %s" % (i, type(var))) global_block.append_op( type='fetch', inputs={'X': [var]}, @@ -346,7 +349,7 @@ class Executor(object): def _fetch_data(self, fetch_list, fetch_var_name, scope): outs = [ core.get_fetch_variable(scope, fetch_var_name, i) - for i in range(len(fetch_list)) + for i in six.moves.range(len(fetch_list)) ] return outs diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py index 45b3abb88c..62682d1032 100644 --- a/python/paddle/fluid/framework.py +++ b/python/paddle/fluid/framework.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import collections import contextlib import re @@ -19,6 +21,7 @@ import six import numpy as np +from .. import compat as cpt from .proto import framework_pb2 try: from . import core @@ -27,7 +30,7 @@ except ImportError as e: """NOTE: You may need to run \"export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH\" if you encounters \"libmkldnn.so not found\" errors. If you have python installed in other directory, replace \"/usr/local/lib\" with your own - directory. The original error is: \n""" + e.message) + directory. The original error is: \n""" + cpt.get_exception_message(e)) except Exception as e: raise e from . import unique_name @@ -87,7 +90,7 @@ def convert_np_dtype_to_dtype_(np_dtype): elif dtype == np.uint8: return core.VarDesc.VarType.UINT8 else: - raise ValueError("Not supported numpy dtype " + six.binary_type(dtype)) + raise ValueError("Not supported numpy dtype %s" % dtype) def dtype_is_floating(dtype): @@ -198,11 +201,11 @@ class Variable(object): if name is None: name = unique_name.generate('_generated_var') is_new_var = False - name = name if isinstance(name, six.binary_type) else name.encode() - self.desc = self.block.desc.find_var(name) + name = cpt.to_text(name) + self.desc = self.block.desc.find_var(cpt.to_bytes(name)) if self.desc is None: - self.desc = self.block.desc.var(name) + self.desc = self.block.desc.var(cpt.to_bytes(name)) is_new_var = True if is_new_var: @@ -325,7 +328,7 @@ class Variable(object): @property def name(self): - return self.desc.name() + return cpt.to_text(self.desc.name()) @name.setter def name(self, new_name): @@ -531,14 +534,7 @@ class Operator(object): elif isinstance(arg, six.binary_type): in_arg_names.append(arg.decode()) else: - if isinstance(arg.name, six.string_types): - in_arg_names.append(arg.name) - elif isinstance(arg.name, six.binary_type): - in_arg_names.append(arg.name.decode()) - else: - raise TypeError( - "arguments require unicode, str or bytes, but get %s instead." - % (type(arg.name))) + in_arg_names.append(cpt.to_text(arg.name)) self.desc.set_input(in_proto.name, in_arg_names) else: self.desc.set_input(in_proto.name, []) @@ -567,14 +563,7 @@ class Operator(object): (out_proto.name, len(out_args))) out_arg_names = [] for arg in out_args: - if isinstance(arg.name, six.string_types): - out_arg_names.append(arg.name) - elif isinstance(arg.name, six.binary_type): - out_arg_names.append(arg.name.decode()) - else: - raise TypeError( - "arguments require unicode, str or bytes, but get %s instead." - % (type(arg.name))) + out_arg_names.append(cpt.to_text(arg.name)) arg.op = self self.desc.set_output(out_proto.name, out_arg_names) @@ -970,10 +959,9 @@ class Block(object): Variable: the Variable with the giving name. """ if not isinstance(name, six.string_types): - if not isinstance(name, six.binary_type): - raise TypeError( - "var require string as parameter, but get %s instead." % - (type(name))) + raise TypeError( + "var require string as parameter, but get %s instead." % + (type(name))) v = self.vars.get(name, None) if v is None: raise ValueError("var %s not in this block" % name) @@ -1024,7 +1012,7 @@ class Block(object): return list(self.iter_parameters()) def iter_parameters(self): - return (item[1] for item in list(self.vars.items()) + return (item[1] for item in six.iteritems(self.vars) if isinstance(item[1], Parameter)) def create_var(self, *args, **kwargs): @@ -1052,6 +1040,9 @@ class Block(object): Returns: Variable: the Variable with the giving name. """ + name = cpt.to_text(name) + new_name = cpt.to_text(new_name) + if not self.has_var(name): raise ValueError("var %s is not in current block" % name) v = self.var(name) @@ -1070,9 +1061,9 @@ class Block(object): else: raise ValueError("unsupported var type: %s", type(v)) orig_var_type = v.type - self.desc._rename_var(name, new_name) + self.desc._rename_var(cpt.to_bytes(name), cpt.to_bytes(new_name)) # NOTE: v is destroyed by C++ after calling _rename_var. - d = self.desc.find_var(new_name) + d = self.desc.find_var(cpt.to_bytes(new_name)) if var_type == "Parameter": var = Parameter( self, @@ -1103,7 +1094,7 @@ class Block(object): def _remove_var(self, name): self._sync_with_cpp() - self.desc._remove_var(name) + self.desc._remove_var(cpt.to_bytes(name)) del self.vars[name] def create_parameter(self, *args, **kwargs): @@ -1205,7 +1196,7 @@ class Block(object): # sync variables removed from c++ end for var in list(self.vars.keys()): - if not self.desc.find_var(var): + if not self.desc.find_var(cpt.to_bytes(var)): self.vars.pop(var) # sync operators from cpp @@ -1372,6 +1363,13 @@ class Program(object): self._current_role = core.op_proto_and_checker_maker.OpRole.Forward self._op_role_var = [] + # for distribute + self._is_distributed = False + self._is_chief = False + self._slice_vars_and_attrs = [] + self._endpoints = [] + self._distributed_lookup_table = None + @property def op_role(self): """ @@ -1576,7 +1574,9 @@ class Program(object): p.current_block_idx = self.current_block_idx p._seed = self._seed p.desc = core.ProgramDesc(self.desc) - p.blocks = [Block(p, i) for i in xrange(self.desc.num_blocks())] + p.blocks = [ + Block(p, i) for i in six.moves.range(self.desc.num_blocks()) + ] p._current_role = self._current_role p._op_role_var = self._op_role_var @@ -1632,7 +1632,9 @@ class Program(object): targets_idx.append([t.block.idx, t.idx]) res = Program() res.desc = core.prune(self.desc, targets_idx) - res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())] + res.blocks = [ + Block(res, i) for i in six.moves.range(res.desc.num_blocks()) + ] res._sync_with_cpp() return res @@ -1675,16 +1677,18 @@ class Program(object): root_block._remove_op(0, read_op_idx + 1) for var in root_block.all_vars(): if var.type() == core.VarDesc.VarType.READER: - root_block._remove_var(var.name()) + root_block._remove_var(cpt.to_bytes(var.name())) # change all `is_test` attributes to True - for i in range(res.desc.num_blocks()): + for i in six.moves.range(res.desc.num_blocks()): block = res.desc.block(i) - for j in range(block.op_size()): + for j in six.moves.range(block.op_size()): op = block.op(j) if op.has_attr('is_test'): op.set_attr('is_test', True) - res.blocks = [Block(res, i) for i in range(res.desc.num_blocks())] + res.blocks = [ + Block(res, i) for i in six.moves.range(res.desc.num_blocks()) + ] res._sync_with_cpp() return res @@ -1704,7 +1708,7 @@ class Program(object): """ p = Program() p.desc = core.ProgramDesc(binary_str) - p.blocks = [Block(p, i) for i in range(p.desc.num_blocks())] + p.blocks = [Block(p, i) for i in six.moves.range(p.desc.num_blocks())] p._sync_with_cpp() return p diff --git a/python/paddle/fluid/graphviz.py b/python/paddle/fluid/graphviz.py index ba67bf5ae6..2b18d854d1 100644 --- a/python/paddle/fluid/graphviz.py +++ b/python/paddle/fluid/graphviz.py @@ -12,9 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import os import random import six +import functools import subprocess import logging @@ -105,8 +108,9 @@ class Graph(object): def _rank_repr(self): ranks = sorted( - list(self.rank_groups.items()), - cmp=lambda a, b: a[1].priority > b[1].priority) + six.iteritems(self.rank_groups), + key=functools.cmp_to_key( + lambda a, b: a[1].priority > b[1].priority)) repr = [] for x in ranks: repr.append(str(x[1])) @@ -149,7 +153,7 @@ class Node(object): name=self.name, label=self.label, extra=',' + ','.join("%s=%s" % (key, crepr(value)) - for key, value in list(self.attrs.items())) + for key, value in six.iteritems(self.attrs)) if self.attrs else "") return reprs @@ -173,7 +177,7 @@ class Edge(object): target=self.target.name, extra="" if not self.attrs else "[" + ','.join("{}={}".format(attr[0], crepr(attr[1])) - for attr in list(self.attrs.items())) + "]") + for attr in six.iteritems(self.attrs)) + "]") return repr diff --git a/python/paddle/fluid/inferencer.py b/python/paddle/fluid/inferencer.py index ff382d8b83..3d2ef56617 100644 --- a/python/paddle/fluid/inferencer.py +++ b/python/paddle/fluid/inferencer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import contextlib from . import core diff --git a/python/paddle/fluid/initializer.py b/python/paddle/fluid/initializer.py index 6dedbae7a6..bd46ed8e50 100644 --- a/python/paddle/fluid/initializer.py +++ b/python/paddle/fluid/initializer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import framework import numpy as np import contextlib diff --git a/python/paddle/fluid/io.py b/python/paddle/fluid/io.py index af73421032..b3ed094c89 100644 --- a/python/paddle/fluid/io.py +++ b/python/paddle/fluid/io.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import os import errno import time @@ -370,6 +372,7 @@ def load_vars(executor, load_vars( executor, dirname=dirname, + main_program=main_program, vars=list(filter(predicate, main_program.list_vars())), filename=filename) else: @@ -401,9 +404,12 @@ def load_vars(executor, inputs={}, outputs={"Out": load_var_list}, attrs={'file_path': os.path.join(dirname, filename)}) - executor.run(load_prog) + # load slice vars on pserver, if have it. + _load_slice_up_vars(executor, dirname, + main_program._slice_vars_and_attrs) + def load_params(executor, dirname, main_program=None, filename=None): """ @@ -603,25 +609,15 @@ def save_inference_model(dirname, # "./infer_model". """ - if isinstance(feeded_var_names, six.binary_type): + if isinstance(feeded_var_names, six.string_types): feeded_var_names = [feeded_var_names] - elif isinstance(feeded_var_names, six.text_type): - feeded_var_names = [feeded_var_names.encode()] else: if len(feeded_var_names) > 0: # TODO(paddle-dev): polish these code blocks if not (bool(feeded_var_names) and all( - isinstance(name, six.binary_type) + isinstance(name, six.string_types) for name in feeded_var_names)): - if not (all( - isinstance(name, six.text_type) - for name in feeded_var_names)): - raise ValueError( - "'feed_var_names' should be a list of str.") - else: - feeded_var_names = [ - name.encode() for name in feeded_var_names - ] + raise ValueError("'feed_var_names' should be a list of str.") if isinstance(target_vars, Variable): target_vars = [target_vars] @@ -667,11 +663,19 @@ def save_inference_model(dirname, save_persistables(executor, dirname, inference_program, params_filename) + # if there is lookup table, the trainer 0 will notify all pserver to save. + if main_program._is_distributed and main_program._is_chief and main_program._distributed_lookup_table: + lookup_table_filename = os.path.join(dirname, "__lookup_table__") + _save_lookup_tables_by_notify(executor, lookup_table_filename, + main_program._distributed_lookup_table, + main_program._endpoints) + def load_inference_model(dirname, executor, model_filename=None, - params_filename=None): + params_filename=None, + pserver_endpoints=None): """ Load inference model from a directory @@ -687,6 +691,10 @@ def load_inference_model(dirname, parameters were saved in a single binary file. If parameters were saved in separate files, set it as 'None'. + pserver_endpoints(list|None): This only need by distributed inference. + When use distributed look up table in training, + We also need it in inference.The parameter is + a list of pserver endpoints. Returns: tuple: The return of this function is a tuple with three elements: @@ -705,12 +713,16 @@ def load_inference_model(dirname, exe = fluid.Executor(fluid.CPUPlace()) path = "./infer_model" + endpoints = ["127.0.0.1:2023","127.0.0.1:2024"] [inference_program, feed_target_names, fetch_targets] = fluid.io.load_inference_model(dirname=path, executor=exe) results = exe.run(inference_program, feed={feed_target_names[0]: tensor_img}, fetch_list=fetch_targets) + # if we need lookup table, we will use: + fluid.io.load_inference_model(dirname=path, executor=exe, pserver_endpoints=endpoints) + # In this exsample, the inference program was saved in the # "./infer_model/__model__" and parameters were saved in # separate files in ""./infer_model". @@ -737,6 +749,9 @@ def load_inference_model(dirname, program = Program.parse_from_string(program_desc_str) load_persistables(executor, dirname, program, params_filename) + if pserver_endpoints: + program = _endpoints_replacement(program, pserver_endpoints) + feed_target_names = program.desc.get_feed_target_names() fetch_target_names = program.desc.get_fetch_target_names() fetch_targets = [ @@ -746,6 +761,61 @@ def load_inference_model(dirname, return [program, feed_target_names, fetch_targets] +def _save_lookup_tables_by_notify(executor, dirname, lookup_table, + pserver_endpoints): + """ + This function will send checkpoint notify message from Trainer 0 + to all the pservers. + The checkpoint notify message contains lookup table name, + the absolute path on pserver to save lookup_table. + + Args: + executor(Executor): The executor to run for send checkpoint notify. + dirname(str): The folder where to save. + lookup_table(string): the lookup table name, when use distribute + lookup table, we can get lookup table name by DistributeTranspiler. + table_name + ps_endpoint_list(list): the parameter server ip:port list. + when use distribute lookup table, we can get ps_endpoint_list by + distribute arguments. + Return: + None + + Examples: + .. code-block:: python + + exe = fluid.Executor(fluid.CPUPlace()) + param_path = "./my_paddle_model" + table_name = "share_w" + ps_endpoints = ["127.0.0.1:6000","127.0.0.1:6001"] + + _save_pserver_vars_by_notify(executor=exe, + dirname=param_path, lookup_table=table_name, + pserver_endpoints=ps_endpoints) + """ + + pserver_notify_program = Program() + pserver_notify_block = pserver_notify_program.global_block() + + attrs = {} + attrs['epmap'] = pserver_endpoints + attrs['dir'] = dirname + attrs['lookup_table'] = lookup_table + + pserver_notify_block.append_op( + type='checkpoint_notify', inputs={}, outputs={}, attrs=attrs) + executor.run(pserver_notify_program) + + +def _endpoints_replacement(program, endpoints): + ENDPOINT_MAP = "epmap" + for op in program.global_block().ops: + if op.has_attr(ENDPOINT_MAP): + op.set_attr(ENDPOINT_MAP, endpoints) + program._sync_with_cpp() + return program + + def get_parameter_value(para, executor): """ Get the LoDTensor value of the given parameter. @@ -807,3 +877,46 @@ def get_parameter_value_by_name(name, executor, program=None): program = default_main_program() var = program.global_block().var(name) return get_parameter_value(var, executor) + + +def _load_slice_up_vars(executor, dirname, slice_vars_and_attrs): + if not slice_vars_and_attrs: + return + + load_prog = Program() + load_block = load_prog.global_block() + + for var_tuple in slice_vars_and_attrs: + orig_var = var_tuple[0] + start = var_tuple[1] + slice_var = var_tuple[2] + end = start + reduce(lambda x, y: x * y, slice_var.shape) + + clone_orig_var = load_block.create_var( + name=orig_var.name, + type=orig_var.type, + shape=orig_var.shape, + dtype=orig_var.dtype, + persistable=True) + + clone_slice_var = load_block.create_var( + name=slice_var.name, + type=slice_var.type, + shape=slice_var.shape, + dtype=slice_var.dtype, + persistable=True) + + load_block.append_op( + type='load', + inputs={}, + outputs={'Out': [clone_orig_var]}, + attrs={'file_path': os.path.join(dirname, clone_orig_var.name)}) + load_block.append_op( + type="slice", + inputs={'Input': clone_orig_var}, + outputs={'Out': clone_slice_var}, + attrs={'axes': [0], + 'starts': [start], + 'ends': [end]}) + + executor.run(load_prog) diff --git a/python/paddle/fluid/layer_helper.py b/python/paddle/fluid/layer_helper.py index 0c2b1eb795..bd9727b6ac 100644 --- a/python/paddle/fluid/layer_helper.py +++ b/python/paddle/fluid/layer_helper.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import copy import itertools import six @@ -85,7 +87,7 @@ class LayerHelper(object): raise ValueError("parameter number mismatch") elif len(param_attr) == 1 and length != 1: tmp = [None] * length - for i in range(length): + for i in six.moves.range(length): tmp[i] = copy.deepcopy(param_attr[0]) param_attr = tmp return param_attr diff --git a/python/paddle/fluid/layers/__init__.py b/python/paddle/fluid/layers/__init__.py index a48e360463..a2a808777d 100644 --- a/python/paddle/fluid/layers/__init__.py +++ b/python/paddle/fluid/layers/__init__.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import ops from .ops import * from . import nn diff --git a/python/paddle/fluid/layers/control_flow.py b/python/paddle/fluid/layers/control_flow.py index 9fb7b4d0ca..8bfe11916b 100644 --- a/python/paddle/fluid/layers/control_flow.py +++ b/python/paddle/fluid/layers/control_flow.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import contextlib from .layer_function_generator import autodoc, templatedoc @@ -22,6 +24,7 @@ from ..initializer import force_init_on_cpu from .ops import logical_and, logical_not, logical_or import numpy import warnings +import six from functools import reduce __all__ = [ @@ -602,7 +605,7 @@ class StaticRNN(object): boot_memories = [] pre_memories = [] memories = [] - for _, mem in list(self.memories.items()): + for _, mem in six.iteritems(self.memories): boot_memories.append(mem.init) pre_memories.append(mem.pre_mem.name) mem_var = rnn_block.var(mem.mem.name) @@ -1269,8 +1272,8 @@ class ConditionalBlock(object): parent_block.append_op( type='conditional_block', inputs={ - 'X': self.inputs, - 'Params': param_list, + 'Cond': self.inputs, + 'Input': param_list, }, outputs={'Out': out_list, 'Scope': [step_scope]}, diff --git a/python/paddle/fluid/layers/detection.py b/python/paddle/fluid/layers/detection.py index b996c83688..7207147884 100644 --- a/python/paddle/fluid/layers/detection.py +++ b/python/paddle/fluid/layers/detection.py @@ -15,13 +15,17 @@ All layers just related to the detection neural network. """ +from __future__ import print_function + from .layer_function_generator import generate_layer_fn from .layer_function_generator import autodoc, templatedoc from ..layer_helper import LayerHelper from . import tensor from . import nn from . import ops +from ... import compat as cpt import math +import six import numpy from functools import reduce @@ -104,7 +108,7 @@ def rpn_target_assign(loc, examples. Returns: - tuple: + tuple: A tuple(predicted_scores, predicted_location, target_label, target_bbox) is returned. The predicted_scores and predicted_location is the predicted result of the RPN. @@ -115,7 +119,7 @@ def rpn_target_assign(loc, anchors. The predicted_scores is a 2D Tensor with shape [F + B, 1], and the shape of target_label is same as the shape of the predicted_scores, B is the number of the background - anchors, the F and B is depends on the input of this operator. + anchors, the F and B is depends on the input of this operator. Examples: .. code-block:: python @@ -232,8 +236,8 @@ def detection_output(loc, nms_eta(float): The parameter for adaptive NMS. Returns: - Variable: - + Variable: + The detection outputs is a LoDTensor with shape [No, 6]. Each row has six values: [label, confidence, xmin, ymin, xmax, ymax]. `No` is the total number of detections in this mini-batch. For each @@ -504,7 +508,7 @@ def target_assign(input, Assumed that the row offset for each instance in `neg_indices` is called neg_lod, for i-th instance and each `id` of neg_indices in this instance: - + .. code-block:: text out[i][id][0 : K] = {mismatch_value, mismatch_value, ...} @@ -522,11 +526,11 @@ def target_assign(input, mismatch_value (float32): Fill this value to the mismatched location. Returns: - tuple: - A tuple(out, out_weight) is returned. out is a 3D Tensor with - shape [N, P, K], N and P is the same as they are in - `neg_indices`, K is the same as it in input of X. If - `match_indices[i][j]`. out_weight is the weight for output with + tuple: + A tuple(out, out_weight) is returned. out is a 3D Tensor with + shape [N, P, K], N and P is the same as they are in + `neg_indices`, K is the same as it in input of X. If + `match_indices[i][j]`. out_weight is the weight for output with the shape of [N, P, 1]. Examples: @@ -834,7 +838,7 @@ def prior_box(input, offset(float): Prior boxes center offset. Default: 0.5 name(str): Name of the prior box op. Default: None. min_max_aspect_ratios_order(bool): If set True, the output prior box is - in order of [min, max, aspect_ratios], which is consistent with + in order of [min, max, aspect_ratios], which is consistent with Caffe. Please note, this order affects the weights order of convolution layer followed by and does not affect the final detection results. Default: False. @@ -977,7 +981,7 @@ def multi_box_head(inputs, stride(int|list|tuple): The stride of conv2d. Default:1, name(str): Name of the prior box layer. Default: None. min_max_aspect_ratios_order(bool): If set True, the output prior box is - in order of [min, max, aspect_ratios], which is consistent with + in order of [min, max, aspect_ratios], which is consistent with Caffe. Please note, this order affects the weights order of convolution layer followed by and does not affect the fininal detection results. Default: False. @@ -1039,7 +1043,7 @@ def multi_box_head(inputs, min_sizes = [] max_sizes = [] step = int(math.floor(((max_ratio - min_ratio)) / (num_layer - 2))) - for ratio in range(min_ratio, max_ratio + 1, step): + for ratio in six.moves.range(min_ratio, max_ratio + 1, step): min_sizes.append(base_size * ratio / 100.) max_sizes.append(base_size * (ratio + step) / 100.) min_sizes = [base_size * .10] + min_sizes @@ -1108,8 +1112,8 @@ def multi_box_head(inputs, mbox_loc = nn.transpose(mbox_loc, perm=[0, 2, 3, 1]) compile_shape = [ - mbox_loc.shape[0], - mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3] / 4, 4 + mbox_loc.shape[0], cpt.floor_division( + mbox_loc.shape[1] * mbox_loc.shape[2] * mbox_loc.shape[3], 4), 4 ] run_shape = tensor.assign(numpy.array([0, -1, 4]).astype("int32")) mbox_loc_flatten = nn.reshape( @@ -1127,8 +1131,9 @@ def multi_box_head(inputs, conf_loc = nn.transpose(conf_loc, perm=[0, 2, 3, 1]) new_shape = [0, -1, num_classes] compile_shape = [ - conf_loc.shape[0], conf_loc.shape[1] * conf_loc.shape[2] * - conf_loc.shape[3] / num_classes, num_classes + conf_loc.shape[0], + cpt.floor_division(conf_loc.shape[1] * conf_loc.shape[2] * + conf_loc.shape[3], num_classes), num_classes ] run_shape = tensor.assign( numpy.array([0, -1, num_classes]).astype("int32")) diff --git a/python/paddle/fluid/layers/device.py b/python/paddle/fluid/layers/device.py index bb1fb7fd57..43ebd160de 100644 --- a/python/paddle/fluid/layers/device.py +++ b/python/paddle/fluid/layers/device.py @@ -15,6 +15,8 @@ All util layers. """ +from __future__ import print_function + from .layer_function_generator import autodoc from ..framework import unique_name from ..layer_helper import LayerHelper diff --git a/python/paddle/fluid/layers/io.py b/python/paddle/fluid/layers/io.py index 327ae30981..b03ee514f5 100644 --- a/python/paddle/fluid/layers/io.py +++ b/python/paddle/fluid/layers/io.py @@ -11,8 +11,11 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import contextlib import multiprocessing +import six import threading from ..data_feeder import DataFeeder @@ -21,7 +24,7 @@ from .layer_function_generator import templatedoc from .. import core from ..executor import global_scope from ..framework import convert_np_dtype_to_dtype_, default_main_program, \ - default_startup_program, program_guard, Program + default_startup_program, program_guard, Program, Variable from ..layer_helper import LayerHelper from ..unique_name import generate as unique_name @@ -69,7 +72,7 @@ def data(name, """ helper = LayerHelper('data', **locals()) shape = list(shape) - for i in range(len(shape)): + for i in six.moves.range(len(shape)): if shape[i] is None: shape[i] = -1 append_batch_size = False @@ -206,7 +209,7 @@ class ListenAndServ(object): }) -def Send(endpoints, send_vars, sync=True): +def Send(endpoints, send_vars, dummy_output=None, sync=True): """ Send variables to the server side, and get vars from server side when server have finished running server side program. @@ -220,6 +223,13 @@ def Send(endpoints, send_vars, sync=True): """ assert (type(send_vars) == list) + if dummy_output is None: + dummy_output = [] + elif isinstance(dummy_output, Variable): + dummy_output = [dummy_output] + + assert (type(dummy_output) == list) + epmap = endpoints.split(",") endpoints = list(set(epmap)) @@ -229,6 +239,7 @@ def Send(endpoints, send_vars, sync=True): helper.append_op( type="send", inputs={"X": send_vars}, + outputs={"Out": dummy_output}, attrs={ "endpoints": endpoints, "epmap": epmap, @@ -238,7 +249,7 @@ def Send(endpoints, send_vars, sync=True): helper.append_op(type="send_barrier", attrs={"endpoints": endpoints}) -def Recv(endpoints, get_vars, sync=True): +def Recv(endpoints, get_vars, dummy_input=None, sync=True): """ Receive variables from server side @@ -253,13 +264,20 @@ def Recv(endpoints, get_vars, sync=True): """ assert (type(get_vars) == list) + if dummy_input is None: + dummy_input = [] + elif isinstance(dummy_input, Variable): + dummy_input = [dummy_input] + + assert (type(dummy_input) == list) + epmap = endpoints.split(",") endpoints = list(set(epmap)) helper = LayerHelper("Recv", **locals()) helper.append_op( type="recv", - inputs={"X": get_vars}, + inputs={"X": dummy_input}, outputs={"Out": get_vars}, attrs={"endpoints": endpoints, "epmap": epmap}) @@ -674,7 +692,7 @@ def py_reader(capacity, def __tensor_provider__(): for slots in paddle_reader(): - yield [slots[str(idx)] for idx in xrange(counter)] + yield [slots[str(idx)] for idx in six.moves.xrange(counter)] __set_tensor_provider__(__tensor_provider__) @@ -750,7 +768,7 @@ def open_files(filenames, else: buffer_size = int(buffer_size) - if isinstance(filenames, basestring): + if isinstance(filenames, six.string_types): filenames = [filenames] dtypes = [convert_np_dtype_to_dtype_(dt) for dt in dtypes] shape_concat = [] @@ -1005,7 +1023,7 @@ class Preprocessor(object): source_lod_levels = self.underlying_reader.desc.lod_levels() self.source_var_names = [ unique_name("preprocessor_source") - for _ in range(len(source_shapes)) + for _ in six.moves.range(len(source_shapes)) ] source_vars = [] for var_name, shape, dtype, lod_level in zip( diff --git a/python/paddle/fluid/layers/layer_function_generator.py b/python/paddle/fluid/layers/layer_function_generator.py index c0d72620b1..8963d74de0 100644 --- a/python/paddle/fluid/layers/layer_function_generator.py +++ b/python/paddle/fluid/layers/layer_function_generator.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import re import functools import warnings diff --git a/python/paddle/fluid/layers/learning_rate_scheduler.py b/python/paddle/fluid/layers/learning_rate_scheduler.py index daf91a40f7..be368007dd 100644 --- a/python/paddle/fluid/layers/learning_rate_scheduler.py +++ b/python/paddle/fluid/layers/learning_rate_scheduler.py @@ -20,6 +20,8 @@ User can also implement their own learning_rate_decay strategy according to this module. """ +from __future__ import print_function + from . import control_flow from . import nn from . import ops @@ -72,10 +74,10 @@ def noam_decay(d_model, warmup_steps): def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False): """ - Applies exponential decay to the learning rate. + Applies exponential decay to the learning rate. - When training a model, it is often recommended to lower the learning rate as the - training progresses. By using this function, the learning rate will be decayed by + When training a model, it is often recommended to lower the learning rate as the + training progresses. By using this function, the learning rate will be decayed by 'decay_rate' every 'decay_steps' steps. >>> if staircase == True: @@ -148,8 +150,8 @@ def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False): """ Applies inverse time decay to the initial learning rate. - When training a model, it is often recommended to lower the learning rate as the - training progresses. By using this function, an inverse decay function will be + When training a model, it is often recommended to lower the learning rate as the + training progresses. By using this function, an inverse decay function will be applied to the initial learning rate. >>> if staircase == True: diff --git a/python/paddle/fluid/layers/math_op_patch.py b/python/paddle/fluid/layers/math_op_patch.py index 0e10a91d25..a458cebfb1 100644 --- a/python/paddle/fluid/layers/math_op_patch.py +++ b/python/paddle/fluid/layers/math_op_patch.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from ..framework import Variable, unique_name from .layer_function_generator import OpProtoHolder from ..initializer import force_init_on_cpu diff --git a/python/paddle/fluid/layers/metric_op.py b/python/paddle/fluid/layers/metric_op.py index 49bae1e8af..2c3bdd77e1 100644 --- a/python/paddle/fluid/layers/metric_op.py +++ b/python/paddle/fluid/layers/metric_op.py @@ -15,6 +15,8 @@ All layers just related to metric. """ +from __future__ import print_function + import warnings from ..layer_helper import LayerHelper from ..initializer import Normal, Constant @@ -81,9 +83,9 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1): **Area Under the Curve (AUC) Layer** This implementation computes the AUC according to forward output and label. - It is used very widely in binary classification evaluation. + It is used very widely in binary classification evaluation. - Note: If input label contains values other than 0 and 1, it will be cast + Note: If input label contains values other than 0 and 1, it will be cast to `bool`. Find the relevant definitions `here `_. @@ -93,14 +95,14 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1): 2. PR: Precision Recall Args: - input(Variable): A floating-point 2D Variable, values are in the range - [0, 1]. Each row is sorted in descending order. This - input should be the output of topk. Typically, this + input(Variable): A floating-point 2D Variable, values are in the range + [0, 1]. Each row is sorted in descending order. This + input should be the output of topk. Typically, this Variable indicates the probability of each label. - label(Variable): A 2D int Variable indicating the label of the training + label(Variable): A 2D int Variable indicating the label of the training data. The height is batch size and width is always 1. curve(str): Curve type, can be 'ROC' or 'PR'. Default 'ROC'. - num_thresholds(int): The number of thresholds to use when discretizing + num_thresholds(int): The number of thresholds to use when discretizing the roc curve. Default 200. topk(int): only topk number of prediction output will be used for auc. @@ -109,7 +111,7 @@ def auc(input, label, curve='ROC', num_thresholds=200, topk=1): Examples: .. code-block:: python - + # network is a binary classification model and label the ground truth prediction = network(image, is_infer=True) auc_out=fluid.layers.auc(input=prediction, label=label) diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index be852b6711..71592618f5 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -11,24 +11,12 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. - -# Copyright (c ) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. """ All layers just related to the neural network. """ +from __future__ import print_function + from ..layer_helper import LayerHelper from ..initializer import Normal, Constant from ..framework import Variable @@ -106,6 +94,7 @@ __all__ = [ 'image_resize_short', 'resize_bilinear', 'gather', + 'scatter', 'random_crop', 'mean_iou', 'relu', @@ -362,7 +351,7 @@ def dynamic_lstm(input, """ helper = LayerHelper('lstm', **locals()) - size = size / 4 + size = size // 4 weight = helper.create_parameter( attr=helper.param_attr, shape=[size, 4 * size], dtype=dtype) bias_size = [1, 7 * size] @@ -552,7 +541,7 @@ def dynamic_lstmp(input, """ helper = LayerHelper('lstmp', **locals()) - size = size / 4 + size = size // 4 weight = helper.create_parameter( attr=helper.param_attr, shape=[proj_size, 4 * size], dtype=dtype) proj_weight = helper.create_parameter( @@ -780,7 +769,7 @@ def gru_unit(input, helper = LayerHelper('gru_unit', **locals()) dtype = helper.input_dtype() - size = size / 3 + size = size // 3 # create weight weight = helper.create_parameter( @@ -1264,7 +1253,7 @@ def sequence_conv(input, outputs={"Out": pre_bias}, attrs={ 'contextStride': filter_stride, - 'contextStart': -int(filter_size / 2), + 'contextStart': -int(filter_size // 2), 'contextLength': filter_size }) pre_act = helper.append_bias_op(pre_bias) @@ -1320,15 +1309,15 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): def softmax(input, param_attr=None, bias_attr=None, use_cudnn=True, name=None): """ - The input of the softmax operator is a tensor of any rank. The output tensor + The input of the softmax operator is a tensor of any rank. The output tensor has the same shape as the input. - The input tensor will first be logically flattened to a 2-D matrix. The matrix's - second dimension(row length) is as same as the last dimension of the input - tensor, and the first dimension(column length) is the product of all other - dimensions of the input tensor. For each row of the matrix, the softmax operator - squashes the K-dimensional(K is the width of the matrix, which is also the size - of the input tensor's last dimension) vector of arbitrary real values to a + The input tensor will first be logically flattened to a 2-D matrix. The matrix's + second dimension(row length) is as same as the last dimension of the input + tensor, and the first dimension(column length) is the product of all other + dimensions of the input tensor. For each row of the matrix, the softmax operator + squashes the K-dimensional(K is the width of the matrix, which is also the size + of the input tensor's last dimension) vector of arbitrary real values to a K-dimensional vector of real values in the range [0, 1] that add up to 1. It computes the exponential of the given dimension and the sum of exponential @@ -1496,7 +1485,7 @@ def conv2d(input, else: if num_channels % groups != 0: raise ValueError("num_channels must be divisible by groups.") - num_filter_channels = num_channels / groups + num_filter_channels = num_channels // groups filter_size = utils.convert_to_list(filter_size, 2, 'filter_size') stride = utils.convert_to_list(stride, 2, 'stride') @@ -1507,7 +1496,7 @@ def conv2d(input, raise ValueError("use_cudnn should be True or False") input_shape = input.shape - filter_shape = [num_filters, num_filter_channels] + filter_size + filter_shape = [num_filters, int(num_filter_channels)] + filter_size def _get_default_param_initializer(): std = (2.0 / (filter_size[0]**2 * num_channels))**0.5 @@ -1658,7 +1647,7 @@ def conv3d(input, else: if num_channels % groups != 0: raise ValueError("num_channels must be divisible by groups.") - num_filter_channels = num_channels / groups + num_filter_channels = num_channels // groups filter_size = utils.convert_to_list(filter_size, 3, 'filter_size') stride = utils.convert_to_list(stride, 3, 'stride') @@ -2393,16 +2382,16 @@ def conv2d_transpose(input, w_in = input.shape[3] filter_size_h = (output_size[0] - (h_in - 1) * stride[0] + 2 * - padding[0] - 1) / dilation[0] + 1 + padding[0] - 1) // dilation[0] + 1 filter_size_w = (output_size[1] - (w_in - 1) * stride[1] + 2 * - padding[1] - 1) / dilation[1] + 1 + padding[1] - 1) // dilation[1] + 1 filter_size = [filter_size_h, filter_size_w] else: filter_size = utils.convert_to_list(filter_size, 2, 'conv2d_transpose.filter_size') groups = 1 if groups is None else groups - filter_shape = [input_channel, num_filters / groups] + filter_size + filter_shape = [input_channel, num_filters // groups] + filter_size img_filter = helper.create_parameter( dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) @@ -2560,18 +2549,18 @@ def conv3d_transpose(input, w_in = input.shape[4] filter_size_d = (output_size[0] - (d_in - 1) * stride[0] + 2 * - padding[0] - 1) / dilation[0] + 1 + padding[0] - 1) // dilation[0] + 1 filter_size_h = (output_size[1] - (h_in - 1) * stride[1] + 2 * - padding[1] - 1) / dilation[1] + 1 + padding[1] - 1) // dilation[1] + 1 filter_size_w = (output_size[2] - (w_in - 1) * stride[2] + 2 * - padding[2] - 1) / dilation[2] + 1 + padding[2] - 1) // dilation[2] + 1 filter_size = [filter_size_d, filter_size_h, filter_size_w] else: filter_size = utils.convert_to_list(filter_size, 3, 'conv3d_transpose.filter_size') groups = 1 if groups is None else groups - filter_shape = [input_channel, num_filters / groups] + filter_size + filter_shape = [input_channel, num_filters // groups] + filter_size img_filter = helper.create_parameter( dtype=input.dtype, shape=filter_shape, attr=helper.param_attr) @@ -2678,15 +2667,15 @@ def beam_search(pre_ids, Refer to `Beam search `_ for more details. - - This layer does the search in beams for one time step. Specifically, it + + This layer does the search in beams for one time step. Specifically, it selects the top-K candidate word ids of current step from :attr:`ids` according to their :attr:`scores` for all source sentences, where K is :attr:`beam_size` and :attr:`ids, scores` are predicted results from the computation cell. Additionally, :attr:`pre_ids` and :attr:`pre_scores` are the output of beam_search at previous step, they are needed for special use to handle ended candidate translations. - + Note that the :attr:`scores` passed in should be accumulated scores, and length penalty should be done with extra operators before calculating the accumulated scores if needed, also suggest finding top-K before it and @@ -3887,7 +3876,7 @@ def nce(input, def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): """ The hierarchical sigmoid operator is used to accelerate the training - process of language model. This operator organizes the classes into a + process of language model. This operator organizes the classes into a complete binary tree, each leaf node represents a class(a word) and each internal node acts as a binary classifier. For each word there's a unique path from root to it's leaf node, hsigmoid calculate the cost for each @@ -3897,9 +3886,9 @@ def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): Refer to `Hierarchical Probabilistic Neural Network Language Model `_ - + Args: - input (Variable): The input tensor variable with shape + input (Variable): The input tensor variable with shape :math:`[N \\times D]`, where :math:`N` is the size of mini-batch, and :math:`D` is the feature size. label (Variable): The tensor variable contains labels of training data. @@ -3907,7 +3896,7 @@ def hsigmoid(input, label, num_classes, param_attr=None, bias_attr=None): num_classes: (int), The number of classes, must not be less than 2. param_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for learnable parameters/weights of this layer. - bias_attr (ParamAttr|list of ParamAttr, default None): The parameter + bias_attr (ParamAttr|list of ParamAttr, default None): The parameter attribute for the bias of this layer. If it is set to False, no bias will be applied. @@ -5048,6 +5037,47 @@ def gather(input, index): return out +def scatter(input, index, updates, name=None): + """ + **Scatter Layer** + + Output is obtained by updating the input on selected indices on the first + axis. + + .. math:: + + Out = X + Out[Ids] = Updates + + Args: + input (Variable): The source input with rank>=1. + index (Variable): The index input with rank=1. Its dtype should be + int32 or int64 as it is used as indexes. + updates (Variable): The updated value of scatter op. + name (str|None): The output variable name. Default None. + + Returns: + output (Variable): The output is a tensor with the same shape as input. + + Examples: + + .. code-block:: python + + output = fluid.layers.scatter(input, index, updates) + + """ + helper = LayerHelper('scatter', **locals()) + dtype = helper.input_dtype() + out = helper.create_tmp_variable(dtype) + helper.append_op( + type="scatter", + inputs={"X": input, + "Ids": index, + "Updates": updates}, + outputs={"Out": out}) + return out + + @templatedoc() def random_crop(x, shape, seed=None): """ @@ -5306,23 +5336,23 @@ def rank_loss(label, left, right, name=None): is a pairwise ranking model with a training sample consisting of a pair of documents, A and B. Label P indicates whether A is ranked higher than B or not: - + P = {0, 1} or {0, 0.5, 1}, where 0.5 means that there is no information about the rank of the input pair. - + Rank loss layer takes three inputs: left (o_i), right (o_j) and label (P_{i,j}). The inputs respectively represent RankNet's output scores for documents A and B and the value of label P. The following equation computes rank loss C_{i,j} from the inputs: - + $$ C_{i,j} = -\tilde{P_{ij}} * o_{i,j} + \log(1 + e^{o_{i,j}}) \\ o_{i,j} = o_i - o_j \\ \tilde{P_{i,j}} = \left \{0, 0.5, 1 \right \} \ or \ \left \{0, 1 \right \} $$ - - Rank loss layer takes batch inputs with size batch_size (batch_size >= 1). - + + Rank loss layer takes batch inputs with size batch_size (batch_size >= 1). + Args: label (Variable): Indicats whether A ranked higher than B or not. left (Variable): RankNet's output score for doc A. @@ -5435,7 +5465,7 @@ def flatten(x, axis=1, name=None): axis = 2 We get: Out.shape = (3 * 100, 4 * 100) - + Case 2: Given X.shape = (3, 100, 100, 4) @@ -5446,8 +5476,8 @@ def flatten(x, axis=1, name=None): Args: x (Variable): A tensor of rank >= axis. - axis (int): Indicate up to which input dimensions (exclusive) should - be flattened to the outer dimension of the output. + axis (int): Indicate up to which input dimensions (exclusive) should + be flattened to the outer dimension of the output. The value for axis must be in the range [0, R], where R is the rank of the input tensor. When axis = 0, the shape of the output tensor is (1, (d_0 X d_1 ... d_n), where the @@ -5463,7 +5493,7 @@ def flatten(x, axis=1, name=None): Raises: ValueError: If x is not a variable. - ValueError: If axis is not in range [0, rank(x)]. + ValueError: If axis is not in range [0, rank(x)]. Examples: diff --git a/python/paddle/fluid/layers/ops.py b/python/paddle/fluid/layers/ops.py index f70c7f2258..7cd62efda8 100644 --- a/python/paddle/fluid/layers/ops.py +++ b/python/paddle/fluid/layers/ops.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function from .layer_function_generator import generate_layer_fn __activations__ = [ @@ -63,7 +65,6 @@ __all__ = [ 'uniform_random_batch_size_like', 'gaussian_random', 'gaussian_random_batch_size_like', - 'scatter', 'sum', 'slice', 'shape', diff --git a/python/paddle/fluid/layers/tensor.py b/python/paddle/fluid/layers/tensor.py index b93d721c12..04e71497aa 100644 --- a/python/paddle/fluid/layers/tensor.py +++ b/python/paddle/fluid/layers/tensor.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from ..layer_helper import LayerHelper from ..param_attr import ParamAttr from ..framework import convert_np_dtype_to_dtype_ diff --git a/python/paddle/fluid/layers/utils.py b/python/paddle/fluid/layers/utils.py index 49ec308883..5688f04ab2 100644 --- a/python/paddle/fluid/layers/utils.py +++ b/python/paddle/fluid/layers/utils.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import numpy as np diff --git a/python/paddle/fluid/lod_tensor.py b/python/paddle/fluid/lod_tensor.py index 53c33616f5..a9de09f31f 100644 --- a/python/paddle/fluid/lod_tensor.py +++ b/python/paddle/fluid/lod_tensor.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import core import numpy as np diff --git a/python/paddle/fluid/metrics.py b/python/paddle/fluid/metrics.py index cd89345227..592cb23eb9 100644 --- a/python/paddle/fluid/metrics.py +++ b/python/paddle/fluid/metrics.py @@ -14,11 +14,15 @@ """ Fluid Metrics -The metrics are accomplished via Python natively. +The metrics are accomplished via Python natively. """ + +from __future__ import print_function + import numpy as np import copy import warnings +import six __all__ = [ 'MetricBase', @@ -79,10 +83,10 @@ class MetricBase(object): """ states = { attr: value - for attr, value in list(self.__dict__.items()) + for attr, value in six.iteritems(self.__dict__) if not attr.startswith("_") } - for attr, value in list(states.items()): + for attr, value in six.iteritems(states): if isinstance(value, int): setattr(self, attr, 0) elif isinstance(value, float): @@ -105,7 +109,7 @@ class MetricBase(object): """ states = { attr: value - for attr, value in list(self.__dict__.items()) + for attr, value in six.iteritems(self.__dict__) if not attr.startswith("_") } config = {} @@ -141,10 +145,10 @@ class CompositeMetric(MetricBase): """ Composite multiple metrics in one instance. for example, merge F1, accuracy, recall into one Metric. - + Examples: .. code-block:: python - + labels = fluid.layers.data(name="data", shape=[1], dtype="int32") data = fluid.layers.data(name="data", shape=[32, 32], dtype="int32") pred = fluid.layers.fc(input=data, size=1000, act="tanh") diff --git a/python/paddle/fluid/net_drawer.py b/python/paddle/fluid/net_drawer.py index 623a7d3fd0..0b61c23d07 100644 --- a/python/paddle/fluid/net_drawer.py +++ b/python/paddle/fluid/net_drawer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import argparse import json import logging diff --git a/python/paddle/fluid/nets.py b/python/paddle/fluid/nets.py index 08480671d8..051fe84364 100644 --- a/python/paddle/fluid/nets.py +++ b/python/paddle/fluid/nets.py @@ -11,6 +11,9 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function +import six from . import layers __all__ = [ @@ -210,7 +213,7 @@ def img_conv_group(input, conv_with_batchnorm = __extend_list__(conv_with_batchnorm) conv_batchnorm_drop_rate = __extend_list__(conv_batchnorm_drop_rate) - for i in range(len(conv_num_filter)): + for i in six.moves.range(len(conv_num_filter)): local_conv_act = conv_act if conv_with_batchnorm[i]: local_conv_act = None diff --git a/python/paddle/fluid/op.py b/python/paddle/fluid/op.py index 93f021a360..667db10d3e 100644 --- a/python/paddle/fluid/op.py +++ b/python/paddle/fluid/op.py @@ -12,6 +12,9 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + +import numpy as np import six import paddle.fluid.core as core @@ -99,6 +102,8 @@ class OpDescCreationMethod(object): new_attr = op_desc.attrs.add() new_attr.name = attr.name new_attr.type = attr.type + if isinstance(user_defined_attr, np.ndarray): + user_defined_attr = user_defined_attr.tolist() if attr.type == framework_pb2.INT: new_attr.i = user_defined_attr elif attr.type == framework_pb2.FLOAT: diff --git a/python/paddle/fluid/optimizer.py b/python/paddle/fluid/optimizer.py index a07325f46a..031ddd09a0 100644 --- a/python/paddle/fluid/optimizer.py +++ b/python/paddle/fluid/optimizer.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import re from collections import defaultdict from paddle.fluid.framework import Program, Variable diff --git a/python/paddle/fluid/parallel_executor.py b/python/paddle/fluid/parallel_executor.py index 2a3555ebdd..a7765c9591 100644 --- a/python/paddle/fluid/parallel_executor.py +++ b/python/paddle/fluid/parallel_executor.py @@ -17,8 +17,10 @@ import multiprocessing from . import core from . import framework from . import executor +from .. import compat as cpt import warnings import sys +import six import os __all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy'] @@ -95,7 +97,7 @@ class ParallelExecutor(object): self._places = [] self._act_places = [] if use_cuda: - for i in range(core.get_cuda_device_count()): + for i in six.moves.range(core.get_cuda_device_count()): p = core.Place() self._act_places.append(core.CUDAPlace(i)) p.set_place(self._act_places[-1]) @@ -103,7 +105,7 @@ class ParallelExecutor(object): else: cpu_num = int( os.environ.get('CPU_NUM', multiprocessing.cpu_count())) - for i in range(cpu_num): + for i in six.moves.range(cpu_num): p = core.Place() self._act_places.append(core.CPUPlace()) p.set_place(self._act_places[-1]) @@ -153,11 +155,13 @@ class ParallelExecutor(object): self.executor = core.ParallelExecutor( self._places, set([ - p.name for p in main.global_block().iter_parameters() + cpt.to_text(p.name) + for p in main.global_block().iter_parameters() if not p.stop_gradient ]), - set(self.persistable_vars), main.desc, loss_name - if loss_name else '', scope, local_scopes, exec_strategy, + set(cpt.to_text(var) for var in self.persistable_vars), main.desc, + cpt.to_text(loss_name) + if loss_name else six.u(''), scope, local_scopes, exec_strategy, build_strategy, num_trainers, trainer_id) self.scope = scope diff --git a/python/paddle/fluid/param_attr.py b/python/paddle/fluid/param_attr.py index afae577656..f0be794327 100644 --- a/python/paddle/fluid/param_attr.py +++ b/python/paddle/fluid/param_attr.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import six from .initializer import Initializer, Xavier, Constant diff --git a/python/paddle/fluid/profiler.py b/python/paddle/fluid/profiler.py index 01983a8303..e05885f5f5 100644 --- a/python/paddle/fluid/profiler.py +++ b/python/paddle/fluid/profiler.py @@ -12,9 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import core from contextlib import contextmanager import os +import six __all__ = [ 'cuda_profiler', 'reset_profiler', 'profiler', 'start_profiler', @@ -88,7 +91,7 @@ def cuda_profiler(output_file, output_mode=None, config=None): config = NVPROF_CONFIG if config is None else config config_file = 'nvprof_config_file' with open(config_file, 'wb') as fp: - fp.writelines(["%s\n" % item for item in config]) + fp.writelines([six.b("%s\n" % item) for item in config]) core.nvprof_init(output_file, output_mode, config_file) # Enables profiler collection by the active CUDA profiling tool. core.nvprof_start() diff --git a/python/paddle/fluid/recordio_writer.py b/python/paddle/fluid/recordio_writer.py index 93b38ad3fa..a69c0c29d4 100644 --- a/python/paddle/fluid/recordio_writer.py +++ b/python/paddle/fluid/recordio_writer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import os import contextlib from . import core diff --git a/python/paddle/fluid/regularizer.py b/python/paddle/fluid/regularizer.py index 6eaac4432d..da38626111 100644 --- a/python/paddle/fluid/regularizer.py +++ b/python/paddle/fluid/regularizer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from . import framework from . import core diff --git a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py index 36a1a223cf..f6017a455d 100644 --- a/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import contextlib diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py index 9e4c384d92..48c0f3d361 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/cifar10_small_test_set.py @@ -28,12 +28,14 @@ images per class. """ +from __future__ import print_function + import itertools import numpy -import paddle.v2.dataset.common +import paddle.dataset.common import tarfile +import six from six.moves import cPickle as pickle -from six.moves import zip __all__ = ['train10'] @@ -44,20 +46,25 @@ CIFAR10_MD5 = 'c58f30108f718f92721af3b95e74349a' def reader_creator(filename, sub_name, batch_size=None): def read_batch(batch): - data = batch['data'] - labels = batch.get('labels', batch.get('fine_labels', None)) + data = batch[six.b('data')] + labels = batch.get( + six.b('labels'), batch.get(six.b('fine_labels'), None)) assert labels is not None - for sample, label in zip(data, labels): + for sample, label in six.moves.zip(data, labels): yield (sample / 255.0).astype(numpy.float32), int(label) def reader(): with tarfile.open(filename, mode='r') as f: - names = (each_item.name for each_item in f - if sub_name in each_item.name) + names = [ + each_item.name for each_item in f if sub_name in each_item.name + ] batch_count = 0 for name in names: - batch = pickle.load(f.extractfile(name)) + if six.PY2: + batch = pickle.load(f.extractfile(name)) + else: + batch = pickle.load(f.extractfile(name), encoding='bytes') for item in read_batch(batch): if isinstance(batch_size, int) and batch_count > batch_size: break @@ -78,6 +85,6 @@ def train10(batch_size=None): :rtype: callable """ return reader_creator( - paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), + paddle.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), 'data_batch', batch_size=batch_size) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py index a1f62db093..be494a0d34 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_resnet.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import numpy @@ -55,7 +57,7 @@ def resnet_cifar10(input, depth=32): return tmp assert (depth - 2) % 6 == 0 - n = (depth - 2) / 6 + n = (depth - 2) // 6 conv1 = conv_bn_layer( input=input, ch_out=16, filter_size=3, stride=1, padding=1) res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) diff --git a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py index 8429551765..dbc7bc06c9 100644 --- a/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py +++ b/python/paddle/fluid/tests/book/high-level-api/image_classification/test_image_classification_vgg.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import numpy diff --git a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py index e3602e2d56..ec4e1c768c 100755 --- a/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py +++ b/python/paddle/fluid/tests/book/high-level-api/label_semantic_roles/test_label_semantic_roles_newapi.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import numpy as np diff --git a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py index 6fb0c85a8b..560f118958 100644 --- a/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/high-level-api/machine_translation/test_machine_translation.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import contextlib import numpy as np diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py index 898807db6f..187bef1b0c 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import argparse import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py index 6dd64be315..b95e7db122 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py +++ b/python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import argparse import paddle.fluid as fluid import paddle diff --git a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py index 60f3d8e105..9e2767783b 100644 --- a/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py +++ b/python/paddle/fluid/tests/book/high-level-api/recommender_system/test_recommender_system_newapi.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import math import sys import numpy as np diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py index 24e65d1bd5..097c2a468f 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_conv.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid from functools import partial diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py index b3b1505a0f..5f74cd1425 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_dynamic_rnn.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid from functools import partial diff --git a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py index 25f99ff0fd..284a6ca168 100644 --- a/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py +++ b/python/paddle/fluid/tests/book/high-level-api/understand_sentiment/test_understand_sentiment_stacked_lstm.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid from functools import partial diff --git a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py index 02e65cf56c..1c7cf3199a 100644 --- a/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py +++ b/python/paddle/fluid/tests/book/high-level-api/word2vec/test_word2vec_new_api.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import numpy as np diff --git a/python/paddle/fluid/tests/book/notest_understand_sentiment.py b/python/paddle/fluid/tests/book/notest_understand_sentiment.py index ce6342c2da..82f1c6615f 100644 --- a/python/paddle/fluid/tests/book/notest_understand_sentiment.py +++ b/python/paddle/fluid/tests/book/notest_understand_sentiment.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from paddle.fluid.layers.device import get_places import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/book/test_fit_a_line.py b/python/paddle/fluid/tests/book/test_fit_a_line.py index 37b64fa94a..334294ab48 100644 --- a/python/paddle/fluid/tests/book/test_fit_a_line.py +++ b/python/paddle/fluid/tests/book/test_fit_a_line.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import contextlib diff --git a/python/paddle/fluid/tests/book/test_image_classification.py b/python/paddle/fluid/tests/book/test_image_classification.py index de6fe5f140..9fe361425c 100644 --- a/python/paddle/fluid/tests/book/test_image_classification.py +++ b/python/paddle/fluid/tests/book/test_image_classification.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import contextlib @@ -60,7 +62,7 @@ def resnet_cifar10(input, depth=32): return tmp assert (depth - 2) % 6 == 0 - n = (depth - 2) / 6 + n = (depth - 2) // 6 conv1 = conv_bn_layer( input=input, ch_out=16, filter_size=3, stride=1, padding=1) res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) diff --git a/python/paddle/fluid/tests/book/test_label_semantic_roles.py b/python/paddle/fluid/tests/book/test_label_semantic_roles.py index b7ac911caf..f63387a906 100644 --- a/python/paddle/fluid/tests/book/test_label_semantic_roles.py +++ b/python/paddle/fluid/tests/book/test_label_semantic_roles.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import contextlib import math import numpy as np diff --git a/python/paddle/fluid/tests/book/test_machine_translation.py b/python/paddle/fluid/tests/book/test_machine_translation.py index 462faad3e1..5e241aaa32 100644 --- a/python/paddle/fluid/tests/book/test_machine_translation.py +++ b/python/paddle/fluid/tests/book/test_machine_translation.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import contextlib import numpy as np diff --git a/python/paddle/fluid/tests/book/test_recognize_digits.py b/python/paddle/fluid/tests/book/test_recognize_digits.py index 3e5f76d12d..da216d0cc4 100644 --- a/python/paddle/fluid/tests/book/test_recognize_digits.py +++ b/python/paddle/fluid/tests/book/test_recognize_digits.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid.core as core import math import os diff --git a/python/paddle/fluid/tests/book/test_recommender_system.py b/python/paddle/fluid/tests/book/test_recommender_system.py index b30c8771fc..cf8c48f346 100644 --- a/python/paddle/fluid/tests/book/test_recommender_system.py +++ b/python/paddle/fluid/tests/book/test_recommender_system.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import math import sys import os diff --git a/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py b/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py index 2e79be2bd0..91c8705aa4 100644 --- a/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py +++ b/python/paddle/fluid/tests/book/test_rnn_encoder_decoder.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import paddle import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/book/test_word2vec.py b/python/paddle/fluid/tests/book/test_word2vec.py index e761e05795..fe063eb462 100644 --- a/python/paddle/fluid/tests/book/test_word2vec.py +++ b/python/paddle/fluid/tests/book/test_word2vec.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid from paddle.fluid.layers.device import get_places diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py index ccc62b442f..f530f8f488 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_fit_a_line.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import math import sys diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py index b2a59d27da..3951e7b8ca 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_image_classification_train.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import sys import paddle @@ -56,7 +58,7 @@ def resnet_cifar10(input, depth=32): return tmp assert (depth - 2) % 6 == 0 - n = (depth - 2) / 6 + n = (depth - 2) // 6 conv1 = conv_bn_layer( input=input, ch_out=16, filter_size=3, stride=1, padding=1) res1 = layer_warp(basicblock, conv1, 16, 16, n, 1) diff --git a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py index 323ddfb691..1ad51936b5 100644 --- a/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py +++ b/python/paddle/fluid/tests/book_memory_optimization/test_memopt_machine_translation.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import paddle import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/demo/fc_gan.py b/python/paddle/fluid/tests/demo/fc_gan.py index 3d92f50f0a..bd77779ce6 100644 --- a/python/paddle/fluid/tests/demo/fc_gan.py +++ b/python/paddle/fluid/tests/demo/fc_gan.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import errno import math import os diff --git a/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py b/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py index a00325d79b..45a104ec96 100644 --- a/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py +++ b/python/paddle/fluid/tests/demo/file_reader/convert_data_to_recordio.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import sys import paddle.fluid as fluid import paddle.v2 as paddle diff --git a/python/paddle/fluid/tests/demo/file_reader/train.py b/python/paddle/fluid/tests/demo/file_reader/train.py index bc3a6dc81d..5f5d2848da 100644 --- a/python/paddle/fluid/tests/demo/file_reader/train.py +++ b/python/paddle/fluid/tests/demo/file_reader/train.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import numpy import sys diff --git a/python/paddle/fluid/tests/demo/pyreader.py b/python/paddle/fluid/tests/demo/pyreader.py index 8206540193..ec61e0ebae 100644 --- a/python/paddle/fluid/tests/demo/pyreader.py +++ b/python/paddle/fluid/tests/demo/pyreader.py @@ -12,7 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy +import six import paddle import paddle.dataset.mnist as mnist @@ -31,7 +34,7 @@ def network(is_train): hidden = img - for i in xrange(2): + for i in six.moves.xrange(2): hidden = fluid.layers.fc(input=hidden, size=100, act='tanh') hidden = fluid.layers.dropout( hidden, dropout_prob=0.5, is_test=not is_train) @@ -74,7 +77,7 @@ def main(): test_reader.decorate_paddle_reader(paddle.batch(mnist.test(), 512)) - for epoch_id in xrange(10): + for epoch_id in six.moves.xrange(10): train_reader.start() try: while True: diff --git a/python/paddle/fluid/tests/no_test_concurrency.py b/python/paddle/fluid/tests/no_test_concurrency.py index 3bc0c9808e..b5d7676f4a 100644 --- a/python/paddle/fluid/tests/no_test_concurrency.py +++ b/python/paddle/fluid/tests/no_test_concurrency.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/notest_concurrency.py b/python/paddle/fluid/tests/notest_concurrency.py index 77107f8b36..fd9da4cce0 100644 --- a/python/paddle/fluid/tests/notest_concurrency.py +++ b/python/paddle/fluid/tests/notest_concurrency.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/test_beam_search_decoder.py b/python/paddle/fluid/tests/test_beam_search_decoder.py index 8bf750940d..fe8a9daa3b 100644 --- a/python/paddle/fluid/tests/test_beam_search_decoder.py +++ b/python/paddle/fluid/tests/test_beam_search_decoder.py @@ -15,6 +15,8 @@ A simple machine translation demo using beam search decoder. """ +from __future__ import print_function + import contextlib import numpy as np import paddle diff --git a/python/paddle/fluid/tests/test_cpp_reader.py b/python/paddle/fluid/tests/test_cpp_reader.py index 6cc291dfcf..b2a5253b95 100644 --- a/python/paddle/fluid/tests/test_cpp_reader.py +++ b/python/paddle/fluid/tests/test_cpp_reader.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import numpy as np diff --git a/python/paddle/fluid/tests/test_data_feeder.py b/python/paddle/fluid/tests/test_data_feeder.py index 30b7a634a2..01de564aa4 100644 --- a/python/paddle/fluid/tests/test_data_feeder.py +++ b/python/paddle/fluid/tests/test_data_feeder.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import unittest diff --git a/python/paddle/fluid/tests/test_detection.py b/python/paddle/fluid/tests/test_detection.py index fd45abd0a7..1467e72caa 100644 --- a/python/paddle/fluid/tests/test_detection.py +++ b/python/paddle/fluid/tests/test_detection.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import paddle.fluid.layers as layers from paddle.fluid.framework import Program, program_guard diff --git a/python/paddle/fluid/tests/test_error_clip.py b/python/paddle/fluid/tests/test_error_clip.py index e8edd7fbbb..3c977afc7c 100644 --- a/python/paddle/fluid/tests/test_error_clip.py +++ b/python/paddle/fluid/tests/test_error_clip.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import paddle import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/test_gradient_clip.py b/python/paddle/fluid/tests/test_gradient_clip.py index d530601f13..266687fcd0 100644 --- a/python/paddle/fluid/tests/test_gradient_clip.py +++ b/python/paddle/fluid/tests/test_gradient_clip.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import paddle import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/test_if_else_op.py b/python/paddle/fluid/tests/test_if_else_op.py index 082f64c146..61d81f4836 100644 --- a/python/paddle/fluid/tests/test_if_else_op.py +++ b/python/paddle/fluid/tests/test_if_else_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid.layers as layers from paddle.fluid.framework import Program, program_guard @@ -28,7 +30,8 @@ import numpy as np class TestMNISTIfElseOp(unittest.TestCase): - def test_raw_api(self): + # FIXME: https://github.com/PaddlePaddle/Paddle/issues/12245#issuecomment-406462379 + def not_test_raw_api(self): prog = Program() startup_prog = Program() with program_guard(prog, startup_prog): @@ -89,7 +92,8 @@ class TestMNISTIfElseOp(unittest.TestCase): return self.assertFalse(True) - def test_ifelse(self): + # FIXME: https://github.com/PaddlePaddle/Paddle/issues/12245#issuecomment-406462379 + def not_test_ifelse(self): prog = Program() startup_prog = Program() with program_guard(prog, startup_prog): @@ -151,6 +155,13 @@ class TestIfElse(unittest.TestCase): self.cond_value = 0.5 self.data = np.random.rand(25, 1).astype(np.float32) + def numpy_cal(self): + s1 = self.data[np.where(self.data < self.cond_value)] + res = np.sum(np.exp(s1)) + s2 = self.data[np.where(self.data >= self.cond_value)] + res += np.sum(np.tanh(s2)) + return res + def compare_ifelse_op_and_numpy(self, place): self.set_test_case() @@ -164,10 +175,12 @@ class TestIfElse(unittest.TestCase): ie = layers.IfElse(ifcond) with ie.true_block(): true_target = ie.input(src) + true_target = fluid.layers.exp(true_target) ie.output(true_target) with ie.false_block(): false_target = ie.input(src) + false_target = fluid.layers.tanh(false_target) ie.output(false_target) if_out = ie() out = layers.reduce_sum(if_out) @@ -178,7 +191,8 @@ class TestIfElse(unittest.TestCase): o1, = exe.run(fluid.default_main_program(), feed={'data': self.data}, fetch_list=[out]) - o2 = np.sum(self.data) + o2 = self.numpy_cal() + self.assertTrue( np.allclose( o1, o2, atol=1e-8), diff --git a/python/paddle/fluid/tests/test_lod_tensor.py b/python/paddle/fluid/tests/test_lod_tensor.py index f7a9dd4129..722b5f07b0 100644 --- a/python/paddle/fluid/tests/test_lod_tensor.py +++ b/python/paddle/fluid/tests/test_lod_tensor.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid from paddle.fluid.lod_tensor import create_lod_tensor, create_random_int_lodtensor import numpy as np diff --git a/python/paddle/fluid/tests/test_python_operator_overriding.py b/python/paddle/fluid/tests/test_python_operator_overriding.py index b5ac97eac5..5f92c437ec 100644 --- a/python/paddle/fluid/tests/test_python_operator_overriding.py +++ b/python/paddle/fluid/tests/test_python_operator_overriding.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/benchmark.py b/python/paddle/fluid/tests/unittests/benchmark.py index b98a92dcbe..9ea95f3e87 100644 --- a/python/paddle/fluid/tests/unittests/benchmark.py +++ b/python/paddle/fluid/tests/unittests/benchmark.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import unittest import time @@ -54,7 +56,7 @@ class BenchmarkSuite(OpTest): def _get_input_names(self): inputs = [] - for name, value in list(self.inputs.items()): + for name, value in six.iteritems(self.inputs): if isinstance(value, list): inputs.extend([sub_name for sub_name, _ in value]) inputs.append(name) @@ -62,7 +64,7 @@ class BenchmarkSuite(OpTest): def _get_output_names(self): outputs = [] - for var_name, var in list(self.outputs.items()): + for var_name, var in six.iteritems(self.outputs): if isinstance(var, list): for sub_var_name, sub_var in var: outputs.append(sub_var_name) diff --git a/python/paddle/fluid/tests/unittests/benchmark_sum_op.py b/python/paddle/fluid/tests/unittests/benchmark_sum_op.py index 91a5f1bca4..0e7338b839 100644 --- a/python/paddle/fluid/tests/unittests/benchmark_sum_op.py +++ b/python/paddle/fluid/tests/unittests/benchmark_sum_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/decorators.py b/python/paddle/fluid/tests/unittests/decorators.py index d1165e2a91..1a5f4540cf 100644 --- a/python/paddle/fluid/tests/unittests/decorators.py +++ b/python/paddle/fluid/tests/unittests/decorators.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid __all__ = ['many_times', 'prog_scope'] diff --git a/python/paddle/fluid/tests/unittests/dist_mnist.py b/python/paddle/fluid/tests/unittests/dist_mnist.py index 8f5ba33f7c..85a96c0b53 100644 --- a/python/paddle/fluid/tests/unittests/dist_mnist.py +++ b/python/paddle/fluid/tests/unittests/dist_mnist.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import argparse import time @@ -44,7 +46,8 @@ def cnn_model(data): pool_size=2, pool_stride=2, act="relu", - param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant())) + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant( + value=0.3))) conv_pool_2 = fluid.nets.simple_img_conv_pool( input=conv_pool_1, filter_size=5, @@ -52,7 +55,8 @@ def cnn_model(data): pool_size=2, pool_stride=2, act="relu", - param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant())) + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Constant( + value=0.2))) SIZE = 10 input_shape = conv_pool_2.shape @@ -64,8 +68,7 @@ def cnn_model(data): size=SIZE, act="softmax", param_attr=fluid.param_attr.ParamAttr( - initializer=fluid.initializer.NormalInitializer( - loc=0.0, scale=scale, seed=1))) + initializer=fluid.initializer.Constant(value=0.1))) return predict diff --git a/python/paddle/fluid/tests/unittests/dist_se_resnext.py b/python/paddle/fluid/tests/unittests/dist_se_resnext.py index d576a173ce..0387e91188 100644 --- a/python/paddle/fluid/tests/unittests/dist_se_resnext.py +++ b/python/paddle/fluid/tests/unittests/dist_se_resnext.py @@ -12,9 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import argparse -import six import time import math @@ -128,7 +129,12 @@ class SE_ResNeXt(): input=conv, pool_size=7, pool_type='avg', global_pooling=True) drop = fluid.layers.dropout(x=pool, dropout_prob=0.2) stdv = 1.0 / math.sqrt(drop.shape[1] * 1.0) - out = fluid.layers.fc(input=drop, size=class_dim, act='softmax') + out = fluid.layers.fc( + input=drop, + size=class_dim, + act='softmax', + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Constant(value=0.2))) return out def shortcut(self, input, ch_out, stride): @@ -173,12 +179,12 @@ class SE_ResNeXt(): num_filters=num_filters, filter_size=filter_size, stride=stride, - padding=(filter_size - 1) / 2, + padding=(filter_size - 1) // 2, groups=groups, act=None, # avoid pserver CPU init differs from GPU param_attr=fluid.ParamAttr( - initializer=fluid.initializer.Constant()), + initializer=fluid.initializer.Constant(value=0.2)), bias_attr=False) return fluid.layers.batch_norm(input=conv, act=act) @@ -187,7 +193,7 @@ class SE_ResNeXt(): input=input, pool_size=0, pool_type='avg', global_pooling=True) stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0) squeeze = fluid.layers.fc(input=pool, - size=num_channels / reduction_ratio, + size=num_channels // reduction_ratio, act='relu') stdv = 1.0 / math.sqrt(squeeze.shape[1] * 1.0) excitation = fluid.layers.fc(input=squeeze, @@ -227,10 +233,8 @@ class DistSeResneXt2x2(TestDistRunnerBase): lr = [base_lr * (0.1**i) for i in range(len(bd) + 1)] optimizer = fluid.optimizer.Momentum( - # FIXME(typhoonzero): add back LR decay once ParallelExecutor fixed. - #learning_rate=fluid.layers.piecewise_decay( - # boundaries=bd, values=lr), - learning_rate=base_lr, + learning_rate=fluid.layers.piecewise_decay( + boundaries=bd, values=lr), momentum=0.9, regularization=fluid.regularizer.L2Decay(1e-4)) optimizer.minimize(avg_cost) diff --git a/python/paddle/fluid/tests/unittests/dist_transformer.py b/python/paddle/fluid/tests/unittests/dist_transformer.py index ee8020a735..239adcb9d5 100644 --- a/python/paddle/fluid/tests/unittests/dist_transformer.py +++ b/python/paddle/fluid/tests/unittests/dist_transformer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import argparse import time @@ -22,6 +24,7 @@ import paddle.fluid as fluid from paddle.fluid import core import os import sys +import six import transformer_model import paddle.dataset.wmt16 as wmt16 @@ -159,6 +162,7 @@ def get_model(): avg_cost = transformer(use_feed=False) optimizer = fluid.optimizer.Adam() optimizer.minimize(avg_cost) + fluid.memory_optimize(fluid.default_main_program()) return avg_cost @@ -222,7 +226,7 @@ class DistTransformer2x2(object): first_loss, = exe.run(fetch_list=[avg_cost.name]) print(first_loss) - for i in xrange(5): + for i in six.moves.xrange(5): _ = exe.run(fetch_list=[avg_cost.name]) last_loss, = exe.run(fetch_list=[avg_cost.name]) print(last_loss) @@ -261,9 +265,9 @@ def main(role="pserver", if __name__ == "__main__": - if len(sys.argv) != 7: + if len(sys.argv) != 8: print( - "Usage: python dist_transformer.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist]" + "Usage: python dist_transformer.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist] [sync_mode]" ) role = sys.argv[1] endpoints = sys.argv[2] @@ -271,6 +275,8 @@ if __name__ == "__main__": current_endpoint = sys.argv[4] trainers = int(sys.argv[5]) is_dist = True if sys.argv[6] == "TRUE" else False + # FIXME(typhoonzero): refine this test. + is_async = True if sys.argv[7] == "TRUE" else False main( role=role, endpoints=endpoints, diff --git a/python/paddle/fluid/tests/unittests/dist_word2vec.py b/python/paddle/fluid/tests/unittests/dist_word2vec.py index 54a70f4adb..0ad994a258 100644 --- a/python/paddle/fluid/tests/unittests/dist_word2vec.py +++ b/python/paddle/fluid/tests/unittests/dist_word2vec.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import argparse import time diff --git a/python/paddle/fluid/tests/unittests/op_test.py b/python/paddle/fluid/tests/unittests/op_test.py index b27d773f09..972e44c952 100644 --- a/python/paddle/fluid/tests/unittests/op_test.py +++ b/python/paddle/fluid/tests/unittests/op_test.py @@ -12,9 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import random +import six import time import itertools import collections @@ -26,15 +29,13 @@ from paddle.fluid.op import Operator from paddle.fluid.executor import Executor from paddle.fluid.framework import Program, OpProtoHolder, Variable from testsuite import create_op, set_input, append_input_output, append_loss_ops -from functools import reduce -from six.moves import zip def randomize_probability(batch_size, class_num, dtype='float32'): prob = np.random.uniform( 0.1, 1.0, size=(batch_size, class_num)).astype(dtype) prob_sum = prob.sum(axis=1) - for i in range(len(prob)): + for i in six.moves.xrange(len(prob)): prob[i] /= prob_sum[i] return prob @@ -51,7 +52,7 @@ def get_numeric_gradient(place, set_input(scope, op, inputs, place) def product(dim): - return reduce(lambda a, b: a * b, dim, 1) + return six.moves.reduce(lambda a, b: a * b, dim, 1) def get_output(): sum = [] @@ -103,7 +104,7 @@ def get_numeric_gradient(place, # we only compute gradient of one element each time. # we use a for loop to compute the gradient of every element. - for i in range(tensor_size): + for i in six.moves.xrange(tensor_size): if in_place: set_input(scope, op, inputs, place) @@ -161,7 +162,7 @@ class OpTest(unittest.TestCase): assert isinstance( numpy_dict, dict), "self.inputs, self.outputs must be numpy_dict" - for var_name, var_value in numpy_dict.items(): + for var_name, var_value in six.iteritems(numpy_dict): if isinstance(var_value, (np.ndarray, np.generic)): self.try_call_once(var_value.dtype) elif isinstance(var_value, (list, tuple)): @@ -225,7 +226,7 @@ class OpTest(unittest.TestCase): def _get_io_vars(self, block, numpy_inputs): inputs = {} - for name, value in numpy_inputs.items(): + for name, value in six.iteritems(numpy_inputs): if isinstance(value, list): var_list = [ block.var(sub_name) for sub_name, sub_value in value @@ -268,7 +269,7 @@ class OpTest(unittest.TestCase): # if the fetch_list is customized by user, we use it directly. # if not, fill the fetch_list by the user configured outputs in test. if len(fetch_list) == 0: - for var_name, var in outputs.items(): + for var_name, var in six.iteritems(outputs): if isinstance(var, list): for v in var: fetch_list.append(v) @@ -366,12 +367,13 @@ class OpTest(unittest.TestCase): for place in places: outs = self.calc_output(place) outs = [np.array(out) for out in outs] + outs.sort(key=len) checker(outs) def __assert_is_close(self, numeric_grads, analytic_grads, names, max_relative_error, msg_prefix): - for a, b, name in zip(numeric_grads, analytic_grads, names): + for a, b, name in six.moves.zip(numeric_grads, analytic_grads, names): abs_a = np.abs(a) abs_a[abs_a < 1e-3] = 1 diff --git a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py index 67c35e9de7..74e9d5c5f9 100644 --- a/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py +++ b/python/paddle/fluid/tests/unittests/parallel_executor_test_base.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import multiprocessing import os import unittest @@ -36,7 +38,8 @@ class TestParallelExecutorBase(unittest.TestCase): seed=None, use_parallel_executor=True, use_reduce=False, - optimizer=fluid.optimizer.Adam): + optimizer=fluid.optimizer.Adam, + use_fast_executor=False): def run_executor(exe, feed, fetch_list, program=None): if isinstance(exe, fluid.ParallelExecutor): res = exe.run(fetch_list=fetch_list, feed=feed) @@ -69,6 +72,8 @@ class TestParallelExecutorBase(unittest.TestCase): startup_exe.run(startup) exec_strategy = fluid.ExecutionStrategy() exec_strategy.allow_op_delay = allow_op_delay + if use_fast_executor: + exec_strategy.use_experimental_executor = True build_strategy = fluid.BuildStrategy() build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce \ diff --git a/python/paddle/fluid/tests/unittests/test_accuracy_op.py b/python/paddle/fluid/tests/unittests/test_accuracy_op.py index db1861fd10..1b2b53f2d4 100644 --- a/python/paddle/fluid/tests/unittests/test_accuracy_op.py +++ b/python/paddle/fluid/tests/unittests/test_accuracy_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py index 7d554c2276..611d0dd076 100644 --- a/python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_activation_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_activation_op.py b/python/paddle/fluid/tests/unittests/test_activation_op.py index 34f9cf0620..30651c1326 100644 --- a/python/paddle/fluid/tests/unittests/test_activation_op.py +++ b/python/paddle/fluid/tests/unittests/test_activation_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_adadelta_op.py b/python/paddle/fluid/tests/unittests/test_adadelta_op.py index 1b892e64c7..969a7da3b7 100644 --- a/python/paddle/fluid/tests/unittests/test_adadelta_op.py +++ b/python/paddle/fluid/tests/unittests/test_adadelta_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_adagrad_op.py b/python/paddle/fluid/tests/unittests/test_adagrad_op.py index 2f0ea79f4d..fc3b7ce2fd 100644 --- a/python/paddle/fluid/tests/unittests/test_adagrad_op.py +++ b/python/paddle/fluid/tests/unittests/test_adagrad_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_adam_op.py b/python/paddle/fluid/tests/unittests/test_adam_op.py index fa4b39879c..5318d2f976 100644 --- a/python/paddle/fluid/tests/unittests/test_adam_op.py +++ b/python/paddle/fluid/tests/unittests/test_adam_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_adamax_op.py b/python/paddle/fluid/tests/unittests/test_adamax_op.py index 8099beefa5..a6d1be7616 100644 --- a/python/paddle/fluid/tests/unittests/test_adamax_op.py +++ b/python/paddle/fluid/tests/unittests/test_adamax_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py b/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py index 9c7d5d41f0..d31eaa0114 100644 --- a/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py +++ b/python/paddle/fluid/tests/unittests/test_anchor_generator_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import sys diff --git a/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py b/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py index e04412f809..0712e102b3 100644 --- a/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py +++ b/python/paddle/fluid/tests/unittests/test_arg_min_max_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_argsort_op.py b/python/paddle/fluid/tests/unittests/test_argsort_op.py index b29a102a38..7bc6f2599d 100644 --- a/python/paddle/fluid/tests/unittests/test_argsort_op.py +++ b/python/paddle/fluid/tests/unittests/test_argsort_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_array_read_write_op.py b/python/paddle/fluid/tests/unittests/test_array_read_write_op.py index 0000fb0958..b86d0bc43a 100644 --- a/python/paddle/fluid/tests/unittests/test_array_read_write_op.py +++ b/python/paddle/fluid/tests/unittests/test_array_read_write_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core import paddle.fluid.layers as layers diff --git a/python/paddle/fluid/tests/unittests/test_assign_op.py b/python/paddle/fluid/tests/unittests/test_assign_op.py index e93c02bd3e..ba2eecfaf1 100644 --- a/python/paddle/fluid/tests/unittests/test_assign_op.py +++ b/python/paddle/fluid/tests/unittests/test_assign_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import op_test import numpy import unittest diff --git a/python/paddle/fluid/tests/unittests/test_assign_value_op.py b/python/paddle/fluid/tests/unittests/test_assign_value_op.py index 02f2e6eddc..5a9d8efef1 100644 --- a/python/paddle/fluid/tests/unittests/test_assign_value_op.py +++ b/python/paddle/fluid/tests/unittests/test_assign_value_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import paddle.fluid.layers as layers import op_test diff --git a/python/paddle/fluid/tests/unittests/test_auc_op.py b/python/paddle/fluid/tests/unittests/test_auc_op.py index 6580c70ca6..5393a17e67 100644 --- a/python/paddle/fluid/tests/unittests/test_auc_op.py +++ b/python/paddle/fluid/tests/unittests/test_auc_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py index 18fa546159..1286cee8dc 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py index f805fdc35f..80261eff4e 100644 --- a/python/paddle/fluid/tests/unittests/test_batch_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_batch_norm_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py index 4a3ac2a31e..51eee41ab2 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_decode_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_beam_search_op.py b/python/paddle/fluid/tests/unittests/test_beam_search_op.py index e8283fc942..c28dda4b53 100644 --- a/python/paddle/fluid/tests/unittests/test_beam_search_op.py +++ b/python/paddle/fluid/tests/unittests/test_beam_search_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import logging from paddle.fluid.op import Operator, DynamicRecurrentOp import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py index b04f25ef87..bed847c3c1 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_interp_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_bilinear_tensor_product_op.py b/python/paddle/fluid/tests/unittests/test_bilinear_tensor_product_op.py index d20a11e27e..46831119c5 100644 --- a/python/paddle/fluid/tests/unittests/test_bilinear_tensor_product_op.py +++ b/python/paddle/fluid/tests/unittests/test_bilinear_tensor_product_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py b/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py index ceeca25b74..5cc8e2ba15 100644 --- a/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py +++ b/python/paddle/fluid/tests/unittests/test_bipartite_match_op.py @@ -11,6 +11,8 @@ #WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. #See the License for the specific language governing permissions and #limitations under the License. + +from __future__ import print_function import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_box_coder_op.py b/python/paddle/fluid/tests/unittests/test_box_coder_op.py index 4ce9a4783e..2511c5c22e 100644 --- a/python/paddle/fluid/tests/unittests/test_box_coder_op.py +++ b/python/paddle/fluid/tests/unittests/test_box_coder_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import sys diff --git a/python/paddle/fluid/tests/unittests/test_calc_gradient.py b/python/paddle/fluid/tests/unittests/test_calc_gradient.py index 7f2a9e6971..4120a18b72 100644 --- a/python/paddle/fluid/tests/unittests/test_calc_gradient.py +++ b/python/paddle/fluid/tests/unittests/test_calc_gradient.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_cast_op.py b/python/paddle/fluid/tests/unittests/test_cast_op.py index b8d3ed3aa3..71a2ccb6da 100644 --- a/python/paddle/fluid/tests/unittests/test_cast_op.py +++ b/python/paddle/fluid/tests/unittests/test_cast_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import op_test import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py b/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py index 354110f1f9..48eb8e9f75 100644 --- a/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py +++ b/python/paddle/fluid/tests/unittests/test_chunk_eval_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py b/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py index 129958fa28..6103c3aafc 100644 --- a/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_clip_op.py b/python/paddle/fluid/tests/unittests/test_clip_op.py index 3df80c8ec8..32677bdb4c 100644 --- a/python/paddle/fluid/tests/unittests/test_clip_op.py +++ b/python/paddle/fluid/tests/unittests/test_clip_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_compare_op.py b/python/paddle/fluid/tests/unittests/test_compare_op.py index 405afebae8..437ad35538 100644 --- a/python/paddle/fluid/tests/unittests/test_compare_op.py +++ b/python/paddle/fluid/tests/unittests/test_compare_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import op_test import unittest import numpy diff --git a/python/paddle/fluid/tests/unittests/test_compat.py b/python/paddle/fluid/tests/unittests/test_compat.py new file mode 100644 index 0000000000..1c2c46f99a --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_compat.py @@ -0,0 +1,505 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import paddle.compat as cpt +import six + + +class TestCompatible(unittest.TestCase): + def test_type(self): + if six.PY2: + self.assertEqual(cpt.int_type, int) + self.assertEqual(cpt.long_type, long) + else: + self.assertEqual(cpt.int_type, int) + self.assertEqual(cpt.long_type, int) + + def test_to_text(self): + # Only support python2.x and python3.x now + self.assertTrue(six.PY2 | six.PY3) + + if six.PY2: + # check None + self.assertIsNone(cpt.to_text(None)) + + # check all string related types + self.assertTrue(isinstance(cpt.to_text(str("")), unicode)) + self.assertTrue(isinstance(cpt.to_text(str("123")), unicode)) + self.assertTrue(isinstance(cpt.to_text(b""), unicode)) + self.assertTrue(isinstance(cpt.to_text(b""), unicode)) + self.assertTrue(isinstance(cpt.to_text(u""), unicode)) + self.assertTrue(isinstance(cpt.to_text(u""), unicode)) + + self.assertEqual(u"", cpt.to_text(str(""))) + self.assertEqual(u"123", cpt.to_text(str("123"))) + self.assertEqual(u"", cpt.to_text(b"")) + self.assertEqual(u"123", cpt.to_text(b"123")) + self.assertEqual(u"", cpt.to_text(u"")) + self.assertEqual(u"123", cpt.to_text(u"123")) + + # check list types, not inplace + l = [""] + l2 = cpt.to_text(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([u""], l2) + l = ["", "123"] + l2 = cpt.to_text(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([u"", u"123"], l2) + l = ["", b'123', u"321"] + l2 = cpt.to_text(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([u"", u"123", u"321"], l2) + for i in l2: + self.assertTrue(isinstance(i, unicode)) + + # check list types, inplace + l = [""] + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([u""], l2) + l = ["", "123"] + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([u"", u"123"], l2) + l = ["", b"123", u"321"] + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([u"", u"123", u"321"], l2) + + # check set types, not inplace + l = set("") + l2 = cpt.to_text(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(u""), l2) + l = set([b"", b"123"]) + l2 = cpt.to_text(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([u"", u"123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_text(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([u"", u"123", u"321"]), l2) + for i in l2: + self.assertTrue(isinstance(i, unicode)) + + # check set types, inplace + l = set("") + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(u""), l2) + l = set([b"", b"123"]) + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([u"", u"123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([u"", u"123", u"321"]), l2) + + elif six.PY3: + self.assertIsNone(cpt.to_text(None)) + + self.assertTrue(isinstance(cpt.to_text(str("")), str)) + self.assertTrue(isinstance(cpt.to_text(str("123")), str)) + self.assertTrue(isinstance(cpt.to_text(b""), str)) + self.assertTrue(isinstance(cpt.to_text(b""), str)) + self.assertTrue(isinstance(cpt.to_text(u""), str)) + self.assertTrue(isinstance(cpt.to_text(u""), str)) + + self.assertEqual("", cpt.to_text(str(""))) + self.assertEqual("123", cpt.to_text(str("123"))) + self.assertEqual("", cpt.to_text(b"")) + self.assertEqual("123", cpt.to_text(b"123")) + self.assertEqual("", cpt.to_text(u"")) + self.assertEqual("123", cpt.to_text(u"123")) + + # check list types, not inplace + l = [""] + l2 = cpt.to_text(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([""], l2) + l = ["", "123"] + l2 = cpt.to_text(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(["", "123"], l2) + l = ["", b"123", u"321"] + l2 = cpt.to_text(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual(["", "123", "321"], l2) + + # check list types, inplace + l = [""] + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([""], l2) + l = ["", b"123"] + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(["", "123"], l2) + l = ["", b"123", u"321"] + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(["", "123", "321"], l2) + for i in l2: + self.assertTrue(isinstance(i, str)) + + # check set types, not inplace + l = set("") + l2 = cpt.to_text(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(""), l2) + l = set([b"", b"123"]) + l2 = cpt.to_text(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual(set(["", "123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_text(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual(set(["", "123", "321"]), l2) + + # check set types, inplace + l = set("") + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(""), l2) + l = set([b"", b"123"]) + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(["", "123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_text(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(["", "123", "321"]), l2) + for i in l2: + self.assertTrue(isinstance(i, str)) + + def test_to_bytes(self): + # Only support python2.x and python3.x now + self.assertTrue(six.PY2 | six.PY3) + + if six.PY2: + # check None + self.assertIsNone(cpt.to_bytes(None)) + + # check all string related types + self.assertTrue(isinstance(cpt.to_bytes(str("")), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(str("123")), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(b""), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(b""), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(u""), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(u""), bytes)) + + self.assertEqual(b"", cpt.to_bytes(str(""))) + self.assertEqual(b"123", cpt.to_bytes(str("123"))) + self.assertEqual(b"", cpt.to_bytes(b"")) + self.assertEqual(b"123", cpt.to_bytes(b"123")) + self.assertEqual(b"", cpt.to_bytes(u"")) + self.assertEqual(b"123", cpt.to_bytes(u"123")) + + # check list types, not inplace + l = [""] + l2 = cpt.to_bytes(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b""], l2) + l = ["", "123"] + l2 = cpt.to_bytes(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b"", b"123"], l2) + l = ["", b'123', u"321"] + l2 = cpt.to_bytes(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b"", b"123", b"321"], l2) + for i in l2: + self.assertTrue(isinstance(i, bytes)) + + # check list types, inplace + l = [""] + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b""], l2) + l = ["", "123"] + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b"", b"123"], l2) + l = ["", b"123", u"321"] + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b"", b"123", b"321"], l2) + + # check set types, not inplace + l = set("") + l2 = cpt.to_bytes(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(b""), l2) + l = set([b"", b"123"]) + l2 = cpt.to_bytes(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([b"", b"123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_bytes(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([b"", b"123", b"321"]), l2) + for i in l2: + self.assertTrue(isinstance(i, bytes)) + + # check set types, inplace + l = set("") + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(b""), l2) + l = set([b"", b"123"]) + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([b"", b"123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([b"", b"123", b"321"]), l2) + + elif six.PY3: + self.assertIsNone(cpt.to_bytes(None)) + + self.assertTrue(isinstance(cpt.to_bytes(str("")), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(str("123")), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(b""), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(b""), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(u""), bytes)) + self.assertTrue(isinstance(cpt.to_bytes(u""), bytes)) + + self.assertEqual(b"", cpt.to_bytes(str(""))) + self.assertEqual(b"123", cpt.to_bytes(str("123"))) + self.assertEqual(b"", cpt.to_bytes(b"")) + self.assertEqual(b"123", cpt.to_bytes(b"123")) + self.assertEqual(b"", cpt.to_bytes(u"")) + self.assertEqual(b"123", cpt.to_bytes(u"123")) + + # check list types, not inplace + l = [""] + l2 = cpt.to_bytes(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual([b""], l2) + l = ["", "123"] + l2 = cpt.to_bytes(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual([b"", b"123"], l2) + l = ["", b"123", u"321"] + l2 = cpt.to_bytes(l) + self.assertTrue(isinstance(l2, list)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual([b"", b"123", b"321"], l2) + + # check list types, inplace + l = [""] + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b""], l2) + l = ["", b"123"] + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b"", b"123"], l2) + l = ["", b"123", u"321"] + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, list)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual([b"", b"123", b"321"], l2) + for i in l2: + self.assertTrue(isinstance(i, bytes)) + + # check set types, not inplace + l = set([""]) + l2 = cpt.to_bytes(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual(set([b""]), l2) + l = set([u"", u"123"]) + l2 = cpt.to_bytes(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual(set([b"", b"123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_bytes(l, inplace=False) + self.assertTrue(isinstance(l2, set)) + self.assertFalse(l is l2) + self.assertNotEqual(l, l2) + self.assertEqual(set([b"", b"123", b"321"]), l2) + + # check set types, inplace + l = set("") + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set(b""), l2) + l = set([u"", u"123"]) + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([b"", b"123"]), l2) + l = set(["", b"123", u"321"]) + l2 = cpt.to_bytes(l, inplace=True) + self.assertTrue(isinstance(l2, set)) + self.assertTrue(l is l2) + self.assertEqual(l, l2) + self.assertEqual(set([b"", b"123", b"321"]), l2) + for i in l2: + self.assertTrue(isinstance(i, bytes)) + + def test_round(self): + self.assertEqual(3.0, cpt.round(3.4)) + self.assertEqual(4.0, cpt.round(3.5)) + self.assertEqual(0.0, cpt.round(0.1)) + self.assertEqual(0.0, cpt.round(0.0)) + self.assertEqual(-0.0, cpt.round(-0.0)) + self.assertEqual(-0.0, cpt.round(-0.1)) + self.assertEqual(-3.0, cpt.round(-3.4)) + self.assertEqual(-4.0, cpt.round(-3.5)) + self.assertEqual(5.0, cpt.round(5)) + self.assertRaises(TypeError, cpt.round, None) + + def test_floor_division(self): + self.assertEqual(0.0, cpt.floor_division(3, 4)) + self.assertEqual(1.0, cpt.floor_division(4, 3)) + self.assertEqual(2.0, cpt.floor_division(6, 3)) + self.assertEqual(-2.0, cpt.floor_division(-4, 3)) + self.assertEqual(-2.0, cpt.floor_division(-6, 3)) + self.assertRaises(ZeroDivisionError, cpt.floor_division, 3, 0) + self.assertRaises(TypeError, cpt.floor_division, None, None) + + def test_get_exception_message(self): + exception_message = "test_message" + self.assertRaises(AssertionError, cpt.get_exception_message, None) + if six.PY2: + self.assertRaises(AttributeError, cpt.get_exception_message, + exception_message) + try: + raise RuntimeError(exception_message) + except Exception as e: + self.assertEqual(exception_message, + cpt.get_exception_message(e)) + self.assertIsNotNone(e) + + try: + raise Exception(exception_message) + except Exception as e: + self.assertEqual(exception_message, + cpt.get_exception_message(e)) + self.assertIsNotNone(e) + + if six.PY3: + try: + raise RuntimeError(exception_message) + except Exception as e: + self.assertEqual(exception_message, + cpt.get_exception_message(e)) + self.assertIsNotNone(e) + + try: + raise Exception(exception_message) + except Exception as e: + self.assertEqual(exception_message, + cpt.get_exception_message(e)) + self.assertIsNotNone(e) + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_concat_op.py b/python/paddle/fluid/tests/unittests/test_concat_op.py index e9f3c45dc4..436ab7d49f 100644 --- a/python/paddle/fluid/tests/unittests/test_concat_op.py +++ b/python/paddle/fluid/tests/unittests/test_concat_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_conditional_block.py b/python/paddle/fluid/tests/unittests/test_conditional_block.py index 77869a1242..5b2b71d050 100644 --- a/python/paddle/fluid/tests/unittests/test_conditional_block.py +++ b/python/paddle/fluid/tests/unittests/test_conditional_block.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.layers as layers import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_const_value.py b/python/paddle/fluid/tests/unittests/test_const_value.py index 58ac6fa0a9..0b2431d772 100644 --- a/python/paddle/fluid/tests/unittests/test_const_value.py +++ b/python/paddle/fluid/tests/unittests/test_const_value.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.framework as framework diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py index d0de7ad52c..1902a98698 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from test_conv2d_op import TestConv2dOp, TestWithPad, TestWithStride diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_op.py index bb1cd87d61..6a2732e939 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -24,12 +26,12 @@ def conv2d_forward_naive(input, filter, group, conv_param): out_c, f_c, f_h, f_w = filter.shape assert f_c * group == in_c assert np.mod(out_c, group) == 0 - sub_out_c = out_c / group + sub_out_c = out_c // group stride, pad, dilation = conv_param['stride'], conv_param['pad'], conv_param[ 'dilation'] - out_h = 1 + (in_h + 2 * pad[0] - (dilation[0] * (f_h - 1) + 1)) / stride[0] - out_w = 1 + (in_w + 2 * pad[1] - (dilation[1] * (f_w - 1) + 1)) / stride[1] + out_h = 1 + (in_h + 2 * pad[0] - (dilation[0] * (f_h - 1) + 1)) // stride[0] + out_w = 1 + (in_w + 2 * pad[1] - (dilation[1] * (f_w - 1) + 1)) // stride[1] out = np.zeros((in_n, out_c, out_h, out_w)) d_bolck_h = (dilation[0] * (f_h - 1) + 1) @@ -138,7 +140,7 @@ class TestConv2dOp(OpTest): self.stride = [1, 1] self.input_size = [2, 3, 5, 5] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3] def init_dilation(self): @@ -157,7 +159,7 @@ class TestWithPad(TestConv2dOp): self.stride = [1, 1] self.input_size = [2, 3, 5, 5] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3] @@ -167,7 +169,7 @@ class TestWithStride(TestConv2dOp): self.stride = [2, 2] self.input_size = [2, 3, 6, 6] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3] @@ -182,7 +184,7 @@ class TestWith1x1(TestConv2dOp): self.stride = [1, 1] self.input_size = [2, 3, 5, 5] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 1, 1] def init_group(self): @@ -195,7 +197,7 @@ class TestWithDilation(TestConv2dOp): self.stride = [1, 1] self.input_size = [2, 3, 10, 10] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3] def init_dilation(self): @@ -211,7 +213,7 @@ class TestWithInput1x1Filter1x1(TestConv2dOp): self.stride = [1, 1] self.input_size = [2, 3, 1, 1] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 1, 1] def init_group(self): @@ -328,7 +330,7 @@ class TestDepthwiseConv(TestConv2dOp): self.input_size = [2, 3, 5, 5] # NCHW self.groups = 3 assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3] self.op_type = "depthwise_conv2d" @@ -340,7 +342,7 @@ class TestDepthwiseConv2(TestConv2dOp): self.input_size = [2, 3, 5, 5] # NCHW self.groups = 3 assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3] self.op_type = "depthwise_conv2d" diff --git a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py b/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py index af6cd99b0d..2a320e735b 100644 --- a/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -25,7 +27,7 @@ def conv2dtranspose_forward_naive(input_, filter_, attrs): groups = attrs['groups'] assert in_c == f_c out_c = f_out_c * groups - sub_in_c = in_c / groups + sub_in_c = in_c // groups stride, pad, dilations = attrs['strides'], attrs['paddings'], attrs[ 'dilations'] @@ -258,7 +260,7 @@ class TestDepthwiseConvTranspose(TestConv2dTransposeOp): self.input_size = [2, 8, 16, 16] # NCHW self.groups = 8 assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [self.input_size[1], f_c, 4, 4] self.op_type = "depthwise_conv2d_transpose" diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_op.py b/python/paddle/fluid/tests/unittests/test_conv3d_op.py index dd4ef7cc94..ddaf99fe06 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv3d_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -24,14 +26,14 @@ def conv3d_forward_naive(input, filter, group, conv_param): out_c, f_c, f_d, f_h, f_w = filter.shape assert f_c * group == in_c assert np.mod(out_c, group) == 0 - sub_out_c = out_c / group + sub_out_c = out_c // group stride, pad, dilation = conv_param['stride'], conv_param['pad'], conv_param[ 'dilations'] - out_d = 1 + (in_d + 2 * pad[0] - (dilation[0] * (f_d - 1) + 1)) / stride[0] - out_h = 1 + (in_h + 2 * pad[1] - (dilation[1] * (f_h - 1) + 1)) / stride[1] - out_w = 1 + (in_w + 2 * pad[2] - (dilation[2] * (f_w - 1) + 1)) / stride[2] + out_d = 1 + (in_d + 2 * pad[0] - (dilation[0] * (f_d - 1) + 1)) // stride[0] + out_h = 1 + (in_h + 2 * pad[1] - (dilation[1] * (f_h - 1) + 1)) // stride[1] + out_w = 1 + (in_w + 2 * pad[2] - (dilation[2] * (f_w - 1) + 1)) // stride[2] out = np.zeros((in_n, out_c, out_d, out_h, out_w)) @@ -166,7 +168,7 @@ class TestConv3dOp(OpTest): self.stride = [1, 1, 1] self.input_size = [2, 3, 4, 4, 4] # NCDHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3, 3] def init_dilation(self): @@ -185,7 +187,7 @@ class TestCase1(TestConv3dOp): self.stride = [1, 1, 1] self.input_size = [2, 3, 4, 4, 4] # NCDHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 3, 3, 3] @@ -205,7 +207,7 @@ class TestWith1x1(TestConv3dOp): self.stride = [1, 1, 1] self.input_size = [2, 3, 4, 4, 4] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 1, 1, 1] def init_dilation(self): @@ -221,7 +223,7 @@ class TestWithInput1x1Filter1x1(TestConv3dOp): self.stride = [1, 1, 1] self.input_size = [2, 3, 1, 1, 1] # NCHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 1, 1, 1] def init_dilation(self): @@ -237,7 +239,7 @@ class TestWithDilation(TestConv3dOp): self.stride = [1, 1, 1] self.input_size = [2, 3, 6, 6, 6] # NCDHW assert np.mod(self.input_size[1], self.groups) == 0 - f_c = self.input_size[1] / self.groups + f_c = self.input_size[1] // self.groups self.filter_size = [6, f_c, 2, 2, 2] def init_dilation(self): diff --git a/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py b/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py index 300fa5e8bd..8d9075961c 100644 --- a/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv3d_transpose_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -25,7 +27,7 @@ def conv3dtranspose_forward_naive(input_, filter_, attrs): groups = attrs['groups'] assert in_c == f_c out_c = f_out_c * groups - sub_in_c = in_c / groups + sub_in_c = in_c // groups stride, pad, dilations = attrs['strides'], attrs['paddings'], attrs[ 'dilations'] diff --git a/python/paddle/fluid/tests/unittests/test_conv_shift_op.py b/python/paddle/fluid/tests/unittests/test_conv_shift_op.py index 9fdb7baa90..b7364e869e 100644 --- a/python/paddle/fluid/tests/unittests/test_conv_shift_op.py +++ b/python/paddle/fluid/tests/unittests/test_conv_shift_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest @@ -21,7 +23,7 @@ def conv_shift_forward(x, y): out = np.zeros_like(x) M = x.shape[1] N = y.shape[1] - y_half_width = (N - 1) / 2 + y_half_width = (N - 1) // 2 for i in range(M): for j in range(N): out[:, i] += x[:, (i + j + M - y_half_width) % M] * y[:, j] diff --git a/python/paddle/fluid/tests/unittests/test_cos_sim_op.py b/python/paddle/fluid/tests/unittests/test_cos_sim_op.py index 1b27cd5767..3c3fd6d4d7 100644 --- a/python/paddle/fluid/tests/unittests/test_cos_sim_op.py +++ b/python/paddle/fluid/tests/unittests/test_cos_sim_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py b/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py index 07c89eefc3..fd34c8fc93 100644 --- a/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py +++ b/python/paddle/fluid/tests/unittests/test_create_op_doc_string.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.layers as layers diff --git a/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py b/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py index 122b076c2d..51bd1300e6 100644 --- a/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py +++ b/python/paddle/fluid/tests/unittests/test_crf_decoding_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import random import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_crop_op.py b/python/paddle/fluid/tests/unittests/test_crop_op.py index 4016089c01..d7bcfba8de 100644 --- a/python/paddle/fluid/tests/unittests/test_crop_op.py +++ b/python/paddle/fluid/tests/unittests/test_crop_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py index 86ac159323..fa367f95fc 100644 --- a/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py +++ b/python/paddle/fluid/tests/unittests/test_cross_entropy_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest, randomize_probability diff --git a/python/paddle/fluid/tests/unittests/test_ctc_align.py b/python/paddle/fluid/tests/unittests/test_ctc_align.py index 131b4076f4..5f17d2d407 100644 --- a/python/paddle/fluid/tests/unittests/test_ctc_align.py +++ b/python/paddle/fluid/tests/unittests/test_ctc_align.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import sys import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_cumsum_op.py b/python/paddle/fluid/tests/unittests/test_cumsum_op.py index 04e7f0b945..13a4eacece 100644 --- a/python/paddle/fluid/tests/unittests/test_cumsum_op.py +++ b/python/paddle/fluid/tests/unittests/test_cumsum_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_data_balance.py b/python/paddle/fluid/tests/unittests/test_data_balance.py index 951282e8ba..e39eedd282 100644 --- a/python/paddle/fluid/tests/unittests/test_data_balance.py +++ b/python/paddle/fluid/tests/unittests/test_data_balance.py @@ -12,9 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid -import paddle.v2 as paddle +import paddle import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_debugger.py b/python/paddle/fluid/tests/unittests/test_debugger.py index 870952f2f9..f4c9466d63 100644 --- a/python/paddle/fluid/tests/unittests/test_debugger.py +++ b/python/paddle/fluid/tests/unittests/test_debugger.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_decayed_adagrad_op.py b/python/paddle/fluid/tests/unittests/test_decayed_adagrad_op.py index 84c44d4817..a664a1529f 100644 --- a/python/paddle/fluid/tests/unittests/test_decayed_adagrad_op.py +++ b/python/paddle/fluid/tests/unittests/test_decayed_adagrad_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py b/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py index 868bcca881..01a7b68248 100644 --- a/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py +++ b/python/paddle/fluid/tests/unittests/test_default_scope_funcs.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from paddle.fluid.default_scope_funcs import * import unittest diff --git a/python/paddle/fluid/tests/unittests/test_desc_clone.py b/python/paddle/fluid/tests/unittests/test_desc_clone.py index 8603d3a5b3..fa6b679562 100644 --- a/python/paddle/fluid/tests/unittests/test_desc_clone.py +++ b/python/paddle/fluid/tests/unittests/test_desc_clone.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import argparse import time @@ -25,6 +27,7 @@ import unittest from multiprocessing import Process import os import signal +import six import collections SEED = 1 @@ -53,7 +56,8 @@ def cnn_model(data): # TODO(dzhwinter) : refine the initializer and random seed settting SIZE = 10 input_shape = conv_pool_2.shape - param_shape = [reduce(lambda a, b: a * b, input_shape[1:], 1)] + [SIZE] + param_shape = [six.moves.reduce(lambda a, b: a * b, input_shape[1:], 1) + ] + [SIZE] scale = (2.0 / (param_shape[0]**2 * SIZE))**0.5 predict = fluid.layers.fc( @@ -106,7 +110,7 @@ def get_transpiler(trainer_id, main_program, pserver_endpoints, trainers): def operator_equal(a, b): - for k, v in a.__dict__.iteritems(): + for k, v in six.iteritems(a.__dict__): if isinstance(v, fluid.framework.Program) or \ isinstance(v, fluid.framework.Block): continue @@ -116,8 +120,8 @@ def operator_equal(a, b): raise ValueError("In operator_equal not equal:{0}\n".format(k)) elif isinstance(v, collections.OrderedDict): - v0 = sorted(v.iteritems(), key=lambda x: x[0]) - v1 = sorted(b.__dict__[k].iteritems(), key=lambda x: x[0]) + v0 = sorted(six.iteritems(v), key=lambda x: x[0]) + v1 = sorted(six.iteritems(b.__dict__[k]), key=lambda x: x[0]) if v0 != v1: raise ValueError("In operator_equal not equal:{0}\n".format(k)) @@ -129,7 +133,7 @@ def operator_equal(a, b): def block_equal(a, b): - for k, v in a.__dict__.iteritems(): + for k, v in six.iteritems(a.__dict__): if isinstance(v, core.ProgramDesc) or isinstance( v, fluid.framework.Program) or isinstance(v, core.BlockDesc): continue @@ -141,8 +145,8 @@ def block_equal(a, b): assert (len(a.ops) == len(b.ops)) elif isinstance(v, collections.OrderedDict): - v0 = sorted(v.iteritems(), key=lambda x: x[0]) - v1 = sorted(b.__dict__[k].iteritems(), key=lambda x: x[0]) + v0 = sorted(six.iteritems(v), key=lambda x: x[0]) + v1 = sorted(six.iteritems(b.__dict__[k]), key=lambda x: x[0]) if v0 != v1: raise ValueError("In block_equal not equal:{0}\n".format(k)) @@ -154,7 +158,7 @@ def block_equal(a, b): def program_equal(a, b): - for k, v in a.__dict__.iteritems(): + for k, v in six.iteritems(a.__dict__): if isinstance(v, core.ProgramDesc): continue diff --git a/python/paddle/fluid/tests/unittests/test_detection_map_op.py b/python/paddle/fluid/tests/unittests/test_detection_map_op.py index 8b66d1b270..f6eb8f2c6d 100644 --- a/python/paddle/fluid/tests/unittests/test_detection_map_op.py +++ b/python/paddle/fluid/tests/unittests/test_detection_map_op.py @@ -12,8 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np +import six import sys import collections import math @@ -176,7 +179,7 @@ class TestDetectionMAPOp(OpTest): true_pos[label].append([score, tp]) false_pos[label].append([score, fp]) - for (label, label_pos_num) in list(label_count.items()): + for (label, label_pos_num) in six.iteritems(label_count): if label_pos_num == 0 or label not in true_pos: continue label_true_pos = true_pos[label] label_false_pos = false_pos[label] diff --git a/python/paddle/fluid/tests/unittests/test_dist_base.py b/python/paddle/fluid/tests/unittests/test_dist_base.py index 4379463aca..0e815c9144 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_base.py +++ b/python/paddle/fluid/tests/unittests/test_dist_base.py @@ -11,11 +11,14 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import time import unittest import os import sys +import six import signal import subprocess import six @@ -27,7 +30,7 @@ class TestDistRunnerBase(object): "get_model should be implemented by child classes.") def get_transpiler(self, trainer_id, main_program, pserver_endpoints, - trainers): + trainers, sync_mode): # NOTE: import fluid until runtime, or else forking processes will cause error. import paddle import paddle.fluid as fluid @@ -36,17 +39,22 @@ class TestDistRunnerBase(object): trainer_id=trainer_id, program=main_program, pservers=pserver_endpoints, - trainers=trainers) + trainers=trainers, + sync_mode=sync_mode) return t - def run_pserver(self, pserver_endpoints, trainers, current_endpoint, - trainer_id): + def run_pserver(self, + pserver_endpoints, + trainers, + current_endpoint, + trainer_id, + sync_mode=True): import paddle import paddle.fluid as fluid self.get_model(batch_size=2) t = self.get_transpiler(trainer_id, fluid.default_main_program(), pserver_endpoints, - trainers) + trainers, sync_mode) pserver_prog = t.get_pserver_program(current_endpoint) startup_prog = t.get_startup_program(current_endpoint, pserver_prog) place = fluid.CPUPlace() @@ -54,7 +62,13 @@ class TestDistRunnerBase(object): exe.run(startup_prog) exe.run(pserver_prog) - def run_trainer(self, place, endpoints, trainer_id, trainers, is_dist=True): + def run_trainer(self, + place, + endpoints, + trainer_id, + trainers, + is_dist=True, + sync_mode=True): import paddle import paddle.fluid as fluid test_program, avg_cost, train_reader, test_reader, batch_acc, predict = \ @@ -62,7 +76,7 @@ class TestDistRunnerBase(object): if is_dist: t = self.get_transpiler(trainer_id, fluid.default_main_program(), endpoints, - trainers) + trainers, sync_mode) trainer_prog = t.get_trainer_program() else: trainer_prog = fluid.default_main_program() @@ -103,9 +117,9 @@ def runtime_main(test_class): import paddle.fluid as fluid import paddle.fluid.core as core - if len(sys.argv) != 7: + if len(sys.argv) != 8: print( - "Usage: python dist_se_resnext.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist]" + "Usage: python dist_se_resnext.py [pserver/trainer] [endpoints] [trainer_id] [current_endpoint] [trainers] [is_dist] [sync_mode]" ) role = sys.argv[1] endpoints = sys.argv[2] @@ -113,31 +127,43 @@ def runtime_main(test_class): current_endpoint = sys.argv[4] trainers = int(sys.argv[5]) is_dist = True if sys.argv[6] == "TRUE" else False + sync_mode = True if sys.argv[7] == "TRUE" else False model = test_class() if role == "pserver": - model.run_pserver(endpoints, trainers, current_endpoint, trainer_id) + model.run_pserver(endpoints, trainers, current_endpoint, trainer_id, + sync_mode) else: p = fluid.CUDAPlace(0) if core.is_compiled_with_cuda( ) else fluid.CPUPlace() - model.run_trainer(p, endpoints, trainer_id, trainers, is_dist) + model.run_trainer(p, endpoints, trainer_id, trainers, is_dist, + sync_mode) + + +import paddle.compat as cpt class TestDistBase(unittest.TestCase): + def _setup_config(self): + raise NotImplementedError("tests should have _setup_config implemented") + def setUp(self): self._trainers = 2 self._pservers = 2 self._ps_endpoints = "127.0.0.1:9123,127.0.0.1:9124" self._python_interp = "python" + self._sync_mode = True + self._setup_config() def start_pserver(self, model_file, check_error_log): + sync_mode_str = "TRUE" if self._sync_mode else "FALSE" ps0_ep, ps1_ep = self._ps_endpoints.split(",") - ps0_cmd = "%s %s pserver %s 0 %s %d TRUE" % \ + ps0_cmd = "%s %s pserver %s 0 %s %d TRUE %s" % \ (self._python_interp, model_file, self._ps_endpoints, ps0_ep, - self._trainers) - ps1_cmd = "%s %s pserver %s 0 %s %d TRUE" % \ + self._trainers, sync_mode_str) + ps1_cmd = "%s %s pserver %s 0 %s %d TRUE %s" % \ (self._python_interp, model_file, self._ps_endpoints, ps1_ep, - self._trainers) + self._trainers, sync_mode_str) ps0_pipe = subprocess.PIPE ps1_pipe = subprocess.PIPE @@ -189,9 +215,10 @@ class TestDistBase(unittest.TestCase): # Run local to get a base line env_local = {"CUDA_VISIBLE_DEVICES": "0"} env_local.update(required_envs) - local_cmd = "%s %s trainer %s 0 %s %d FLASE" % \ + sync_mode_str = "TRUE" if self._sync_mode else "FALSE" + local_cmd = "%s %s trainer %s 0 %s %d FLASE %s" % \ (self._python_interp, model_file, - "127.0.0.1:1234", "127.0.0.1:1234", 1) + "127.0.0.1:1234", "127.0.0.1:1234", 1, sync_mode_str) if not check_error_log: local_proc = subprocess.Popen( local_cmd.split(" "), @@ -209,7 +236,7 @@ class TestDistBase(unittest.TestCase): local_proc.wait() out, err = local_proc.communicate() - local_ret = out + local_ret = cpt.to_text(out) sys.stderr.write('local_loss: %s\n' % local_ret) sys.stderr.write('local_stderr: %s\n' % err) @@ -220,12 +247,12 @@ class TestDistBase(unittest.TestCase): self._wait_ps_ready(ps1.pid) ps0_ep, ps1_ep = self._ps_endpoints.split(",") - tr0_cmd = "%s %s trainer %s 0 %s %d TRUE" % \ + tr0_cmd = "%s %s trainer %s 0 %s %d TRUE %s" % \ (self._python_interp, model_file, self._ps_endpoints, ps0_ep, - self._trainers) - tr1_cmd = "%s %s trainer %s 1 %s %d TRUE" % \ + self._trainers, sync_mode_str) + tr1_cmd = "%s %s trainer %s 1 %s %d TRUE %s" % \ (self._python_interp, model_file, self._ps_endpoints, ps1_ep, - self._trainers) + self._trainers, sync_mode_str) env0 = {"CUDA_VISIBLE_DEVICES": "0"} env1 = {"CUDA_VISIBLE_DEVICES": "1"} @@ -256,7 +283,7 @@ class TestDistBase(unittest.TestCase): tr1_proc.wait() out, err = tr0_proc.communicate() sys.stderr.write('dist_stderr: %s\n' % err) - loss_data0 = out + loss_data0 = cpt.to_text(out) sys.stderr.write('dist_loss: %s\n' % loss_data0) lines = loss_data0.split("\n") dist_first_loss = eval(lines[0].replace(" ", ","))[0] diff --git a/python/paddle/fluid/tests/unittests/test_dist_mnist.py b/python/paddle/fluid/tests/unittests/test_dist_mnist.py index b3ccec9a7d..36bab6f046 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_dist_mnist.py @@ -11,14 +11,27 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest from test_dist_base import TestDistBase -class TestDistSeResneXt2x2(TestDistBase): +class TestDistMnist2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + def test_se_resnext(self): self.check_with_place("dist_mnist.py", delta=1e-7) +class TestDistMnistAsync(TestDistBase): + def _setup_config(self): + self._sync_mode = False + + def test_se_resnext(self): + self.check_with_place("dist_mnist.py", delta=200) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py index a33a338fc1..c0e9fa38e7 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py +++ b/python/paddle/fluid/tests/unittests/test_dist_se_resnext.py @@ -11,14 +11,27 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest from test_dist_base import TestDistBase class TestDistSeResneXt2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + def test_se_resnext(self): self.check_with_place("dist_se_resnext.py", delta=1e-7) +class TestDistSeResneXt2x2Async(TestDistBase): + def _setup_config(self): + self._sync_mode = False + + def test_se_resnext(self): + self.check_with_place("dist_se_resnext.py", delta=100) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_train.py b/python/paddle/fluid/tests/unittests/test_dist_train.py index 55aa923f5a..9581abdf39 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_train.py +++ b/python/paddle/fluid/tests/unittests/test_dist_train.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import os import time import unittest diff --git a/python/paddle/fluid/tests/unittests/test_dist_transformer.py b/python/paddle/fluid/tests/unittests/test_dist_transformer.py index 68cd35d751..62fcf5953f 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transformer.py @@ -12,11 +12,16 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from test_dist_base import TestDistBase class TestDistTransformer2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + def test_transformer(self): # TODO(paddle-dev): check if the delta is OK. # Usually start around ~8000 and converge to ~5000 diff --git a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py index 124abf4ccd..9f04d290f7 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_transpiler.py +++ b/python/paddle/fluid/tests/unittests/test_dist_transpiler.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import math import unittest @@ -45,10 +47,10 @@ class TranspilerTest(unittest.TestCase): avg_cost = fluid.layers.mean(cost) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) sgd_optimizer.minimize(avg_cost) - return def get_main_program(self): main = fluid.Program() + main.random_seed = 1 with fluid.program_guard(main): self.net_conf() self.origin_prog = main.clone() @@ -92,8 +94,9 @@ class TranspilerTest(unittest.TestCase): def test_transpiler(self): main = fluid.Program() startup = fluid.Program() - with fluid.program_guard(main, startup): - self.transpiler_test_impl() + with fluid.unique_name.guard(): + with fluid.program_guard(main, startup): + self.transpiler_test_impl() class TestBasicModel(TranspilerTest): @@ -246,7 +249,6 @@ class TestLRDecay(TranspilerTest): decay_rate=0.1, staircase=True)) sgd_optimizer.minimize(avg_cost) - return def transpiler_test_impl(self): pserver, startup = self.get_pserver(self.pserver1_ep) @@ -276,7 +278,6 @@ class TestLRDecayConditional(TranspilerTest): learning_rate=fluid.layers.piecewise_decay([10000, 20000], [1.0, 0.5, 1.0])) sgd_optimizer.minimize(avg_cost) - return def transpiler_test_impl(self): pserver, startup = self.get_pserver(self.pserver1_ep) @@ -325,7 +326,6 @@ class TestL2Decay(TranspilerTest): avg_cost = fluid.layers.mean(cost) sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1) sgd_optimizer.minimize(avg_cost) - return def transpiler_test_impl(self): pserver, startup = self.get_pserver(self.pserver1_ep) @@ -360,7 +360,6 @@ class TestL2DecayWithPiecewise(TranspilerTest): momentum=0.9, regularization=fluid.regularizer.L2Decay(1e-4)) sgd_optimizer.minimize(avg_cost) - return def transpiler_test_impl(self): pserver, startup = self.get_pserver(self.pserver1_ep) @@ -390,13 +389,14 @@ class TestDistLookupTableBase(TranspilerTest): def network_with_table(self, is_sparse, is_distributed): self.table_size = 1000 self.emb_size = 64 + self.lookup_table_name = 'shared_w' def emb_pool(ids): emb = fluid.layers.embedding( input=ids, size=[self.table_size, self.emb_size], dtype='float32', - param_attr='shared_w', # share parameter + param_attr=self.lookup_table_name, # share parameter is_sparse=is_sparse, is_distributed=is_distributed) pool = fluid.layers.sequence_pool(input=emb, pool_type='average') @@ -569,7 +569,7 @@ class TestDistLookupTableSliceSize(TestDistLookupTableBase): def transpiler_test_impl(self): config = fluid.DistributeTranspilerConfig() - pserver1, startup1 = self.get_pserver(self.pserver1_ep, config) + pserver1, _ = self.get_pserver(self.pserver1_ep, config) self.assertTrue(self.transpiler.has_distributed_lookup_table) lookup_table_var = pserver1.global_block().vars[ @@ -579,6 +579,21 @@ class TestDistLookupTableSliceSize(TestDistLookupTableBase): self.assertEqual(row_size, calc_row_size) +class TestDistArgsInProgram(TestDistLookupTableBase): + def net_conf(self): + self.network_with_table(is_sparse=True, is_distributed=True) + + def transpiler_test_impl(self): + trainer, _ = self.get_trainer() + + self.assertTrue(trainer._is_distributed) + self.assertTrue(trainer._is_chief) + self.assertEqual(trainer._distributed_lookup_table, + self.lookup_table_name) + self.assertEqual(trainer._endpoints, + [self.pserver1_ep, self.pserver2_ep]) + + class TestRMSPropOptimizer(TranspilerTest): def net_conf(self): x = fluid.layers.data(name='x', shape=[1000], dtype='float32') @@ -592,7 +607,6 @@ class TestRMSPropOptimizer(TranspilerTest): avg_cost = fluid.layers.mean(cost) optimizer = fluid.optimizer.RMSProp(learning_rate=0.1) optimizer.minimize(avg_cost) - return def transpiler_test_impl(self): pserver, startup = self.get_pserver(self.pserver1_ep) @@ -609,5 +623,40 @@ class TestRMSPropOptimizer(TranspilerTest): self.assertEqual(moment_var.shape, (500, 1000)) +class TestLoadSliceVar(TranspilerTest): + def net_conf(self): + x = fluid.layers.data(name='x', shape=[1000], dtype='float32') + y_predict = fluid.layers.fc(input=x, + size=1000, + act=None, + param_attr=fluid.ParamAttr(name='fc_w'), + bias_attr=fluid.ParamAttr(name='fc_b')) + y = fluid.layers.data(name='y', shape=[1], dtype='float32') + cost = fluid.layers.square_error_cost(input=y_predict, label=y) + avg_cost = fluid.layers.mean(cost) + optimizer = fluid.optimizer.RMSProp(learning_rate=0.1) + optimizer.minimize(avg_cost) + + def transpiler_test_impl(self): + pserver, _ = self.get_pserver(self.pserver1_ep) + pserver2, _ = self.get_pserver(self.pserver2_ep) + + self.assertTrue(pserver._slice_vars_and_attrs) + self.assertTrue(pserver2._slice_vars_and_attrs) + + for idx in xrange(len(pserver._slice_vars_and_attrs)): + self.assertEqual(pserver._slice_vars_and_attrs[idx][0], + pserver2._slice_vars_and_attrs[idx][0]) + + total_numel = reduce(lambda x, y: x * y, + pserver._slice_vars_and_attrs[idx][0].shape) + self.assertEqual( + total_numel, + reduce(lambda x, y: x * y, + pserver._slice_vars_and_attrs[idx][2].shape) + reduce( + lambda x, y: x * y, + pserver2._slice_vars_and_attrs[idx][2].shape)) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dist_word2vec.py b/python/paddle/fluid/tests/unittests/test_dist_word2vec.py index 543d0f9dc2..38af149ad3 100644 --- a/python/paddle/fluid/tests/unittests/test_dist_word2vec.py +++ b/python/paddle/fluid/tests/unittests/test_dist_word2vec.py @@ -11,14 +11,27 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest from test_dist_base import TestDistBase class TestDistSeResneXt2x2(TestDistBase): + def _setup_config(self): + self._sync_mode = True + def test_se_resnext(self): self.check_with_place("dist_word2vec.py", delta=1e-7) +class TestDistSeResneXt2x2Async(TestDistBase): + def _setup_config(self): + self._sync_mode = False + + def test_se_resnext(self): + self.check_with_place("dist_word2vec.py", delta=1) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_dropout_op.py b/python/paddle/fluid/tests/unittests/test_dropout_op.py index eaa3435a86..0296bc2af4 100644 --- a/python/paddle/fluid/tests/unittests/test_dropout_op.py +++ b/python/paddle/fluid/tests/unittests/test_dropout_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py index fdc6adc93b..d84dab1499 100644 --- a/python/paddle/fluid/tests/unittests/test_dyn_rnn.py +++ b/python/paddle/fluid/tests/unittests/test_dyn_rnn.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import paddle import unittest diff --git a/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py b/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py index 7756885166..9d635f36fe 100644 --- a/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py +++ b/python/paddle/fluid/tests/unittests/test_dynrnn_gradient_check.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy import random import collections diff --git a/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py b/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py index d182889a97..b4359fc69a 100644 --- a/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py +++ b/python/paddle/fluid/tests/unittests/test_dynrnn_static_input.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_edit_distance_op.py b/python/paddle/fluid/tests/unittests/test_edit_distance_op.py index 816562621b..4d03523025 100644 --- a/python/paddle/fluid/tests/unittests/test_edit_distance_op.py +++ b/python/paddle/fluid/tests/unittests/test_edit_distance_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py index bcdbfc8e52..d85cc1f856 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_mkldnn_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py index fb9a496126..5aec5d8e38 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_add_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_div_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_div_op.py index bfe022af6d..cadaf1df53 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_div_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_div_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py index 6f35004489..9f452ffde7 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_gradient_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_max_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_max_op.py index b6cd18a579..43c58710ba 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_max_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_max_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_min_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_min_op.py index 92099724fe..45c861e2c3 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_min_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_min_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py index 2742bb21d9..775c2253ab 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_pow_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_pow_op.py index a3fd18669c..7bf642f03f 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_pow_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_pow_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py index 1854232194..6cb88a8bb1 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_sub_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_exception.py b/python/paddle/fluid/tests/unittests/test_exception.py index bb7c0f88f6..798ed53cdd 100644 --- a/python/paddle/fluid/tests/unittests/test_exception.py +++ b/python/paddle/fluid/tests/unittests/test_exception.py @@ -12,19 +12,23 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + +import paddle.compat as cpt import paddle.fluid.core as core import unittest class TestException(unittest.TestCase): def test_exception(self): - ex = None + exception = None try: core.__unittest_throw_exception__() except core.EnforceNotMet as ex: - self.assertIn("test exception", ex.message) + self.assertIn("test exception", cpt.get_exception_message(ex)) + exception = ex - self.assertIsNotNone(ex) + self.assertIsNotNone(exception) if __name__ == "__main__": diff --git a/python/paddle/fluid/tests/unittests/test_executor_and_mul.py b/python/paddle/fluid/tests/unittests/test_executor_and_mul.py index e1272c1d6d..b1f89eca6e 100644 --- a/python/paddle/fluid/tests/unittests/test_executor_and_mul.py +++ b/python/paddle/fluid/tests/unittests/test_executor_and_mul.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy diff --git a/python/paddle/fluid/tests/unittests/test_expand_op.py b/python/paddle/fluid/tests/unittests/test_expand_op.py index a91e3aef5a..67a8d8f072 100644 --- a/python/paddle/fluid/tests/unittests/test_expand_op.py +++ b/python/paddle/fluid/tests/unittests/test_expand_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_extract_rows_op.py b/python/paddle/fluid/tests/unittests/test_extract_rows_op.py index 6a41c44fe6..8629bcf0f2 100644 --- a/python/paddle/fluid/tests/unittests/test_extract_rows_op.py +++ b/python/paddle/fluid/tests/unittests/test_extract_rows_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py b/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py index 026ac2112b..d84ebed3fa 100644 --- a/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py +++ b/python/paddle/fluid/tests/unittests/test_fake_dequantize_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import math diff --git a/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py b/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py index 6c6aa9d3bb..cc0494774a 100644 --- a/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py +++ b/python/paddle/fluid/tests/unittests/test_fake_quantize_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py index 099e6e6064..45951a34d6 100644 --- a/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_fc_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_fc_op.py b/python/paddle/fluid/tests/unittests/test_fc_op.py index 2bb920710a..ff417ad2f1 100644 --- a/python/paddle/fluid/tests/unittests/test_fc_op.py +++ b/python/paddle/fluid/tests/unittests/test_fc_op.py @@ -64,27 +64,47 @@ class TestFCOp(OpTest): self.check_output() -class TestFCOpBiasBoth(TestFCOp): +class TestFCOpNoBias(TestFCOp): def init_shapes(self, mb, ic, oc, h, w): - for with_bias in {True, False}: - self.with_bias = with_bias - self.matrix = MatrixGenerate(mb, ic, oc, h, w) + self.with_bias = False + self.matrix = MatrixGenerate(mb, ic, oc, h, w) -class TestFCOp1(TestFCOpBiasBoth): +class TestFCOpWithBias(TestFCOp): + def init_shapes(self, mb, ic, oc, h, w): + self.with_bias = True + self.matrix = MatrixGenerate(mb, ic, oc, h, w) + + +class TestFCOp1(TestFCOpNoBias): def init_op_type(self): self.init_shapes(2, 8, 10, 1, 1) -class TestFCOp2(TestFCOpBiasBoth): +class TestFCOp2(TestFCOpNoBias): def init_op_type(self): self.init_shapes(4, 5, 6, 2, 2) -class TestFCOp4(TestFCOpBiasBoth): +class TestFCOp4(TestFCOpNoBias): def init_op_type(self): self.init_shapes(1, 32, 64, 3, 3) +class TestFCOpWithBias1(TestFCOpWithBias): + def init_op_type(self): + self.init_shapes(3, 8, 10, 2, 1) + + +class TestFCOpWithBias2(TestFCOpWithBias): + def init_op_type(self): + self.init_shapes(4, 5, 6, 2, 2) + + +class TestFCOpWithBias3(TestFCOpWithBias): + def init_op_type(self): + self.init_shapes(1, 64, 32, 3, 3) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py b/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py index 8b9da84311..b823d397e9 100644 --- a/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py +++ b/python/paddle/fluid/tests/unittests/test_feed_fetch_method.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid.core as core import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_fetch_var.py b/python/paddle/fluid/tests/unittests/test_fetch_var.py index e6f37f0b4c..de339d821b 100644 --- a/python/paddle/fluid/tests/unittests/test_fetch_var.py +++ b/python/paddle/fluid/tests/unittests/test_fetch_var.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import paddle.fluid.layers as layers import op_test diff --git a/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py b/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py index 0c75cf33f5..fdc8a118e5 100644 --- a/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py +++ b/python/paddle/fluid/tests/unittests/test_fill_constant_batch_size_like_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_fill_constant_op.py b/python/paddle/fluid/tests/unittests/test_fill_constant_op.py index 5e2ddb218a..44fb1d047d 100644 --- a/python/paddle/fluid/tests/unittests/test_fill_constant_op.py +++ b/python/paddle/fluid/tests/unittests/test_fill_constant_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_fill_op.py b/python/paddle/fluid/tests/unittests/test_fill_op.py index 762d29199e..b734ee05b3 100644 --- a/python/paddle/fluid/tests/unittests/test_fill_op.py +++ b/python/paddle/fluid/tests/unittests/test_fill_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_fill_zeros_like_op.py b/python/paddle/fluid/tests/unittests/test_fill_zeros_like_op.py index c9b3e4ba13..eec73d0beb 100644 --- a/python/paddle/fluid/tests/unittests/test_fill_zeros_like_op.py +++ b/python/paddle/fluid/tests/unittests/test_fill_zeros_like_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_flatten_op.py b/python/paddle/fluid/tests/unittests/test_flatten_op.py index f8692ce2ea..17b01e0312 100644 --- a/python/paddle/fluid/tests/unittests/test_flatten_op.py +++ b/python/paddle/fluid/tests/unittests/test_flatten_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_framework_debug_str.py b/python/paddle/fluid/tests/unittests/test_framework_debug_str.py index c906c74afe..72f43e56cc 100644 --- a/python/paddle/fluid/tests/unittests/test_framework_debug_str.py +++ b/python/paddle/fluid/tests/unittests/test_framework_debug_str.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from paddle.fluid.framework import Program diff --git a/python/paddle/fluid/tests/unittests/test_ftrl_op.py b/python/paddle/fluid/tests/unittests/test_ftrl_op.py index 5f7581391a..a6390b054f 100644 --- a/python/paddle/fluid/tests/unittests/test_ftrl_op.py +++ b/python/paddle/fluid/tests/unittests/test_ftrl_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py b/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py index ec0a939e9e..97e1b9061a 100644 --- a/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py +++ b/python/paddle/fluid/tests/unittests/test_fused_elemwise_activation_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_fusion_lstm_op.py b/python/paddle/fluid/tests/unittests/test_fusion_lstm_op.py new file mode 100644 index 0000000000..9d8bef677f --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_fusion_lstm_op.py @@ -0,0 +1,151 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import print_function + +import unittest +import numpy as np +from op_test import OpTest +from test_lstm_op import lstm, ACTIVATION + + +def fc(x, w, b): + return np.dot(x, w) + b + + +def fusion_lstm( + x, # T x M + lod, # 1 x N + wx=None, # M x 4D + bx=None, # 1 x 4D + h0=None, # N x D + c0=None, # N x D + w_h=None, # D x 4D + w_b=None, # 1 x 4D + w_c=None, # 1 x 3D + is_reverse=False, + act_gate=None, + act_cell=None, + act_cand=None): + return lstm( + fc(x, wx, bx), lod, h0, c0, w_h, w_b, w_c, is_reverse, act_gate, + act_cell, act_cand) + + +class TestLstmOp(OpTest): + def set_argument(self): + self.lod = [[2, 3, 2]] + + def setUp(self): + self.op_type = 'fusion_lstm' + self.lod = [[2, 3, 2]] + self.M = 8 + self.D = 16 + self.has_initial_state = False + self.is_reverse = False + self.act_gate = 'sigmoid' + self.act_cell = 'tanh' + self.act_cand = 'tanh' + self.use_peepholes = False + self.set_argument() + + T = sum(self.lod[0]) + bs = len(self.lod[0]) + + x = np.random.normal(size=(T, self.M)).astype('float64') + if self.has_initial_state: + h0 = np.random.normal(size=(bs, self.D)).astype('float64') + c0 = np.random.normal(size=(bs, self.D)).astype('float64') + else: + h0 = np.zeros((bs, self.D)).astype('float64') + c0 = np.zeros((bs, self.D)).astype('float64') + + wh = np.random.normal(size=(self.D, 4 * self.D)).astype('float64') + + if self.use_peepholes: + b = np.random.normal(size=(1, 7 * self.D)).astype('float64') + else: + b = np.random.normal(size=(1, 4 * self.D)).astype('float64') + w_b = np.copy(b[:, 0:4 * self.D]) + w_c = b[:, 4 * self.D:] if self.use_peepholes else None + + # this is the weight of fc + wx = np.random.normal(size=(self.M, 4 * self.D)).astype('float64') + # this is the bias of fc + # and it should be manually added into the bias of this fusion LSTM + bx = np.random.normal(size=(1, 4 * self.D)).astype('float64') + b[0, 0:4 * self.D] += bx[0, :] + h, c = fusion_lstm(x, self.lod, wx, bx, h0, c0, wh, w_b, w_c, + self.is_reverse, ACTIVATION[self.act_gate], + ACTIVATION[self.act_cell], ACTIVATION[self.act_cand]) + + self.inputs = { + 'X': (x, self.lod), + 'WeightX': wx, + 'WeightH': wh, + 'Bias': b + } + + if self.has_initial_state: + self.inputs['H0'] = h0 + self.inputs['C0'] = c0 + + self.outputs = { + 'Hidden': (h, self.lod), + 'Cell': (c, self.lod), + } + self.attrs = { + 'use_peepholes': self.use_peepholes, + 'is_reverse': self.is_reverse, + 'gate_activation': self.act_gate, + 'cell_activation': self.act_cell, + 'candidate_activation': self.act_cand + } + + def test_check_output(self): + self.check_output(atol=1e-8) + + +class TestLstmOpInitReverse(TestLstmOp): + def set_argument(self): + self.has_initial_state = True + self.is_reverse = True + + +class TestLstmOpMD1(TestLstmOp): + def set_argument(self): + self.M = 36 + self.D = 8 + + +class TestLstmOpMD2(TestLstmOp): + def set_argument(self): + self.M = 8 + self.D = 8 + + +class TestLstmOpMD3(TestLstmOp): + def set_argument(self): + self.M = 15 + self.D = 3 + + +class TestLstmOpBS1(TestLstmOp): + def set_argument(self): + self.lod = [[3]] + self.D = 16 + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_gather_op.py b/python/paddle/fluid/tests/unittests/test_gather_op.py index 4ae9086480..bd5785aa55 100644 --- a/python/paddle/fluid/tests/unittests/test_gather_op.py +++ b/python/paddle/fluid/tests/unittests/test_gather_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_batch_size_like_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_batch_size_like_op.py index 1398166a74..9a0631fa26 100644 --- a/python/paddle/fluid/tests/unittests/test_gaussian_random_batch_size_like_op.py +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_batch_size_like_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py index 3ae877a608..9777ec3906 100644 --- a/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from test_gaussian_random_op import TestGaussianRandomOp diff --git a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py index 8481500fd7..496aa41110 100644 --- a/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py +++ b/python/paddle/fluid/tests/unittests/test_gaussian_random_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy diff --git a/python/paddle/fluid/tests/unittests/test_get_places_op.py b/python/paddle/fluid/tests/unittests/test_get_places_op.py index 964423e2d2..441666a97b 100644 --- a/python/paddle/fluid/tests/unittests/test_get_places_op.py +++ b/python/paddle/fluid/tests/unittests/test_get_places_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid from paddle.fluid.layers.device import get_places import decorators diff --git a/python/paddle/fluid/tests/unittests/test_gru_op.py b/python/paddle/fluid/tests/unittests/test_gru_op.py index 86a2c674d0..001fd7efb1 100644 --- a/python/paddle/fluid/tests/unittests/test_gru_op.py +++ b/python/paddle/fluid/tests/unittests/test_gru_op.py @@ -12,9 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import math +import functools from op_test import OpTest from test_lstm_op import identity, sigmoid, tanh, relu @@ -38,7 +41,8 @@ class TestGRUOp(OpTest): for i in range(len(seq_lens)): seq_starts.append(seq_starts[-1] + seq_lens[i]) sorted_seqs = sorted( - list(range(len(seq_lens))), lambda x, y: seq_lens[y] - seq_lens[x]) + list(range(len(seq_lens))), + key=functools.cmp_to_key(lambda x, y: seq_lens[y] - seq_lens[x])) num_batch = seq_lens[sorted_seqs[0]] for batch_idx in range(num_batch): idx_in_seq = [] diff --git a/python/paddle/fluid/tests/unittests/test_gru_unit_op.py b/python/paddle/fluid/tests/unittests/test_gru_unit_op.py index 87a9eba4d9..b5a66fdf08 100644 --- a/python/paddle/fluid/tests/unittests/test_gru_unit_op.py +++ b/python/paddle/fluid/tests/unittests/test_gru_unit_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import math import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_hinge_loss_op.py b/python/paddle/fluid/tests/unittests/test_hinge_loss_op.py index 70586c6be3..1eb441e2c5 100644 --- a/python/paddle/fluid/tests/unittests/test_hinge_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_hinge_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py index daa5da8d95..6948ae3002 100644 --- a/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py +++ b/python/paddle/fluid/tests/unittests/test_hsigmoid_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import math diff --git a/python/paddle/fluid/tests/unittests/test_huber_loss_op.py b/python/paddle/fluid/tests/unittests/test_huber_loss_op.py index a8d0a77625..0055ef0052 100644 --- a/python/paddle/fluid/tests/unittests/test_huber_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_huber_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_im2sequence_op.py b/python/paddle/fluid/tests/unittests/test_im2sequence_op.py index 13bc576874..833e46483c 100644 --- a/python/paddle/fluid/tests/unittests/test_im2sequence_op.py +++ b/python/paddle/fluid/tests/unittests/test_im2sequence_op.py @@ -11,6 +11,8 @@ #WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. #See the License for the specific language governing permissions and #limitations under the License. + +from __future__ import print_function import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_image_classification_layer.py b/python/paddle/fluid/tests/unittests/test_image_classification_layer.py index 23b1ed957a..405637969a 100644 --- a/python/paddle/fluid/tests/unittests/test_image_classification_layer.py +++ b/python/paddle/fluid/tests/unittests/test_image_classification_layer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_infer_shape.py b/python/paddle/fluid/tests/unittests/test_infer_shape.py index 699a2d4246..a3d700aad8 100644 --- a/python/paddle/fluid/tests/unittests/test_infer_shape.py +++ b/python/paddle/fluid/tests/unittests/test_infer_shape.py @@ -12,8 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest +import six import paddle.fluid.core as core @@ -27,14 +30,14 @@ class TestInferShape(unittest.TestCase): shape = [10, 20] # prepare input/output - x1 = block.var("x1") + x1 = block.var(six.b("x1")) x1.set_type(core.VarDesc.VarType.LOD_TENSOR) x1.set_shape(shape) - x2 = block.var("x2") + x2 = block.var(six.b("x2")) x2.set_type(core.VarDesc.VarType.LOD_TENSOR) x2.set_shape(shape) - out = block.var("out") + out = block.var(six.b("out")) out.set_type(core.VarDesc.VarType.LOD_TENSOR) # prepare the operator @@ -57,14 +60,14 @@ class TestInferShape(unittest.TestCase): y_shape = [20, 30] # prepare input/output - x1 = block.var("x") + x1 = block.var(six.b("x")) x1.set_type(core.VarDesc.VarType.LOD_TENSOR) x1.set_shape(x_shape) - x2 = block.var("y") + x2 = block.var(six.b("y")) x2.set_type(core.VarDesc.VarType.LOD_TENSOR) x2.set_shape(y_shape) - out = block.var("out") + out = block.var(six.b("out")) out.set_type(core.VarDesc.VarType.LOD_TENSOR) # prepare the operator diff --git a/python/paddle/fluid/tests/unittests/test_inference_model_io.py b/python/paddle/fluid/tests/unittests/test_inference_model_io.py index 4cd203155f..9962702f69 100644 --- a/python/paddle/fluid/tests/unittests/test_inference_model_io.py +++ b/python/paddle/fluid/tests/unittests/test_inference_model_io.py @@ -12,8 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest +import six import numpy as np import paddle.fluid.core as core @@ -48,7 +51,7 @@ class TestBook(unittest.TestCase): exe.run(init_program, feed={}, fetch_list=[]) - for i in range(100): + for i in six.moves.xrange(100): tensor_x = np.array( [[1, 1], [1, 2], [3, 4], [5, 2]]).astype("float32") tensor_y = np.array([[-2], [-3], [-7], [-7]]).astype("float32") @@ -64,7 +67,7 @@ class TestBook(unittest.TestCase): 'y': tensor_y}, fetch_list=[avg_cost])[0] - reload(executor) # reload to build a new scope + six.moves.reload_module(executor) # reload to build a new scope exe = executor.Executor(place) [infer_prog, feed_var_names, fetch_vars] = load_inference_model( diff --git a/python/paddle/fluid/tests/unittests/test_initializer.py b/python/paddle/fluid/tests/unittests/test_initializer.py index b215e37986..ab7183f88d 100644 --- a/python/paddle/fluid/tests/unittests/test_initializer.py +++ b/python/paddle/fluid/tests/unittests/test_initializer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import unittest diff --git a/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py b/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py index eff4212d91..7c1808cf99 100644 --- a/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py +++ b/python/paddle/fluid/tests/unittests/test_iou_similarity_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import numpy.random as random diff --git a/python/paddle/fluid/tests/unittests/test_is_empty_op.py b/python/paddle/fluid/tests/unittests/test_is_empty_op.py index 11121d9b65..26d607718a 100644 --- a/python/paddle/fluid/tests/unittests/test_is_empty_op.py +++ b/python/paddle/fluid/tests/unittests/test_is_empty_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_l1_norm_op.py b/python/paddle/fluid/tests/unittests/test_l1_norm_op.py index fa5b18a16f..4e24a78ee5 100644 --- a/python/paddle/fluid/tests/unittests/test_l1_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_l1_norm_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import unittest from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_label_smooth_op.py b/python/paddle/fluid/tests/unittests/test_label_smooth_op.py index ca21289a0d..62d385bc52 100644 --- a/python/paddle/fluid/tests/unittests/test_label_smooth_op.py +++ b/python/paddle/fluid/tests/unittests/test_label_smooth_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_layer_norm_op.py b/python/paddle/fluid/tests/unittests/test_layer_norm_op.py index 295887ccd1..fb6c43136f 100644 --- a/python/paddle/fluid/tests/unittests/test_layer_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_layer_norm_op.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_layers.py b/python/paddle/fluid/tests/unittests/test_layers.py index 07fd0575d3..8e707c8b00 100644 --- a/python/paddle/fluid/tests/unittests/test_layers.py +++ b/python/paddle/fluid/tests/unittests/test_layers.py @@ -159,7 +159,7 @@ class TestBook(unittest.TestCase): input=crf_decode, label=label, chunk_scheme="IOB", - num_chunk_types=(label_dict_len - 1) / 2) + num_chunk_types=(label_dict_len - 1) // 2) self.assertFalse(crf is None) self.assertFalse(crf_decode is None) @@ -286,7 +286,7 @@ class TestBook(unittest.TestCase): name='word_{0}'.format(i), shape=[1], dtype='int64')) dict_size = 10000 - label_word = int(window_size / 2) + 1 + label_word = int(window_size // 2) + 1 embs = [] for i in range(window_size): @@ -347,6 +347,25 @@ class TestBook(unittest.TestCase): self.assertIsNotNone(loss) print(str(program)) + def test_scatter(self): + program = Program() + with program_guard(program): + x = layers.data( + name='x', + shape=[3, 3], + append_batch_size=False, + dtype='float32') + idx = layers.data( + name='idx', shape=[2], append_batch_size=False, dtype='int32') + updates = layers.data( + name='updates', + shape=[2, 3], + append_batch_size=False, + dtype='float32') + out = layers.scatter(input=x, index=idx, updates=updates) + self.assertIsNotNone(out) + print(str(program)) + def test_lod_reset(self): program = Program() with program_guard(program): diff --git a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py index e628195e72..0d3e6d73e0 100644 --- a/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py +++ b/python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import copy import math import unittest diff --git a/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py b/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py index 696d0ab4fa..6e31e9204e 100644 --- a/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py +++ b/python/paddle/fluid/tests/unittests/test_linear_chain_crf_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import random import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py index 1cdc695010..48b52a5412 100644 --- a/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py +++ b/python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle import paddle.fluid as fluid import os diff --git a/python/paddle/fluid/tests/unittests/test_lod_array_length_op.py b/python/paddle/fluid/tests/unittests/test_lod_array_length_op.py index d8b4e40662..15485df5ac 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_array_length_op.py +++ b/python/paddle/fluid/tests/unittests/test_lod_array_length_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.layers as layers from paddle.fluid.executor import Executor diff --git a/python/paddle/fluid/tests/unittests/test_lod_rank_table.py b/python/paddle/fluid/tests/unittests/test_lod_rank_table.py index d53ead381d..865ca118d5 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_rank_table.py +++ b/python/paddle/fluid/tests/unittests/test_lod_rank_table.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from paddle.fluid.layers import data from paddle.fluid.layers.control_flow import lod_rank_table from paddle.fluid.executor import Executor diff --git a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py index 77905c4b96..31f364a42f 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_reset_op.py +++ b/python/paddle/fluid/tests/unittests/test_lod_reset_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py b/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py index 0ac6d9b81d..6ad27de9a0 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py +++ b/python/paddle/fluid/tests/unittests/test_lod_tensor_array.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core import numpy diff --git a/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py b/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py index 9789ff4af6..6a78ef5078 100644 --- a/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py +++ b/python/paddle/fluid/tests/unittests/test_lod_tensor_array_ops.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core import numpy diff --git a/python/paddle/fluid/tests/unittests/test_log_loss_op.py b/python/paddle/fluid/tests/unittests/test_log_loss_op.py index d3980b8db9..784f4f648d 100644 --- a/python/paddle/fluid/tests/unittests/test_log_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_log_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_logical_op.py b/python/paddle/fluid/tests/unittests/test_logical_op.py index 1d7dfe60f2..521851a3d5 100644 --- a/python/paddle/fluid/tests/unittests/test_logical_op.py +++ b/python/paddle/fluid/tests/unittests/test_logical_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import op_test import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py b/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py index aa9eae1e88..11e5d8b536 100644 --- a/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py +++ b/python/paddle/fluid/tests/unittests/test_lookup_sparse_table_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest @@ -19,36 +21,27 @@ import paddle.fluid.core as core from paddle.fluid.op import Operator -def output_hist(out): - hist, _ = np.histogram(out, range=(-5, 10)) - hist = hist.astype("float32") - hist /= float(out.size) - prob = 0.1 * np.ones((10)) - return hist, prob - - class TestLookupSpraseTable(OpTest): def check_with_place(self, place): scope = core.Scope() - # create and initialize Id Variable - ids = scope.var("Ids").get_tensor() - ids_array = np.array([0, 2, 3, 5, 100]).astype("int64") - ids.set(ids_array, place) - # create and initialize W Variable - rows = [0, 1, 2, 3, 4, 5, 6] - row_numel = 10000 + table_size = 10000 + row_numel = 8 w_selected_rows = scope.var('W').get_selected_rows() - w_selected_rows.set_height(len(rows)) - w_selected_rows.set_rows(rows) - w_array = np.ones((len(rows), row_numel)).astype("float32") - for i in range(len(rows)): + w_selected_rows.set_height(table_size) + w_array = np.ones((table_size, row_numel)).astype("float32") + for i in range(table_size): w_array[i] *= i w_tensor = w_selected_rows.get_tensor() w_tensor.set(w_array, place) + # create and initialize Id Variable + ids = scope.var("Ids").get_tensor() + ids_array1 = np.array([0, 2, 3, 2, 5, 0, 100]).astype("int64") + ids.set(ids_array1, place) + # create Out Variable out_tensor = scope.var('Out').get_tensor() @@ -64,16 +57,28 @@ class TestLookupSpraseTable(OpTest): lookup_table.run(scope, place) # get result from Out - result_array = np.array(out_tensor) + result_array1 = np.array(out_tensor) # all(): return True if all elements of the iterable are true (or if the iterable is empty) - for idx, row in enumerate(ids_array[:-2]): - assert (row == result_array[idx]).all() + assert (result_array1[0] == w_array[0]).all() + assert (result_array1[1] == w_array[1]).all() + assert (result_array1[2] == w_array[2]).all() + assert (result_array1[3] == w_array[1]).all() + assert (result_array1[4] == w_array[3]).all() + assert (result_array1[5] == w_array[0]).all() + assert (result_array1[6] == w_array[4]).all() + + # create and initialize Id Variable + ids = scope.var("Ids").get_tensor() + ids_array2 = np.array([4, 2, 3, 7, 100000]).astype("int64") + ids.set(ids_array2, place) + lookup_table.run(scope, place) - # check the random value - hist, prob = output_hist(result_array[-1]) - self.assertTrue( - np.allclose( - hist, prob, rtol=0, atol=0.01), "hist: " + str(hist)) + result_array2 = np.array(out_tensor) + assert (result_array2[0] == w_array[5]).all() + assert (result_array2[1] == w_array[1]).all() + assert (result_array2[2] == w_array[2]).all() + assert (result_array2[3] == w_array[6]).all() + assert (result_array2[4] == w_array[7]).all() def test_w_is_selected_rows(self): places = [core.CPUPlace()] diff --git a/python/paddle/fluid/tests/unittests/test_lookup_table_op.py b/python/paddle/fluid/tests/unittests/test_lookup_table_op.py index ac25f432df..4990ee898d 100644 --- a/python/paddle/fluid/tests/unittests/test_lookup_table_op.py +++ b/python/paddle/fluid/tests/unittests/test_lookup_table_op.py @@ -12,11 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest import paddle.fluid.core as core from paddle.fluid.op import Operator +import paddle.compat as cpt class TestLookupTableOp(OpTest): @@ -71,7 +74,7 @@ class TestLookupTableOpWithTensorIdsAndPadding(TestLookupTableOpWithTensorIds): flatten_idx = ids.flatten() padding_idx = np.random.choice(flatten_idx, 1)[0] self.outputs['Out'][np.squeeze(ids == padding_idx)] = np.zeros(31) - self.attrs = {'padding_idx': long(padding_idx)} + self.attrs = {'padding_idx': cpt.long_type(padding_idx)} self.check_output() def test_check_grad(self): diff --git a/python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py index 966a16dc87..f6bb2ab7a6 100644 --- a/python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_lrn_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from test_lrn_op import TestLRNOp diff --git a/python/paddle/fluid/tests/unittests/test_lrn_op.py b/python/paddle/fluid/tests/unittests/test_lrn_op.py index eaff45cbb2..bb91f26bbb 100644 --- a/python/paddle/fluid/tests/unittests/test_lrn_op.py +++ b/python/paddle/fluid/tests/unittests/test_lrn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest @@ -34,7 +36,7 @@ class TestLRNOp(OpTest): return x + 1 def get_out(self): - start = -(self.n - 1) / 2 + start = -(self.n - 1) // 2 end = start + self.n mid = np.empty((self.N, self.C, self.H, self.W)).astype("float32") diff --git a/python/paddle/fluid/tests/unittests/test_lstm_op.py b/python/paddle/fluid/tests/unittests/test_lstm_op.py index 705a24bd8f..76a24123fc 100644 --- a/python/paddle/fluid/tests/unittests/test_lstm_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstm_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_lstm_unit_op.py b/python/paddle/fluid/tests/unittests/test_lstm_unit_op.py index e343265874..eaa6b774c4 100644 --- a/python/paddle/fluid/tests/unittests/test_lstm_unit_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstm_unit_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_lstmp_op.py b/python/paddle/fluid/tests/unittests/test_lstmp_op.py index ed2262da4b..9c3ec45515 100644 --- a/python/paddle/fluid/tests/unittests/test_lstmp_op.py +++ b/python/paddle/fluid/tests/unittests/test_lstmp_op.py @@ -11,6 +11,8 @@ #WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. #See the License for the specific language governing permissions and #limitations under the License. + +from __future__ import print_function import unittest import numpy as np import test_lstm_op as LstmTest diff --git a/python/paddle/fluid/tests/unittests/test_margin_rank_loss_op.py b/python/paddle/fluid/tests/unittests/test_margin_rank_loss_op.py index 97c112487f..4a7e952436 100644 --- a/python/paddle/fluid/tests/unittests/test_margin_rank_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_margin_rank_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_math_op_patch.py b/python/paddle/fluid/tests/unittests/test_math_op_patch.py index 852a80261e..b25d40a3a1 100644 --- a/python/paddle/fluid/tests/unittests/test_math_op_patch.py +++ b/python/paddle/fluid/tests/unittests/test_math_op_patch.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import decorators import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_matmul_op.py b/python/paddle/fluid/tests/unittests/test_matmul_op.py index cae2c8fa87..abf10437d8 100644 --- a/python/paddle/fluid/tests/unittests/test_matmul_op.py +++ b/python/paddle/fluid/tests/unittests/test_matmul_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_maxout_op.py b/python/paddle/fluid/tests/unittests/test_maxout_op.py index f5ddf72516..d588b22fe2 100644 --- a/python/paddle/fluid/tests/unittests/test_maxout_op.py +++ b/python/paddle/fluid/tests/unittests/test_maxout_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest @@ -19,7 +21,7 @@ from op_test import OpTest def maxout_forward_naive(input, groups): s0, s1, s2, s3 = input.shape - return np.ndarray([s0, s1 / groups, groups, s2, s3], \ + return np.ndarray([s0, s1 // groups, groups, s2, s3], \ buffer = input, dtype=input.dtype).max(axis=(2)) diff --git a/python/paddle/fluid/tests/unittests/test_mean_iou.py b/python/paddle/fluid/tests/unittests/test_mean_iou.py index 32b4ee1847..03e9448317 100644 --- a/python/paddle/fluid/tests/unittests/test_mean_iou.py +++ b/python/paddle/fluid/tests/unittests/test_mean_iou.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from __future__ import division import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_mean_op.py b/python/paddle/fluid/tests/unittests/test_mean_op.py index 15472a8fc4..ff338f0e00 100644 --- a/python/paddle/fluid/tests/unittests/test_mean_op.py +++ b/python/paddle/fluid/tests/unittests/test_mean_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_memory_usage.py b/python/paddle/fluid/tests/unittests/test_memory_usage.py index f9daf83652..4cdb5b5d9f 100644 --- a/python/paddle/fluid/tests/unittests/test_memory_usage.py +++ b/python/paddle/fluid/tests/unittests/test_memory_usage.py @@ -34,7 +34,7 @@ def train_simulator(test_batch_size=10): sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001) sgd_optimizer.minimize(avg_cost) - # Calculate memory usage in current network config + # Calculate memory usage in current network config lower_usage, upper_usage, unit = fluid.contrib.memory_usage( fluid.default_main_program(), batch_size=test_batch_size) diff --git a/python/paddle/fluid/tests/unittests/test_merge_ids_op.py b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py index f209bdf30f..26ce702411 100644 --- a/python/paddle/fluid/tests/unittests/test_merge_ids_op.py +++ b/python/paddle/fluid/tests/unittests/test_merge_ids_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py b/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py index 54ee85c1a7..4e5cc91268 100644 --- a/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py +++ b/python/paddle/fluid/tests/unittests/test_mine_hard_examples_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import sys diff --git a/python/paddle/fluid/tests/unittests/test_minus_op.py b/python/paddle/fluid/tests/unittests/test_minus_op.py index ee32bd4992..54253b17b9 100644 --- a/python/paddle/fluid/tests/unittests/test_minus_op.py +++ b/python/paddle/fluid/tests/unittests/test_minus_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_modified_huber_loss_op.py b/python/paddle/fluid/tests/unittests/test_modified_huber_loss_op.py index 62035efe8e..02fecfe47e 100644 --- a/python/paddle/fluid/tests/unittests/test_modified_huber_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_modified_huber_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_momentum_op.py b/python/paddle/fluid/tests/unittests/test_momentum_op.py index c75d3bd276..7137fd0fdb 100644 --- a/python/paddle/fluid/tests/unittests/test_momentum_op.py +++ b/python/paddle/fluid/tests/unittests/test_momentum_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_mul_op.py b/python/paddle/fluid/tests/unittests/test_mul_op.py index bbc782c1bc..fca4ffa88b 100644 --- a/python/paddle/fluid/tests/unittests/test_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_mul_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_multi_file_reader.py b/python/paddle/fluid/tests/unittests/test_multi_file_reader.py index cb0ea96ff6..09788868cc 100644 --- a/python/paddle/fluid/tests/unittests/test_multi_file_reader.py +++ b/python/paddle/fluid/tests/unittests/test_multi_file_reader.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py b/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py index 7fc9f55044..4fae11e928 100644 --- a/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py +++ b/python/paddle/fluid/tests/unittests/test_multi_pass_reader.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py index 10cb78a08d..df0562dcc7 100644 --- a/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py +++ b/python/paddle/fluid/tests/unittests/test_multiclass_nms_op.py @@ -11,6 +11,8 @@ #WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. #See the License for the specific language governing permissions and #limitations under the License. + +from __future__ import print_function import unittest import numpy as np import copy diff --git a/python/paddle/fluid/tests/unittests/test_multihead_attention.py b/python/paddle/fluid/tests/unittests/test_multihead_attention.py index 80c3c67967..f60da862ac 100644 --- a/python/paddle/fluid/tests/unittests/test_multihead_attention.py +++ b/python/paddle/fluid/tests/unittests/test_multihead_attention.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_multiplex_op.py b/python/paddle/fluid/tests/unittests/test_multiplex_op.py index 03cad8b43b..1567a74808 100644 --- a/python/paddle/fluid/tests/unittests/test_multiplex_op.py +++ b/python/paddle/fluid/tests/unittests/test_multiplex_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_nce.py b/python/paddle/fluid/tests/unittests/test_nce.py index 7431a142c5..0745bd274f 100644 --- a/python/paddle/fluid/tests/unittests/test_nce.py +++ b/python/paddle/fluid/tests/unittests/test_nce.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_network_with_dtype.py b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py index d4835dd184..60dcf195da 100644 --- a/python/paddle/fluid/tests/unittests/test_network_with_dtype.py +++ b/python/paddle/fluid/tests/unittests/test_network_with_dtype.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_norm_op.py b/python/paddle/fluid/tests/unittests/test_norm_op.py index 108a665f37..22bc45ff1e 100644 --- a/python/paddle/fluid/tests/unittests/test_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_norm_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py b/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py index 198c68866d..24fdcf8c88 100644 --- a/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py +++ b/python/paddle/fluid/tests/unittests/test_normalization_wrapper.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_nvprof.py b/python/paddle/fluid/tests/unittests/test_nvprof.py index 226e5e5d11..da943d64da 100644 --- a/python/paddle/fluid/tests/unittests/test_nvprof.py +++ b/python/paddle/fluid/tests/unittests/test_nvprof.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import os import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_one_hot_op.py b/python/paddle/fluid/tests/unittests/test_one_hot_op.py index 06fccd39ac..7afdae804a 100644 --- a/python/paddle/fluid/tests/unittests/test_one_hot_op.py +++ b/python/paddle/fluid/tests/unittests/test_one_hot_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import math diff --git a/python/paddle/fluid/tests/unittests/test_op_support_gpu.py b/python/paddle/fluid/tests/unittests/test_op_support_gpu.py index 5fafb8280e..e203fccd03 100644 --- a/python/paddle/fluid/tests/unittests/test_op_support_gpu.py +++ b/python/paddle/fluid/tests/unittests/test_op_support_gpu.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_operator.py b/python/paddle/fluid/tests/unittests/test_operator.py index 5e418fe6ac..544fca8cec 100644 --- a/python/paddle/fluid/tests/unittests/test_operator.py +++ b/python/paddle/fluid/tests/unittests/test_operator.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.op as op diff --git a/python/paddle/fluid/tests/unittests/test_operator_desc.py b/python/paddle/fluid/tests/unittests/test_operator_desc.py index c098a5a0cb..6d01955993 100644 --- a/python/paddle/fluid/tests/unittests/test_operator_desc.py +++ b/python/paddle/fluid/tests/unittests/test_operator_desc.py @@ -12,9 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core +import paddle.compat as cpt from paddle.fluid.framework import Program, default_startup_program @@ -29,14 +32,15 @@ class TestOperator(unittest.TestCase): self.assertFail() except ValueError as v_err: self.assertEqual( - v_err.message, + cpt.get_exception_message(v_err), "`type` to initilized an Operator can not be None.") try: block.append_op(type="no_such_op") self.assertFail() except ValueError as a_err: - self.assertEqual(a_err.message, - "Operator \"no_such_op\" has not been registered.") + self.assertEqual( + cpt.get_exception_message(a_err), + "Operator \"no_such_op\" has not been registered.") def test_op_desc_creation(self): program = Program() diff --git a/python/paddle/fluid/tests/unittests/test_optimizer.py b/python/paddle/fluid/tests/unittests/test_optimizer.py index 18921d727f..4374d198f2 100644 --- a/python/paddle/fluid/tests/unittests/test_optimizer.py +++ b/python/paddle/fluid/tests/unittests/test_optimizer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.framework as framework diff --git a/python/paddle/fluid/tests/unittests/test_pad_op.py b/python/paddle/fluid/tests/unittests/test_pad_op.py index 300f3ffcb8..58e56ca1a4 100644 --- a/python/paddle/fluid/tests/unittests/test_pad_op.py +++ b/python/paddle/fluid/tests/unittests/test_pad_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py index d17e493c36..6d6917300c 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.conll05 as conll05 import paddle.fluid as fluid import unittest diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py index a43f2e7c49..372ef748b2 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_fetch_feed.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.dataset.flowers as flowers import math import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py index 9448d89cd5..5b96d641d6 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from parallel_executor_test_base import TestParallelExecutorBase import paddle.fluid as fluid import paddle.fluid.core as core @@ -181,7 +183,9 @@ class TestMNIST(TestParallelExecutorBase): use_parallel_executor=True) self.assertAlmostEquals( - np.mean(parallel_first_loss), single_first_loss, delta=1e-6) + np.mean(parallel_first_loss), + single_first_loss, + delta=1e-6, ) self.assertAlmostEquals( np.mean(parallel_last_loss), single_last_loss, delta=1e-6) @@ -189,7 +193,7 @@ class TestMNIST(TestParallelExecutorBase): self.check_simple_fc_parallel_accuracy(True) self.check_simple_fc_parallel_accuracy(False) - def check_batchnorm_fc_convergence(self, use_cuda): + def check_batchnorm_fc_convergence(self, use_cuda, use_fast_executor): if use_cuda and not core.is_compiled_with_cuda(): return @@ -201,11 +205,13 @@ class TestMNIST(TestParallelExecutorBase): fc_with_batchnorm, feed_dict={"image": img, "label": label}, - use_cuda=use_cuda) + use_cuda=use_cuda, + use_fast_executor=use_fast_executor) def test_batchnorm_fc(self): - self.check_batchnorm_fc_convergence(True) - self.check_batchnorm_fc_convergence(False) + for use_cuda in (False, True): + for use_fast_executor in (False, True): + self.check_batchnorm_fc_convergence(use_cuda, use_fast_executor) def test_batchnorm_fc_with_new_strategy(self): # FIXME(zcd): close this test temporally. diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py index a28428d8de..cc2d692e18 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import paddle.fluid.layers.ops as ops from paddle.fluid.initializer import init_on_cpu @@ -46,7 +48,7 @@ def squeeze_excitation(input, num_channels, reduction_ratio): pool = fluid.layers.reduce_mean(input=reshape, dim=2) squeeze = fluid.layers.fc(input=pool, - size=num_channels / reduction_ratio, + size=num_channels // reduction_ratio, act='relu') excitation = fluid.layers.fc(input=squeeze, size=num_channels, @@ -62,7 +64,7 @@ def conv_bn_layer(input, num_filters, filter_size, stride=1, groups=1, num_filters=num_filters, filter_size=filter_size, stride=stride, - padding=(filter_size - 1) / 2, + padding=(filter_size - 1) // 2, groups=groups, act=None, bias_attr=False) diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py index fcb5947ff0..f5a0ba6246 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_test_while_train.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import paddle.fluid.core as core import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py index 8203d5d1fc..5ad922725a 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid import transformer_model import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_parallel_op.py b/python/paddle/fluid/tests/unittests/test_parallel_op.py index c9617e3677..d7b9af8bac 100644 --- a/python/paddle/fluid/tests/unittests/test_parallel_op.py +++ b/python/paddle/fluid/tests/unittests/test_parallel_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_parameter.py b/python/paddle/fluid/tests/unittests/test_parameter.py index e09865074e..df42e6cb9a 100644 --- a/python/paddle/fluid/tests/unittests/test_parameter.py +++ b/python/paddle/fluid/tests/unittests/test_parameter.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from paddle.fluid.framework import default_main_program import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py index 8aff4e87f6..dfedf8190f 100644 --- a/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py +++ b/python/paddle/fluid/tests/unittests/test_polygon_box_transform.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py index 003ebba18b..14d7ed9057 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool2d_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from test_pool2d_op import TestPool2d_Op, TestCase1, TestCase2, TestCase3, TestCase4, TestCase5 diff --git a/python/paddle/fluid/tests/unittests/test_pool2d_op.py b/python/paddle/fluid/tests/unittests/test_pool2d_op.py index 1cf70311b4..26969bd523 100644 --- a/python/paddle/fluid/tests/unittests/test_pool2d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool2d_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -29,11 +31,11 @@ def max_pool2D_forward_naive(x, if global_pool == 1: ksize = [H, W] H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1 - ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 * - paddings[0]) / strides[0] + 1 + ) // strides[0] + 1 if ceil_mode else ( + H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 W_out = (W - ksize[1] + 2 * paddings[1] + strides[1] - 1 - ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 * - paddings[1]) / strides[1] + 1 + ) // strides[1] + 1 if ceil_mode else ( + W - ksize[1] + 2 * paddings[1]) // strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) for i in range(H_out): for j in range(W_out): @@ -57,11 +59,11 @@ def avg_pool2D_forward_naive(x, if global_pool == 1: ksize = [H, W] H_out = (H - ksize[0] + 2 * paddings[0] + strides[0] - 1 - ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 * - paddings[0]) / strides[0] + 1 + ) // strides[0] + 1 if ceil_mode else ( + H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 W_out = (W - ksize[1] + 2 * paddings[1] + strides[1] - 1 - ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 * - paddings[1]) / strides[1] + 1 + ) // strides[1] + 1 if ceil_mode else ( + W - ksize[1] + 2 * paddings[1]) // strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) for i in range(H_out): for j in range(W_out): diff --git a/python/paddle/fluid/tests/unittests/test_pool3d_op.py b/python/paddle/fluid/tests/unittests/test_pool3d_op.py index 92c64b3792..77045c1307 100644 --- a/python/paddle/fluid/tests/unittests/test_pool3d_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool3d_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -29,14 +31,14 @@ def max_pool3D_forward_naive(x, if global_pool == 1: ksize = [D, H, W] D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1 - ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 * - paddings[0]) / strides[0] + 1 + ) // strides[0] + 1 if ceil_mode else ( + H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H_out = (H - ksize[1] + 2 * paddings[1] + strides[1] - 1 - ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 * - paddings[1]) / strides[1] + 1 + ) // strides[1] + 1 if ceil_mode else ( + W - ksize[1] + 2 * paddings[1]) // strides[1] + 1 W_out = (W - ksize[2] + 2 * paddings[2] + strides[2] - 1 - ) / strides[2] + 1 if ceil_mode else (W - ksize[2] + 2 * - paddings[2]) / strides[2] + 1 + ) // strides[2] + 1 if ceil_mode else ( + W - ksize[2] + 2 * paddings[2]) // strides[2] + 1 out = np.zeros((N, C, D_out, H_out, W_out)) for k in range(D_out): d_start = np.max((k * strides[0] - paddings[0], 0)) @@ -63,14 +65,14 @@ def avg_pool3D_forward_naive(x, if global_pool == 1: ksize = [D, H, W] D_out = (D - ksize[0] + 2 * paddings[0] + strides[0] - 1 - ) / strides[0] + 1 if ceil_mode else (H - ksize[0] + 2 * - paddings[0]) / strides[0] + 1 + ) // strides[0] + 1 if ceil_mode else ( + H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 H_out = (H - ksize[1] + 2 * paddings[1] + strides[1] - 1 - ) / strides[1] + 1 if ceil_mode else (W - ksize[1] + 2 * - paddings[1]) / strides[1] + 1 + ) // strides[1] + 1 if ceil_mode else ( + W - ksize[1] + 2 * paddings[1]) // strides[1] + 1 W_out = (W - ksize[2] + 2 * paddings[2] + strides[2] - 1 - ) / strides[2] + 1 if ceil_mode else (W - ksize[2] + 2 * - paddings[2]) / strides[2] + 1 + ) // strides[2] + 1 if ceil_mode else ( + W - ksize[2] + 2 * paddings[2]) // strides[2] + 1 out = np.zeros((N, C, D_out, H_out, W_out)) for k in range(D_out): d_start = np.max((k * strides[0] - paddings[0], 0)) diff --git a/python/paddle/fluid/tests/unittests/test_pool_max_op.py b/python/paddle/fluid/tests/unittests/test_pool_max_op.py index e6a9f6f08c..488ff431d4 100644 --- a/python/paddle/fluid/tests/unittests/test_pool_max_op.py +++ b/python/paddle/fluid/tests/unittests/test_pool_max_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest @@ -24,9 +26,9 @@ def max_pool3D_forward_naive(x, ksize, strides, paddings, global_pool=False): ksize = [D, H, W] paddings = [0, 0, 0] - D_out = (D - ksize[0] + 2 * paddings[0]) / strides[0] + 1 - H_out = (H - ksize[1] + 2 * paddings[1]) / strides[1] + 1 - W_out = (W - ksize[2] + 2 * paddings[2]) / strides[2] + 1 + D_out = (D - ksize[0] + 2 * paddings[0]) // strides[0] + 1 + H_out = (H - ksize[1] + 2 * paddings[1]) // strides[1] + 1 + W_out = (W - ksize[2] + 2 * paddings[2]) // strides[2] + 1 out = np.zeros((N, C, D_out, H_out, W_out)) mask = np.zeros((N, C, D_out, H_out, W_out)) for k in range(D_out): @@ -63,8 +65,8 @@ def max_pool2D_forward_naive(x, ksize, strides, paddings, global_pool=False): ksize = [H, W] paddings = [0, 0] - H_out = (H - ksize[0] + 2 * paddings[0]) / strides[0] + 1 - W_out = (W - ksize[1] + 2 * paddings[1]) / strides[1] + 1 + H_out = (H - ksize[0] + 2 * paddings[0]) // strides[0] + 1 + W_out = (W - ksize[1] + 2 * paddings[1]) // strides[1] + 1 out = np.zeros((N, C, H_out, W_out)) mask = np.zeros((N, C, H_out, W_out)) for i in range(H_out): diff --git a/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py b/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py index 8c76393bda..afe8d212d6 100644 --- a/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py +++ b/python/paddle/fluid/tests/unittests/test_positive_negative_pair_op.py @@ -12,9 +12,12 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import itertools import numpy as np +import six from op_test import OpTest @@ -32,7 +35,7 @@ def py_pnpair_op(score, label, query, column=-1, weight=None): # accumulate statistics pos, neg, neu = 0, 0, 0 - for _, ranks in list(predictions.items()): + for _, ranks in six.iteritems(predictions): for e1, e2 in itertools.combinations(ranks, 2): s1, s2, l1, l2, w1, w2 = e1[0], e2[0], e1[1], e2[1], e1[2], e2[2] w = (w1 + w2) * 0.5 diff --git a/python/paddle/fluid/tests/unittests/test_precision_recall_op.py b/python/paddle/fluid/tests/unittests/test_precision_recall_op.py index 5ae425fee1..6456376259 100644 --- a/python/paddle/fluid/tests/unittests/test_precision_recall_op.py +++ b/python/paddle/fluid/tests/unittests/test_precision_recall_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_prelu_op.py b/python/paddle/fluid/tests/unittests/test_prelu_op.py index cb7de3fc93..979be5af3b 100644 --- a/python/paddle/fluid/tests/unittests/test_prelu_op.py +++ b/python/paddle/fluid/tests/unittests/test_prelu_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_preprocessor.py b/python/paddle/fluid/tests/unittests/test_preprocessor.py index cbf1a7e0c5..98e609b769 100644 --- a/python/paddle/fluid/tests/unittests/test_preprocessor.py +++ b/python/paddle/fluid/tests/unittests/test_preprocessor.py @@ -12,12 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np +import paddle import paddle.fluid as fluid -import paddle.v2 as paddle -import paddle.v2.dataset.mnist as mnist +import paddle.dataset.mnist as mnist class TestPreprocessor(unittest.TestCase): diff --git a/python/paddle/fluid/tests/unittests/test_print_op.py b/python/paddle/fluid/tests/unittests/test_print_op.py index b461c5c940..ac682d6181 100644 --- a/python/paddle/fluid/tests/unittests/test_print_op.py +++ b/python/paddle/fluid/tests/unittests/test_print_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core from paddle.fluid.executor import Executor diff --git a/python/paddle/fluid/tests/unittests/test_prior_box_op.py b/python/paddle/fluid/tests/unittests/test_prior_box_op.py index e15554737b..7381b74af7 100644 --- a/python/paddle/fluid/tests/unittests/test_prior_box_op.py +++ b/python/paddle/fluid/tests/unittests/test_prior_box_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import sys diff --git a/python/paddle/fluid/tests/unittests/test_profiler.py b/python/paddle/fluid/tests/unittests/test_profiler.py index 9f8d33f9bb..7934164b84 100644 --- a/python/paddle/fluid/tests/unittests/test_profiler.py +++ b/python/paddle/fluid/tests/unittests/test_profiler.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import os import numpy as np @@ -23,9 +25,6 @@ import paddle.fluid.core as core class TestProfiler(unittest.TestCase): def net_profiler(self, state, profile_path='/tmp/profile'): - enable_if_gpu = state == 'GPU' or state == "All" - if enable_if_gpu and not core.is_compiled_with_cuda(): - return startup_program = fluid.Program() main_program = fluid.Program() @@ -79,8 +78,6 @@ class TestProfiler(unittest.TestCase): pass_acc_calculator.add(value=acc, weight=b_size) pass_acc = pass_acc_calculator.eval() - @unittest.skipIf(not core.is_compiled_with_cuda(), - "profiler is enabled only with GPU") def test_cpu_profiler(self): self.net_profiler('CPU') @@ -93,7 +90,7 @@ class TestProfiler(unittest.TestCase): "profiler is enabled only with GPU") def test_all_profiler(self): self.net_profiler('All', '/tmp/profile_out') - with open('/tmp/profile_out', 'r') as f: + with open('/tmp/profile_out', 'rb') as f: self.assertGreater(len(f.read()), 0) diff --git a/python/paddle/fluid/tests/unittests/test_protobuf.py b/python/paddle/fluid/tests/unittests/test_protobuf.py index c3f1fa8018..7b80927c48 100644 --- a/python/paddle/fluid/tests/unittests/test_protobuf.py +++ b/python/paddle/fluid/tests/unittests/test_protobuf.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid.proto.framework_pb2 as framework_pb2 import unittest diff --git a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py index 9853fb4e9a..d24b5cbd06 100644 --- a/python/paddle/fluid/tests/unittests/test_protobuf_descs.py +++ b/python/paddle/fluid/tests/unittests/test_protobuf_descs.py @@ -12,8 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core +import paddle.compat as cpt from paddle.fluid.framework import Program @@ -108,7 +111,7 @@ class TestVarDesc(unittest.TestCase): def test_shape(self): program_desc = core.ProgramDesc() block = program_desc.block(0) - var = block.var('my_var') + var = block.var(cpt.to_bytes('my_var')) var.set_type(core.VarDesc.VarType.SELECTED_ROWS) src_shape = [3, 2, 10, 8] var.set_shape(src_shape) @@ -119,7 +122,7 @@ class TestVarDesc(unittest.TestCase): def test_multiple_shape(self): program_desc = core.ProgramDesc() block = program_desc.block(0) - var = block.var('my_reader') + var = block.var(cpt.to_bytes('my_reader')) var.set_type(core.VarDesc.VarType.READER) src_shapes = [[2, 3, 3], [4, 5], [6, 7, 8, 9]] var.set_shapes(src_shapes) @@ -130,7 +133,7 @@ class TestVarDesc(unittest.TestCase): def test_dtype(self): program_desc = core.ProgramDesc() block = program_desc.block(0) - var = block.var('my_var') + var = block.var(cpt.to_bytes('my_var')) var.set_type(core.VarDesc.VarType.LOD_TENSOR) var.set_dtype(core.VarDesc.VarType.INT32) self.assertEqual(core.VarDesc.VarType.INT32, var.dtype()) @@ -139,7 +142,7 @@ class TestVarDesc(unittest.TestCase): def test_multiple_dtype(self): program_desc = core.ProgramDesc() block = program_desc.block(0) - var = block.var('my_reader') + var = block.var(cpt.to_bytes('my_reader')) var.set_type(core.VarDesc.VarType.READER) src_types = [ core.VarDesc.VarType.INT32, core.VarDesc.VarType.FP64, @@ -152,7 +155,7 @@ class TestVarDesc(unittest.TestCase): def test_multiple_lod_level(self): program_desc = core.ProgramDesc() block = program_desc.block(0) - var = block.var('my_reader') + var = block.var(cpt.to_bytes('my_reader')) var.set_type(core.VarDesc.VarType.READER) src_types = [3, 1, 2] var.set_lod_levels(src_types) @@ -166,12 +169,12 @@ class TestBlockDesc(unittest.TestCase): self.assertIsNotNone(program_desc) block = program_desc.block(0) self.assertIsNotNone(block) - var1 = block.var("var1") - var2 = block.var("var2") - var3 = block.var("var3") + var1 = block.var(cpt.to_bytes("var1")) + var2 = block.var(cpt.to_bytes("var2")) + var3 = block.var(cpt.to_bytes("var3")) all_vars = block.all_vars() self.assertEqual(set(all_vars), {var1, var2, var3}) - var2_re = block.find_var("var2") + var2_re = block.find_var(cpt.to_bytes("var2")) self.assertEqual(var2_re, var2) def test_add_op(self): diff --git a/python/paddle/fluid/tests/unittests/test_proximal_adagrad_op.py b/python/paddle/fluid/tests/unittests/test_proximal_adagrad_op.py index 3c26895850..57e96f1fa3 100644 --- a/python/paddle/fluid/tests/unittests/test_proximal_adagrad_op.py +++ b/python/paddle/fluid/tests/unittests/test_proximal_adagrad_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_proximal_gd_op.py b/python/paddle/fluid/tests/unittests/test_proximal_gd_op.py index 137594b9a0..067502baec 100644 --- a/python/paddle/fluid/tests/unittests/test_proximal_gd_op.py +++ b/python/paddle/fluid/tests/unittests/test_proximal_gd_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py b/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py index f9bda5e470..3efe5aac88 100644 --- a/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py +++ b/python/paddle/fluid/tests/unittests/test_py_reader_push_pop.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py index 9a379bdbaa..931cac409f 100644 --- a/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py +++ b/python/paddle/fluid/tests/unittests/test_py_reader_using_executor.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_random_crop_op.py b/python/paddle/fluid/tests/unittests/test_random_crop_op.py index 1c708d0386..f29dddff7a 100644 --- a/python/paddle/fluid/tests/unittests/test_random_crop_op.py +++ b/python/paddle/fluid/tests/unittests/test_random_crop_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core @@ -21,11 +23,12 @@ from op_test import OpTest class TestRandomCropOp(OpTest): def setUp(self): to_crop = np.array([[[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12]]] * - 5).astype("float32") + 5).astype(np.int32) self.possible_res = [ - np.array([[1, 2, 3], [5, 6, 7]]), np.array([[2, 3, 4], [6, 7, 8]]), - np.array([[5, 6, 7], [9, 10, 11]]), - np.array([[6, 7, 8], [10, 11, 12]]) + np.array([[1, 2, 3], [5, 6, 7]]).astype(np.int32), + np.array([[2, 3, 4], [6, 7, 8]]).astype(np.int32), + np.array([[5, 6, 7], [9, 10, 11]]).astype(np.int32), + np.array([[6, 7, 8], [10, 11, 12]]).astype(np.int32) ] self.op_type = "random_crop" self.inputs = {'X': to_crop, 'Seed': np.array([10])} diff --git a/python/paddle/fluid/tests/unittests/test_rank_loss_op.py b/python/paddle/fluid/tests/unittests/test_rank_loss_op.py index 7eba1e2077..c9fa24b103 100644 --- a/python/paddle/fluid/tests/unittests/test_rank_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_rank_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_reader_reset.py b/python/paddle/fluid/tests/unittests/test_reader_reset.py index 3ad85d5748..8ad11d76f6 100644 --- a/python/paddle/fluid/tests/unittests/test_reader_reset.py +++ b/python/paddle/fluid/tests/unittests/test_reader_reset.py @@ -12,8 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid as fluid -import paddle.v2 as paddle +import paddle import numpy as np import unittest diff --git a/python/paddle/fluid/tests/unittests/test_recordio_reader.py b/python/paddle/fluid/tests/unittests/test_recordio_reader.py index 69a522e273..c5210bb208 100644 --- a/python/paddle/fluid/tests/unittests/test_recordio_reader.py +++ b/python/paddle/fluid/tests/unittests/test_recordio_reader.py @@ -12,11 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid -import paddle.v2 as paddle -import paddle.v2.dataset.mnist as mnist +import paddle +import paddle.dataset.mnist as mnist class TestRecordIO(unittest.TestCase): diff --git a/python/paddle/fluid/tests/unittests/test_recurrent_op.py b/python/paddle/fluid/tests/unittests/test_recurrent_op.py index 2e22df2beb..6dfc85e301 100644 --- a/python/paddle/fluid/tests/unittests/test_recurrent_op.py +++ b/python/paddle/fluid/tests/unittests/test_recurrent_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.layers as layers diff --git a/python/paddle/fluid/tests/unittests/test_reduce_op.py b/python/paddle/fluid/tests/unittests/test_reduce_op.py index 06d116601b..328f0f0011 100644 --- a/python/paddle/fluid/tests/unittests/test_reduce_op.py +++ b/python/paddle/fluid/tests/unittests/test_reduce_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_registry.py b/python/paddle/fluid/tests/unittests/test_registry.py index a361c4624e..7381bb61eb 100644 --- a/python/paddle/fluid/tests/unittests/test_registry.py +++ b/python/paddle/fluid/tests/unittests/test_registry.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_regularizer.py b/python/paddle/fluid/tests/unittests/test_regularizer.py index 9b1c4ceada..6727335c60 100644 --- a/python/paddle/fluid/tests/unittests/test_regularizer.py +++ b/python/paddle/fluid/tests/unittests/test_regularizer.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.framework as framework diff --git a/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py b/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py index 6e1cd56b3e..28c8c4699a 100644 --- a/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py +++ b/python/paddle/fluid/tests/unittests/test_reorder_lod_tensor.py @@ -12,11 +12,14 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid import paddle.fluid.core as core from paddle.fluid.layers.control_flow import lod_rank_table import numpy +import functools class TestReorderLoDTensor(unittest.TestCase): @@ -101,7 +104,8 @@ class TestReorderLoDTensor(unittest.TestCase): rank_table = [] # list of (index, length) for i in range(len(ref_lod)): rank_table.append((i, ref_lod[i])) - rank_table = sorted(rank_table, lambda x, y: y[1] - x[1]) + rank_table = sorted( + rank_table, key=functools.cmp_to_key(lambda x, y: y[1] - x[1])) # compute the input sequence info according to input_lod input_value, input_lod = self.data[self.data_desc[0][0]] diff --git a/python/paddle/fluid/tests/unittests/test_reshape_op.py b/python/paddle/fluid/tests/unittests/test_reshape_op.py index 2f5558578a..1de35dc35b 100644 --- a/python/paddle/fluid/tests/unittests/test_reshape_op.py +++ b/python/paddle/fluid/tests/unittests/test_reshape_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_reverse_op.py b/python/paddle/fluid/tests/unittests/test_reverse_op.py index f845575a02..e83f548c22 100644 --- a/python/paddle/fluid/tests/unittests/test_reverse_op.py +++ b/python/paddle/fluid/tests/unittests/test_reverse_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py index 0d84a5853e..3d4623c74d 100644 --- a/python/paddle/fluid/tests/unittests/test_rmsprop_op.py +++ b/python/paddle/fluid/tests/unittests/test_rmsprop_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_rnn_memory_helper_op.py b/python/paddle/fluid/tests/unittests/test_rnn_memory_helper_op.py index 178606f059..9bfec8e9bd 100644 --- a/python/paddle/fluid/tests/unittests/test_rnn_memory_helper_op.py +++ b/python/paddle/fluid/tests/unittests/test_rnn_memory_helper_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from paddle.fluid.framework import Program diff --git a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py index df5684ab17..ed7f467835 100644 --- a/python/paddle/fluid/tests/unittests/test_roi_pool_op.py +++ b/python/paddle/fluid/tests/unittests/test_roi_pool_op.py @@ -12,10 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import math import sys +import paddle.compat as cpt from op_test import OpTest @@ -59,10 +62,10 @@ class TestROIPoolOp(OpTest): for i in range(self.rois_num): roi = self.rois[i] roi_batch_id = roi[0] - roi_start_w = int(round(roi[1] * self.spatial_scale)) - roi_start_h = int(round(roi[2] * self.spatial_scale)) - roi_end_w = int(round(roi[3] * self.spatial_scale)) - roi_end_h = int(round(roi[4] * self.spatial_scale)) + roi_start_w = int(cpt.round(roi[1] * self.spatial_scale)) + roi_start_h = int(cpt.round(roi[2] * self.spatial_scale)) + roi_end_w = int(cpt.round(roi[3] * self.spatial_scale)) + roi_end_h = int(cpt.round(roi[4] * self.spatial_scale)) roi_height = int(max(roi_end_h - roi_start_h + 1, 1)) roi_width = int(max(roi_end_w - roi_start_w + 1, 1)) @@ -97,8 +100,8 @@ class TestROIPoolOp(OpTest): for w in range(wstart, wend): if x_i[c, h, w] > out_data[i, c, ph, pw]: out_data[i, c, ph, pw] = x_i[c, h, w] - argmax_data[i, c, ph, pw] = h * \ - self.width + w + argmax_data[i, c, ph, + pw] = h * self.width + w self.outs = out_data.astype('float32') self.argmaxes = argmax_data.astype('int64') @@ -110,14 +113,14 @@ class TestROIPoolOp(OpTest): self.rois_lod[0].append(bno + 1) for i in range(bno + 1): x1 = np.random.random_integers( - 0, self.width / self.spatial_scale - self.pooled_width) + 0, self.width // self.spatial_scale - self.pooled_width) y1 = np.random.random_integers( - 0, self.height / self.spatial_scale - self.pooled_height) + 0, self.height // self.spatial_scale - self.pooled_height) x2 = np.random.random_integers(x1 + self.pooled_width, - self.width / self.spatial_scale) - y2 = np.random.random_integers(y1 + self.pooled_height, - self.height / self.spatial_scale) + self.width // self.spatial_scale) + y2 = np.random.random_integers( + y1 + self.pooled_height, self.height // self.spatial_scale) roi = [bno, x1, y1, x2, y2] rois.append(roi) diff --git a/python/paddle/fluid/tests/unittests/test_row_conv_op.py b/python/paddle/fluid/tests/unittests/test_row_conv_op.py index 07dcd10868..2f13f067ef 100644 --- a/python/paddle/fluid/tests/unittests/test_row_conv_op.py +++ b/python/paddle/fluid/tests/unittests/test_row_conv_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py b/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py index df6e0faaca..08c462d903 100644 --- a/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py +++ b/python/paddle/fluid/tests/unittests/test_rpn_target_assign_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_sampling_id_op.py b/python/paddle/fluid/tests/unittests/test_sampling_id_op.py new file mode 100644 index 0000000000..708265b457 --- /dev/null +++ b/python/paddle/fluid/tests/unittests/test_sampling_id_op.py @@ -0,0 +1,51 @@ +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import unittest +import numpy as np +from op_test import OpTest + +import paddle.fluid.core as core +from paddle.fluid.op import Operator + + +class TestSamplingIdOp(OpTest): + def setUp(self): + self.op_type = "sampling_id" + self.use_mkldnn = False + self.init_kernel_type() + self.X = np.random.random((8, 4)).astype('float32') + self.inputs = {"X": self.X} + self.Y = np.random.random(8).astype('float32') + self.outputs = {'Out': self.Y} + self.attrs = {'max': 1.0, 'min': 0.0, 'seed': 1} + + def test_check_output(self): + self.check_output_customized(self.verify_output) + y1 = self.out + self.check_output_customized(self.verify_output) + y2 = self.out + self.assertTrue(np.array_equal(y1, y2)) + self.assertEqual(len(y1), len(self.Y)) + + def verify_output(self, outs): + out = np.array(outs[0]) + self.out = out + + def init_kernel_type(self): + pass + + +if __name__ == "__main__": + unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_scale_op.py b/python/paddle/fluid/tests/unittests/test_scale_op.py index 53f59c3990..0a8a43253d 100644 --- a/python/paddle/fluid/tests/unittests/test_scale_op.py +++ b/python/paddle/fluid/tests/unittests/test_scale_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_scatter_op.py b/python/paddle/fluid/tests/unittests/test_scatter_op.py index fb17287436..088996f9d7 100644 --- a/python/paddle/fluid/tests/unittests/test_scatter_op.py +++ b/python/paddle/fluid/tests/unittests/test_scatter_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_scope.py b/python/paddle/fluid/tests/unittests/test_scope.py index d249a989a9..45fcbfba6e 100644 --- a/python/paddle/fluid/tests/unittests/test_scope.py +++ b/python/paddle/fluid/tests/unittests/test_scope.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid.core import unittest diff --git a/python/paddle/fluid/tests/unittests/test_selected_rows.py b/python/paddle/fluid/tests/unittests/test_selected_rows.py index f504a06fff..2f34f79b8e 100644 --- a/python/paddle/fluid/tests/unittests/test_selected_rows.py +++ b/python/paddle/fluid/tests/unittests/test_selected_rows.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid.core as core import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_seq_concat_op.py b/python/paddle/fluid/tests/unittests/test_seq_concat_op.py index 11ffa761a6..9d1d139721 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_concat_op.py +++ b/python/paddle/fluid/tests/unittests/test_seq_concat_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import sys diff --git a/python/paddle/fluid/tests/unittests/test_seq_conv.py b/python/paddle/fluid/tests/unittests/test_seq_conv.py index 1a6e1aad79..dcc86382e5 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_conv.py +++ b/python/paddle/fluid/tests/unittests/test_seq_conv.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import random diff --git a/python/paddle/fluid/tests/unittests/test_seq_pool.py b/python/paddle/fluid/tests/unittests/test_seq_pool.py index 0b3659d7a6..66e77714c5 100644 --- a/python/paddle/fluid/tests/unittests/test_seq_pool.py +++ b/python/paddle/fluid/tests/unittests/test_seq_pool.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py b/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py index 8f0765277a..92cd5b0cbc 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_erase_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_sequence_expand.py b/python/paddle/fluid/tests/unittests/test_sequence_expand.py index 5ff0dab23e..ffd4026dba 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_expand.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_expand.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_sequence_reshape.py b/python/paddle/fluid/tests/unittests/test_sequence_reshape.py index 39b02ecf6d..f11fa6c39c 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_reshape.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_reshape.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import math diff --git a/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py b/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py index 313e485d1e..1561490087 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_slice_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import sys diff --git a/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py b/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py index c4fc8b74cf..3e00e7d95f 100644 --- a/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_sequence_softmax_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_sgd_op.py b/python/paddle/fluid/tests/unittests/test_sgd_op.py index 3126293f9d..b46e4bfb86 100644 --- a/python/paddle/fluid/tests/unittests/test_sgd_op.py +++ b/python/paddle/fluid/tests/unittests/test_sgd_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import paddle.fluid.core as core @@ -124,6 +126,7 @@ class TestSGDOpOptimizeSelectedRows(unittest.TestCase): w_selected_rows = scope.var('Param').get_selected_rows() w_selected_rows.set_height(len(param_rows)) w_selected_rows.set_rows(param_rows) + w_selected_rows.sync_index() w_array = np.ones((len(param_rows), row_width)).astype("float32") for i in range(len(param_rows)): w_array[i] *= i diff --git a/python/paddle/fluid/tests/unittests/test_shape_op.py b/python/paddle/fluid/tests/unittests/test_shape_op.py index a62ee05007..02231ea943 100644 --- a/python/paddle/fluid/tests/unittests/test_shape_op.py +++ b/python/paddle/fluid/tests/unittests/test_shape_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py b/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py index a994bf181a..97f79f9421 100644 --- a/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py +++ b/python/paddle/fluid/tests/unittests/test_shrink_rnn_memory.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core from paddle.fluid.executor import Executor diff --git a/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py b/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py index c435796569..97ff203499 100644 --- a/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py +++ b/python/paddle/fluid/tests/unittests/test_sigmoid_cross_entropy_with_logits_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np from op_test import OpTest from scipy.special import logit diff --git a/python/paddle/fluid/tests/unittests/test_sign_op.py b/python/paddle/fluid/tests/unittests/test_sign_op.py index 087a0c575b..85a9d9cae4 100644 --- a/python/paddle/fluid/tests/unittests/test_sign_op.py +++ b/python/paddle/fluid/tests/unittests/test_sign_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_slice_op.py b/python/paddle/fluid/tests/unittests/test_slice_op.py index 1a48bce3bb..134df38eea 100644 --- a/python/paddle/fluid/tests/unittests/test_slice_op.py +++ b/python/paddle/fluid/tests/unittests/test_slice_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_slice_var.py b/python/paddle/fluid/tests/unittests/test_slice_var.py index 82305b23a1..fab63b7d56 100644 --- a/python/paddle/fluid/tests/unittests/test_slice_var.py +++ b/python/paddle/fluid/tests/unittests/test_slice_var.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import math import unittest from paddle.fluid.transpiler.distribute_transpiler import slice_variable diff --git a/python/paddle/fluid/tests/unittests/test_smooth_l1_loss_op.py b/python/paddle/fluid/tests/unittests/test_smooth_l1_loss_op.py index e74664dac4..8ab6833821 100644 --- a/python/paddle/fluid/tests/unittests/test_smooth_l1_loss_op.py +++ b/python/paddle/fluid/tests/unittests/test_smooth_l1_loss_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_softmax_op.py b/python/paddle/fluid/tests/unittests/test_softmax_op.py index 70ad05597c..d88aa1ae1c 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py b/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py index c0d9fc8f22..b7e5ff6d52 100644 --- a/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py +++ b/python/paddle/fluid/tests/unittests/test_softmax_with_cross_entropy_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py b/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py index ea1146166d..5397d5c521 100644 --- a/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_and_merge_lod_tensor_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_split_ids_op.py b/python/paddle/fluid/tests/unittests/test_split_ids_op.py index ca78613098..4c3d025898 100644 --- a/python/paddle/fluid/tests/unittests/test_split_ids_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_ids_op.py @@ -12,8 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np +import six from op_test import OpTest import paddle.fluid.core as core from paddle.fluid.op import Operator @@ -59,7 +62,7 @@ class TestSpliteIds(unittest.TestCase): x_tensor = x.get_tensor() x_tensor.set(np_array, place) - outs_name = ["out%d" % i for i in xrange(3)] + outs_name = ["out%d" % i for i in six.moves.xrange(3)] outs = [ scope.var(var_name).get_selected_rows() for var_name in outs_name ] diff --git a/python/paddle/fluid/tests/unittests/test_split_op.py b/python/paddle/fluid/tests/unittests/test_split_op.py index 6b67a52e81..3c5dd782f8 100644 --- a/python/paddle/fluid/tests/unittests/test_split_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py index 2b261820e0..41a5ee59ea 100644 --- a/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py +++ b/python/paddle/fluid/tests/unittests/test_split_selected_rows_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core import numpy as np diff --git a/python/paddle/fluid/tests/unittests/test_spp_op.py b/python/paddle/fluid/tests/unittests/test_spp_op.py index 3cbfc2a703..a6c2cccd39 100644 --- a/python/paddle/fluid/tests/unittests/test_spp_op.py +++ b/python/paddle/fluid/tests/unittests/test_spp_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_squared_l2_distance_op.py b/python/paddle/fluid/tests/unittests/test_squared_l2_distance_op.py index 78bc300ebe..a8bc1004d9 100644 --- a/python/paddle/fluid/tests/unittests/test_squared_l2_distance_op.py +++ b/python/paddle/fluid/tests/unittests/test_squared_l2_distance_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_squared_l2_norm_op.py b/python/paddle/fluid/tests/unittests/test_squared_l2_norm_op.py index 609445d522..439bae9510 100644 --- a/python/paddle/fluid/tests/unittests/test_squared_l2_norm_op.py +++ b/python/paddle/fluid/tests/unittests/test_squared_l2_norm_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import unittest from numpy import linalg as LA diff --git a/python/paddle/fluid/tests/unittests/test_squeeze_op.py b/python/paddle/fluid/tests/unittests/test_squeeze_op.py index bca6af2fd5..2be8e24a0f 100644 --- a/python/paddle/fluid/tests/unittests/test_squeeze_op.py +++ b/python/paddle/fluid/tests/unittests/test_squeeze_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -39,7 +41,7 @@ class TestSqueezeOp(OpTest): self.new_shape = (3, 5) def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": False} + self.attrs = {"axes": self.axes} # Correct: There is mins axis. @@ -66,49 +68,5 @@ class TestSqueezeOp3(TestSqueezeOp): self.new_shape = (3, 5, 1, 4) -# Correct: Inplace. -class TestSqueezeOpInplace1(TestSqueezeOp): - def init_test_case(self): - self.ori_shape = (1, 3, 1, 5) - self.axes = (0, 2) - self.new_shape = (3, 5) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - -# Correct: Inplace. There is mins axis. -class TestSqueezeOpInplace2(TestSqueezeOp): - def inti_test_case(self): - self.ori_shape = (1, 3, 1, 5) - self.axes = (0, -2) - self.new_shape = (3, 5) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - -# Correct: Inplace. No axes input. -class TestSqueezeOpInplace3(TestSqueezeOp): - def init_test_case(self): - self.ori_shape = (1, 3, 1, 5) - self.axes = () - self.new_shape = (3, 5) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - -# Correct: Inpalce. Just part of axes be squeezed. -class TestSqueezeOpInplace4(TestSqueezeOp): - def init_test_case(self): - self.ori_shape = (3, 1, 5, 1, 4, 1) - self.axes = (1, -1) - self.new_shape = (3, 5, 1, 4) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py b/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py index 7956897d68..55820f31b8 100644 --- a/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py +++ b/python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from test_sum_op import TestSumOp diff --git a/python/paddle/fluid/tests/unittests/test_sum_op.py b/python/paddle/fluid/tests/unittests/test_sum_op.py index 1d90414e13..74797bb656 100644 --- a/python/paddle/fluid/tests/unittests/test_sum_op.py +++ b/python/paddle/fluid/tests/unittests/test_sum_op.py @@ -12,9 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core +from paddle.fluid.op import Operator class TestSumOp(OpTest): @@ -40,5 +44,66 @@ class TestSumOp(OpTest): pass +class TestSelectedRowsSumOp(OpTest): + def check_with_place(self, place): + scope = core.Scope() + self.check_input_and_optput(scope, place, True, True, True) + self.check_input_and_optput(scope, place, False, True, True) + self.check_input_and_optput(scope, place, False, False, True) + self.check_input_and_optput(scope, place, False, False, False) + + def check_input_and_optput(self, + scope, + place, + w1_has_data=False, + w2_has_data=False, + w3_has_data=False): + + self.create_selected_rows(scope, place, "W1", w1_has_data) + self.create_selected_rows(scope, place, "W2", w2_has_data) + self.create_selected_rows(scope, place, "W3", w3_has_data) + + # create Out Variable + out = scope.var('Out').get_selected_rows() + + # create and run sum operator + sum_op = Operator("sum", X=["W1", "W2", "W3"], Out='Out') + sum_op.run(scope, place) + + has_data_w_num = 0 + for w in [w1_has_data, w2_has_data, w3_has_data]: + if not w: + has_data_w_num += 1 + + self.assertEqual(7 * has_data_w_num, len(out.rows())) + + def create_selected_rows(self, scope, place, var_name, isEmpty): + # create and initialize W Variable + if not isEmpty: + rows = [0, 1, 2, 3, 4, 5, 6] + row_numel = 12 + else: + rows = [] + row_numel = 12 + + var = scope.var(var_name) + w_selected_rows = var.get_selected_rows() + w_selected_rows.set_height(len(rows)) + w_selected_rows.set_rows(rows) + w_array = np.ones((len(rows), row_numel)).astype("float32") + for i in range(len(rows)): + w_array[i] *= i + w_tensor = w_selected_rows.get_tensor() + w_tensor.set(w_array, place) + + return var + + def test_w_is_selected_rows(self): + places = [core.CPUPlace()] + # currently only support CPU + for place in places: + self.check_with_place(place) + + if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_switch.py b/python/paddle/fluid/tests/unittests/test_switch.py index 528c5cce4b..2a9c07a889 100644 --- a/python/paddle/fluid/tests/unittests/test_switch.py +++ b/python/paddle/fluid/tests/unittests/test_switch.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_target_assign_op.py b/python/paddle/fluid/tests/unittests/test_target_assign_op.py index bd20889752..aec219f806 100644 --- a/python/paddle/fluid/tests/unittests/test_target_assign_op.py +++ b/python/paddle/fluid/tests/unittests/test_target_assign_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np import random diff --git a/python/paddle/fluid/tests/unittests/test_tensor.py b/python/paddle/fluid/tests/unittests/test_tensor.py index 5ccc876ae8..e9d0f8a019 100644 --- a/python/paddle/fluid/tests/unittests/test_tensor.py +++ b/python/paddle/fluid/tests/unittests/test_tensor.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import paddle.fluid.core as core import unittest import numpy diff --git a/python/paddle/fluid/tests/unittests/test_top_k_op.py b/python/paddle/fluid/tests/unittests/test_top_k_op.py index cbc3da5503..e54e170f7f 100644 --- a/python/paddle/fluid/tests/unittests/test_top_k_op.py +++ b/python/paddle/fluid/tests/unittests/test_top_k_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_transpose_op.py b/python/paddle/fluid/tests/unittests/test_transpose_op.py index ebd63fbd49..0853f80b82 100644 --- a/python/paddle/fluid/tests/unittests/test_transpose_op.py +++ b/python/paddle/fluid/tests/unittests/test_transpose_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_uniform_random_batch_size_like_op.py b/python/paddle/fluid/tests/unittests/test_uniform_random_batch_size_like_op.py index e033e86114..7b8be24d9d 100644 --- a/python/paddle/fluid/tests/unittests/test_uniform_random_batch_size_like_op.py +++ b/python/paddle/fluid/tests/unittests/test_uniform_random_batch_size_like_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_uniform_random_op.py b/python/paddle/fluid/tests/unittests/test_uniform_random_op.py index 346a949b6e..d6a5d68765 100644 --- a/python/paddle/fluid/tests/unittests/test_uniform_random_op.py +++ b/python/paddle/fluid/tests/unittests/test_uniform_random_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest diff --git a/python/paddle/fluid/tests/unittests/test_unique_name.py b/python/paddle/fluid/tests/unittests/test_unique_name.py index 49ef335618..b8c751b2e9 100644 --- a/python/paddle/fluid/tests/unittests/test_unique_name.py +++ b/python/paddle/fluid/tests/unittests/test_unique_name.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid as fluid diff --git a/python/paddle/fluid/tests/unittests/test_unpool_op.py b/python/paddle/fluid/tests/unittests/test_unpool_op.py index ecce4cdde2..b0c7c3c866 100644 --- a/python/paddle/fluid/tests/unittests/test_unpool_op.py +++ b/python/paddle/fluid/tests/unittests/test_unpool_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np from op_test import OpTest @@ -27,7 +29,7 @@ def unpool2dmax_forward_naive(input, indices, ksize, strides, paddings): for h in range(s2): for w in range(s3): index = indices[nidx, cidx, h, w] - hidx = (index - index % out_wsize) / out_wsize + hidx = (index - index % out_wsize) // out_wsize widx = index % out_wsize out[nidx, cidx, int(hidx), int(widx)] = \ input[nidx, cidx, h, w] @@ -41,9 +43,9 @@ class TestUnpoolOp(OpTest): self.init_test_case() pre_input = np.random.random(self.shape).astype("float32") nsize, csize, hsize, wsize = pre_input.shape - hsize_out = (hsize - self.ksize[0] + 2 * self.paddings[0]) / \ + hsize_out = (hsize - self.ksize[0] + 2 * self.paddings[0]) // \ self.strides[0] + 1 - wsize_out = (wsize - self.ksize[1] + 2 * self.paddings[1]) / \ + wsize_out = (wsize - self.ksize[1] + 2 * self.paddings[1]) // \ self.strides[1] + 1 input = np.zeros((nsize, csize, hsize_out, wsize_out)) indices = np.zeros((nsize, csize, hsize_out, wsize_out)) @@ -62,7 +64,7 @@ class TestUnpoolOp(OpTest): input[nidx, cidx, i, j] = x_masked.max() arg = x_masked.argmax() indices[nidx, cidx, i, j] = \ - (r_start + arg / self.ksize[1]) * wsize + \ + (r_start + arg // self.ksize[1]) * wsize + \ c_start + arg % self.ksize[1] output = self.unpool2d_forward_naive(input, indices, self.ksize, \ self.strides, self.paddings).astype("float32") diff --git a/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py b/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py index 7a4aa0a40b..a324438ba5 100644 --- a/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py +++ b/python/paddle/fluid/tests/unittests/test_unsqueeze_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy as np @@ -39,7 +41,7 @@ class TestUnsqueezeOp(OpTest): self.new_shape = (3, 1, 1, 5) def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": False} + self.attrs = {"axes": self.axes} # Correct: Single input index. @@ -74,38 +76,5 @@ class TestUnsqueezeOp4(TestUnsqueezeOp): self.new_shape = (3, 1, 1, 2, 5, 1) -# Correct: Inplace. -class TestUnsqueezeOpInplace1(TestUnsqueezeOp): - def init_test_case(self): - self.ori_shape = (3, 5) - self.axes = (0, 2) - self.new_shape = (1, 3, 1, 5) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - -# Correct: Inplace. There is mins index. -class TestUnsqueezeOpInplace2(TestUnsqueezeOp): - def init_test_case(self): - self.ori_shape = (3, 5) - self.axes = (0, -2) - self.new_shape = (1, 3, 1, 5) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - -# Correct: Inplace. There is duplicated axis. -class TestUnsqueezeOpInplace3(TestUnsqueezeOp): - def init_test_case(self): - self.ori_shape = (3, 2, 5) - self.axes = (0, 3, 3) - self.new_shape = (1, 3, 2, 1, 1, 5) - - def init_attrs(self): - self.attrs = {"axes": self.axes, "inplace": True} - - if __name__ == "__main__": unittest.main() diff --git a/python/paddle/fluid/tests/unittests/test_variable.py b/python/paddle/fluid/tests/unittests/test_variable.py index 49784e21c4..b0830e130d 100644 --- a/python/paddle/fluid/tests/unittests/test_variable.py +++ b/python/paddle/fluid/tests/unittests/test_variable.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest from paddle.fluid.framework import default_main_program, Program, convert_np_dtype_to_dtype_ import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/test_version.py b/python/paddle/fluid/tests/unittests/test_version.py index a09c8a759b..42a0e5c802 100644 --- a/python/paddle/fluid/tests/unittests/test_version.py +++ b/python/paddle/fluid/tests/unittests/test_version.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import re diff --git a/python/paddle/fluid/tests/unittests/test_warpctc_op.py b/python/paddle/fluid/tests/unittests/test_warpctc_op.py index 9f1aaee472..5e3aa13546 100644 --- a/python/paddle/fluid/tests/unittests/test_warpctc_op.py +++ b/python/paddle/fluid/tests/unittests/test_warpctc_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import sys import unittest import numpy as np @@ -132,7 +134,7 @@ class CTCForward(object): for k in range(end - start): j = k + start if j & 1 == 1: - label_idx = j / 2 + label_idx = j // 2 label_val = labels_a_sequence[label_idx, 0] fv = self.log_add(forward_vars[i - 1, j], forward_vars[i - 1, j - 1]) diff --git a/python/paddle/fluid/tests/unittests/test_weight_normalization.py b/python/paddle/fluid/tests/unittests/test_weight_normalization.py index 436f9b9f86..e990d8b249 100644 --- a/python/paddle/fluid/tests/unittests/test_weight_normalization.py +++ b/python/paddle/fluid/tests/unittests/test_weight_normalization.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import numpy import collections diff --git a/python/paddle/fluid/tests/unittests/test_while_op.py b/python/paddle/fluid/tests/unittests/test_while_op.py index 790e6afe5f..b75373cf24 100644 --- a/python/paddle/fluid/tests/unittests/test_while_op.py +++ b/python/paddle/fluid/tests/unittests/test_while_op.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import unittest import paddle.fluid.layers as layers from paddle.fluid.executor import Executor diff --git a/python/paddle/fluid/tests/unittests/testsuite.py b/python/paddle/fluid/tests/unittests/testsuite.py index c6e176ca31..31ae25f02c 100644 --- a/python/paddle/fluid/tests/unittests/testsuite.py +++ b/python/paddle/fluid/tests/unittests/testsuite.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import numpy as np import paddle.fluid.core as core diff --git a/python/paddle/fluid/tests/unittests/transformer_model.py b/python/paddle/fluid/tests/unittests/transformer_model.py index 868a0248be..f0e74aff6b 100644 --- a/python/paddle/fluid/tests/unittests/transformer_model.py +++ b/python/paddle/fluid/tests/unittests/transformer_model.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from functools import partial import numpy as np diff --git a/python/paddle/fluid/trainer.py b/python/paddle/fluid/trainer.py index eed9b49ef4..d094647afe 100644 --- a/python/paddle/fluid/trainer.py +++ b/python/paddle/fluid/trainer.py @@ -12,10 +12,13 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import contextlib import os import errno import shutil +import six import time from . import core @@ -282,11 +285,12 @@ class Trainer(object): self._load_checkpoint() if param_path and os.path.isdir(param_path): - # load params from param_path into scope - io.load_persistables( - executor=exe, - dirname=param_path, - main_program=self.startup_program) + with self._prog_and_scope_guard(): + # load params from param_path into scope + io.load_persistables( + executor=exe, + dirname=param_path, + main_program=self.startup_program) def _transpile_nccl2_dist(self): # PADDLE_TRAINER_IPS @@ -618,7 +622,7 @@ def build_feed_var_list(program, feed_order): "The values of 'feed_order' should be a permutation of [0, len(feed_order))" ) sorted_pair_list = sorted( - list(feed_order.items()), key=lambda item: item[1]) + six.iteritems(feed_order), key=lambda item: item[1]) feed_var_list = [ program.global_block().var(pair[0]) for pair in sorted_pair_list ] @@ -1036,7 +1040,7 @@ def _save_trainer_args(dirname, trainer_id, trainer_args): cur_dir = _get_trainer_dir(dirname, trainer_id) - for name, value in list(trainer_args.items()): + for name, value in six.iteritems(trainer_args): args_file = os.path.join(cur_dir, name) with open(args_file, 'w') as f: f.write(str(value)) diff --git a/python/paddle/fluid/transpiler/__init__.py b/python/paddle/fluid/transpiler/__init__.py index a8622ad544..8429e2fd7c 100644 --- a/python/paddle/fluid/transpiler/__init__.py +++ b/python/paddle/fluid/transpiler/__init__.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from .distribute_transpiler import DistributeTranspiler, DistributeTranspilerConfig from .inference_transpiler import InferenceTranspiler from .memory_optimization_transpiler import memory_optimize, release_memory diff --git a/python/paddle/fluid/transpiler/details/__init__.py b/python/paddle/fluid/transpiler/details/__init__.py index 1bfab1f219..5e98266a76 100644 --- a/python/paddle/fluid/transpiler/details/__init__.py +++ b/python/paddle/fluid/transpiler/details/__init__.py @@ -12,5 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from .program_utils import * from .ufind import * diff --git a/python/paddle/fluid/transpiler/details/program_utils.py b/python/paddle/fluid/transpiler/details/program_utils.py index 76d10777f5..640dbf4bbe 100644 --- a/python/paddle/fluid/transpiler/details/program_utils.py +++ b/python/paddle/fluid/transpiler/details/program_utils.py @@ -12,12 +12,16 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + +import six + def delete_ops(block, ops): try: start = list(block.ops).index(ops[0]) end = list(block.ops).index(ops[-1]) - [block._remove_op(start) for _ in range(end - start + 1)] + [block._remove_op(start) for _ in six.moves.range(end - start + 1)] except Exception as e: raise e block.program._sync_with_cpp() diff --git a/python/paddle/fluid/transpiler/details/ufind.py b/python/paddle/fluid/transpiler/details/ufind.py index 0e30d0e3f9..aa63af7dcf 100644 --- a/python/paddle/fluid/transpiler/details/ufind.py +++ b/python/paddle/fluid/transpiler/details/ufind.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + class UnionFind(object): """ Union-find data structure. diff --git a/python/paddle/fluid/transpiler/distribute_transpiler.py b/python/paddle/fluid/transpiler/distribute_transpiler.py index ce4709f23b..540eb8c833 100644 --- a/python/paddle/fluid/transpiler/distribute_transpiler.py +++ b/python/paddle/fluid/transpiler/distribute_transpiler.py @@ -11,6 +11,8 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. + +from __future__ import print_function """ Steps to transpile trainer: 1. split variable to multiple blocks, aligned by product(dim[1:]) (width). @@ -31,6 +33,8 @@ Steps to transpile pserver: import math import random import numpy as np +import collections +import six from .ps_dispatcher import RoundRobin, HashName, PSDispatcher from .. import core, framework @@ -207,7 +211,17 @@ class DistributeTranspiler(object): ps_dispatcher = self.config.split_method(self.pserver_endpoints) self.has_distributed_lookup_table = self._has_distributed_lookup_table() + self.param_name_to_grad_name = dict() + for param_var, grad_var in self.params_grads: + self.param_name_to_grad_name[param_var.name] = grad_var.name + + # add distributed attrs to program + self.origin_program._is_distributed = True + self.origin_program._endpoints = self.pserver_endpoints + self.origin_program._is_chief = self.trainer_id == 0 + self.origin_program._distributed_lookup_table = self.table_name if self.table_name else None + # split and create vars, then put splited vars in dicts for later use. # step 1: split and create vars, then put splited vars in dicts for later use. self._init_splited_vars() @@ -220,39 +234,45 @@ class DistributeTranspiler(object): # fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1 # fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2 # shuffle the map will avoid the uneven distribution above - grad_var_mapping_items = list(self.grad_var_mapping.items()) + grad_var_mapping_items = list(six.iteritems(self.grad_var_mapping)) + if not self.config.slice_var_up: - random.seed(self.trainer_num) + random.seed(self.origin_program.random_seed) random.shuffle(grad_var_mapping_items) - for orig_varname, splited_vars in grad_var_mapping_items: + grad_name_to_send_dummy_out = dict() + for grad_varname, splited_vars in grad_var_mapping_items: eplist = ps_dispatcher.dispatch(splited_vars) if not self.config.slice_var_up: assert (len(splited_vars) == 1) + splited_grad_varname = grad_varname if len(splited_vars) == 1: - orig_varname = splited_vars[0].name + splited_grad_varname = splited_vars[0].name index = find_op_by_output_arg(program.global_block(), - orig_varname) + splited_grad_varname) elif len(splited_vars) > 1: - orig_var = program.global_block().vars[orig_varname] + orig_var = program.global_block().vars[splited_grad_varname] index = find_op_by_output_arg(program.global_block(), - orig_varname) + splited_grad_varname) self._insert_split_op(program, orig_var, index, splited_vars) index += 1 else: AssertionError("Can not insert the send op by original " - "variable name :", orig_varname) + "variable name :", splited_grad_varname) + dummy_output = program.global_block().create_var() + grad_name_to_send_dummy_out[grad_varname] = dummy_output program.global_block()._insert_op( index=index + 1, type="send", inputs={"X": splited_vars}, - outputs={}, + outputs={"Out": dummy_output}, attrs={ "epmap": eplist, - RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, + "sync_mode": not self.sync_mode, }) for _, var in enumerate(splited_vars): send_vars.append(var) @@ -264,7 +284,6 @@ class DistributeTranspiler(object): outputs={}, attrs={ "endpoints": pserver_endpoints, - "sync_mode": self.sync_mode, RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) @@ -280,19 +299,21 @@ class DistributeTranspiler(object): self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i]) # step4: Concat the parameters splits together after recv. - for varname, splited_var in list(self.param_var_mapping.items()): + for param_varname, splited_var in six.iteritems(self.param_var_mapping): eps = [] for var in splited_var: index = [v.name for v in recv_vars].index(var.name) eps.append(eplist[index]) - + grad_send_dummy_out = grad_name_to_send_dummy_out[ + self.param_name_to_grad_name[param_varname]] program.global_block().append_op( type="recv", - inputs={}, + inputs={"X": [grad_send_dummy_out]}, outputs={"Out": splited_var}, attrs={ "epmap": eps, - RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE + RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE, + "sync_mode": not self.sync_mode }) if self.sync_mode: @@ -305,10 +326,10 @@ class DistributeTranspiler(object): RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - for varname, splited_var in list(self.param_var_mapping.items()): + for param_varname, splited_var in six.iteritems(self.param_var_mapping): if len(splited_var) <= 1: continue - orig_param = program.global_block().vars[varname] + orig_param = program.global_block().vars[param_varname] program.global_block().append_op( type="concat", inputs={"X": splited_var}, @@ -355,7 +376,7 @@ class DistributeTranspiler(object): # FIXME(gongwb): delete not need ops. # note that: some parameter is not trainable and those ops can't be deleted. - for varname, splited_var in self.param_var_mapping.iteritems(): + for varname, splited_var in six.iteritems(self.param_var_mapping): # Get the eplist of recv vars eps = [] for var in splited_var: @@ -376,7 +397,7 @@ class DistributeTranspiler(object): op = startup_program.global_block().append_op( type="recv", - inputs={}, + inputs={"X": []}, outputs={"Out": splited_var}, attrs={ "epmap": eps, @@ -392,7 +413,7 @@ class DistributeTranspiler(object): RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE }) - for varname, splited_var in self.param_var_mapping.iteritems(): + for varname, splited_var in six.iteritems(self.param_var_mapping): #add concat ops to merge splited parameters received from parameter servers. if len(splited_var) <= 1: continue @@ -576,6 +597,8 @@ class DistributeTranspiler(object): checkpoint_block_id = self._create_checkpoint_save_block( pserver_program, table_opt_block.idx) + pserver_program._distributed_lookup_table = self.table_name + # NOTE: if has_distributed_lookup_table is False, then prefetch_block will # not be executed, so it's safe to use optimize_block to hold the place if self.has_distributed_lookup_table: @@ -602,6 +625,10 @@ class DistributeTranspiler(object): outputs={}, attrs=attrs) + # add distributed attrs + pserver_program._slice_vars_and_attrs = self._get_slice_vars_and_attrs( + endpoint) + pserver_program._sync_with_cpp() return pserver_program @@ -641,14 +668,14 @@ class DistributeTranspiler(object): # 1. create vars in pserver program to startup program pserver_vars = pserver_program.global_block().vars - created_var_map = dict() - for _, var in list(pserver_vars.items()): + created_var_map = collections.OrderedDict() + for _, var in six.iteritems(pserver_vars): tmpvar = s_prog.global_block()._clone_variable(var) created_var_map[var.name] = tmpvar # 2. rename op outputs for op in orig_s_prog.global_block().ops: - new_outputs = dict() + new_outputs = collections.OrderedDict() # do not append startup op if var is not on this pserver op_on_pserver = False # TODO(gongwb): remove this line. @@ -675,8 +702,31 @@ class DistributeTranspiler(object): inputs=new_inputs, outputs=new_outputs, attrs=op.all_attrs()) + + # add slice vars + s_prog._slice_vars_and_attrs = self._get_slice_vars_and_attrs(endpoint) + return s_prog + def _get_slice_vars_and_attrs(self, endpoint): + slice_vars_and_attrs = [] + block_suffix = "block" + for param in self.param_grad_ep_mapping[endpoint]["params"]: + orig_var_name, block_name, _ = self._get_varname_parts(param.name) + if not block_name: + continue + + block_idx = int(block_name.split(block_suffix)[1]) + orig_var = self.origin_program.global_block().vars[orig_var_name] + + skip_numel = 0 + slice_vars = self.param_var_mapping[orig_var_name] + for slice_var in slice_vars[:block_idx]: + skip_numel += reduce(lambda x, y: x * y, slice_var.shape) + slice_vars_and_attrs.append([orig_var, skip_numel, param]) + + return slice_vars_and_attrs + # ====================== private transpiler functions ===================== def _has_distributed_lookup_table(self): @@ -782,22 +832,24 @@ class DistributeTranspiler(object): self.config.min_block_size) assert (len(grad_blocks) == len(param_blocks)) - # origin_varname -> [splited_var] + # origin_param_name -> [splited_param_vars] self.param_var_mapping = self._create_vars_from_blocklist( self.origin_program, param_blocks) + # origin_grad_name -> [splited_grad_vars] self.grad_var_mapping = self._create_vars_from_blocklist( self.origin_program, grad_blocks, add_trainer_suffix=self.trainer_num > 1) - self.grad_param_mapping = dict() + # dict(grad_splited_var -> param_splited_var) + self.grad_param_mapping = collections.OrderedDict() for g, p in zip(grad_blocks, param_blocks): g_name, g_bid, _ = g.split(":") p_name, p_bid, _ = p.split(":") self.grad_param_mapping[self.grad_var_mapping[g_name][int(g_bid)]] = \ - self.param_var_mapping[p_name][int(p_bid)] + self.param_var_mapping[p_name][int(p_bid)] # create mapping of endpoint -> split var to create pserver side program - self.param_grad_ep_mapping = dict() + self.param_grad_ep_mapping = collections.OrderedDict() [ self.param_grad_ep_mapping.update({ ep: { @@ -915,7 +967,7 @@ class DistributeTranspiler(object): index=op_index + 2, type="send", inputs={'X': self.trainer_side_table_grad_list}, - outputs={}, + outputs={'Out': []}, attrs={ "sync_mode": True, "epmap": pserver_endpoints, @@ -1072,21 +1124,21 @@ class DistributeTranspiler(object): block_list (list[(varname, block_id, block_size)]): List of gradient blocks. add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True. Returns: - var_mapping (dict(varname->[new_varname_variable])):A dict mapping + var_mapping (collections.OrderedDict(varname->[new_varname_variable])):A dict mapping from original var name to each var split. """ # varname->[(block_id, current_block_size)] - block_map = dict() + block_map = collections.OrderedDict() - var_mapping = dict() + var_mapping = collections.OrderedDict() for block_str in block_list: varname, offset, size = block_str.split(":") if varname not in block_map: block_map[varname] = [] block_map[varname].append((int(offset), int(size))) - for varname, splited in list(block_map.items()): + for varname, splited in six.iteritems(block_map): orig_var = program.global_block().var(varname) if len(splited) == 1: if self.sync_mode and add_trainer_suffix: @@ -1107,7 +1159,7 @@ class DistributeTranspiler(object): for i, block in enumerate(splited): size = block[1] - rows = size / orig_dim1_flatten + rows = size // orig_dim1_flatten splited_shape = [rows] if len(orig_shape) >= 2: splited_shape.extend(orig_shape[1:]) @@ -1193,8 +1245,8 @@ class DistributeTranspiler(object): elif op_type == "momentum": if varkey == "Velocity": return param_shape - elif op_type == "": - if varkey == "Moment": + elif op_type == "rmsprop": + if varkey in ["Moment", "MeanSquare"]: return param_shape elif op_type == "sgd": pass @@ -1271,10 +1323,8 @@ class DistributeTranspiler(object): grad_to_block_id, origin_program, merged_var): program = optimize_block.program pserver_block = program.global_block() - new_inputs = dict() + new_inputs = collections.OrderedDict() - # update param/grad shape first, then other inputs like - # moment can use the updated shape def _get_param_block(opt_op): # param is already created on global program param_block = None @@ -1287,22 +1337,6 @@ class DistributeTranspiler(object): for key in opt_op.input_names: if key == "Grad": new_inputs[key] = merged_var - # For RMSProp optimizer - elif key == "Moment" or key == "MeanSquare": - param_block = _get_param_block(opt_op) - if not param_block: - return - moment_var = origin_program.global_block().vars[opt_op.input( - key)[0]] - tmpvar = pserver_block.create_var( - name=moment_var.name, - persistable=moment_var.persistable, - dtype=moment_var.dtype, - # change to use same shape as param - # TODO(typhoonzero): didn't append .block in the var name, - # may affect checkpoint saving? Need to verify. - shape=param_block.shape) - new_inputs[key] = tmpvar elif key == "Param": param_block = _get_param_block(opt_op) if not param_block: @@ -1330,7 +1364,7 @@ class DistributeTranspiler(object): for key in opt_op.input_names: new_shape = None - if key in ["Param", "Grad", "LearningRate", "Moment", "MeanSquare"]: + if key in ["Param", "Grad", "LearningRate"]: continue var = self.origin_program.global_block().vars[opt_op.input(key)[0]] # update accumulator variable shape @@ -1357,9 +1391,7 @@ class DistributeTranspiler(object): def _is_splited_grad_var(self, var, var_dict): grad_block = None - # TODO(minqiyang): replace these items() with six.iteritems() to - # improve memory - for _, g in list(var_dict.items()): + for _, g in six.iteritems(var_dict): if self._orig_varname(g.name) == self._orig_varname(var.name): if g.name.find(".trainer_") == -1: grad_block = g @@ -1369,7 +1401,7 @@ class DistributeTranspiler(object): def _clone_lr_op(self, program, block, op): inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, op) - for key, varlist in list(inputs.items()): + for key, varlist in six.iteritems(inputs): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: @@ -1378,7 +1410,7 @@ class DistributeTranspiler(object): outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, op) - for key, varlist in list(outputs.items()): + for key, varlist in six.iteritems(outputs): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: @@ -1393,7 +1425,7 @@ class DistributeTranspiler(object): # Append the ops for parameters that do not need to be optimized/updated inputs = self._get_input_map_from_op( self.origin_program.global_block().vars, opt_op) - for key, varlist in list(inputs.items()): + for key, varlist in six.iteritems(inputs): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: @@ -1412,7 +1444,7 @@ class DistributeTranspiler(object): outputs = self._get_output_map_from_op( self.origin_program.global_block().vars, opt_op) - for key, varlist in list(outputs.items()): + for key, varlist in six.iteritems(outputs): if not isinstance(varlist, list): varlist = [varlist] for var in varlist: @@ -1470,7 +1502,7 @@ class DistributeTranspiler(object): def _get_input_map_from_op(self, varmap, op): """Returns a dict from op input name to the vars in varmap.""" - iomap = dict() + iomap = collections.OrderedDict() for key in op.input_names: vars = [] for varname in op.input(key): @@ -1483,7 +1515,7 @@ class DistributeTranspiler(object): def _get_output_map_from_op(self, varmap, op): """Returns a dict from op output name to the vars in varmap.""" - iomap = dict() + iomap = collections.OrderedDict() for key in op.output_names: vars = [] for varname in op.output(key): diff --git a/python/paddle/fluid/transpiler/inference_transpiler.py b/python/paddle/fluid/transpiler/inference_transpiler.py index 87f20bbccf..f79fcb24bb 100644 --- a/python/paddle/fluid/transpiler/inference_transpiler.py +++ b/python/paddle/fluid/transpiler/inference_transpiler.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import os import numpy as np from .. import core @@ -57,8 +59,12 @@ class InferenceTranspiler(object): scope = global_scope() if not isinstance(scope, core.Scope): raise TypeError("scope should be as Scope type or None") - self._fuse_batch_norm(program, place, scope) - self._fuse_relu_mkldnn(program) + use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False)) + if use_mkldnn: + self._fuse_relu_mkldnn(program) + self._fuse_conv_bias_mkldnn(program) + else: + self._fuse_batch_norm(program, place, scope) def _fuse_relu_mkldnn(self, program): ''' @@ -80,10 +86,6 @@ class InferenceTranspiler(object): :param program: program to transpile :type program: Program ''' - use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False)) - if not use_mkldnn: - return - self.block = program.block(0) i = 0 @@ -104,6 +106,69 @@ class InferenceTranspiler(object): # And a better solution will be considered later. program = program.clone() + def _fuse_conv_bias_mkldnn(self, program): + ''' + Transpile the program by fused convolution and elementwise_add. + + Replace conv2d and elementwise_add ops with a new conv2d op + based on an old conv2d op and the :math:`Bias` taken from + elementwise_add. + + For input :math:`X`: + + - Conv process: :math:`X = input * W` + - Elementwise_add process: :math` X = X + bias` + + After fuse into one operation: + + .. math:: + + X = input * W + bias + + The operator transformation is: + + - before: + + - conv->elementwise_add->any_other_op + + - after: + + - conv->any_other_op + + The transpile stages are: + + 1. Extract bias and output variables from elementwise_add. + 2. Extract Input, Weight and attributes from conv op. + 3. Create a new convolution op based on extracted params. + 4. Remove old conv op. + 5. Remove elementwise_add. + 5. Remove unused variables. + + Args: + program (Program): program to transpile + + ''' + self.block = program.block(0) + + i = 0 + while i < len(self.block.ops) - 2: + current_op = self.block.ops[i] + next_op = self.block.ops[i + 1] + # conv2d with bias + if current_op.type in ['conv2d'] and \ + next_op.type in ['elementwise_add']: + self._fuse_conv_bias(i, current_op, next_op) + self.block._remove_op(i + 1) # Remove old conv + self.block._remove_op(i + 1) # Remove elementwise_add + i = i + 1 + i = i + 1 + + self._remove_unused_var() + # TODO(luotao): use clone() method to flush the program.desc in force, + # since some large program.desc will not be flushed immediately. + # And a better solution will be considered later. + program = program.clone() + def _fuse_batch_norm(self, program, place, scope): ''' Transpile the program by fused batch normalization. @@ -183,7 +248,6 @@ class InferenceTranspiler(object): self.block._remove_op(i + 2) i = i + 1 i = i + 1 - self._adjust_input() self._remove_unused_var() # TODO(luotao): use clone() method to flush the program.desc in force, @@ -286,6 +350,33 @@ class InferenceTranspiler(object): # collect the renamed input self.input_map[bn_op.output("Y")[0]] = bias_op.output("Out")[0] + def _fuse_conv_bias(self, index, conv_op, elementwise_add_op): + ''' + fuse the conv op with elementwise_add + + :param index: index of the conv_op in ops list + :type index: Int + :param conv_op: convolution operator + :type conv_op: Operator + :param elementwise_add_op: convolution's bias operator + :type elementwise_add_op: Operator + ''' + + bias_var = self.block.var(elementwise_add_op.input("Y")[0]) + out_var = self.block.var(elementwise_add_op.output("Out")[0]) + filter_var = self.block.var(conv_op.input("Filter")[0]) + in_var = self.block.var(conv_op.input("Input")[0]) + attrs = {name: conv_op.attr(name) for name in conv_op.attr_names} + + self.block._insert_op( + index, + type="conv2d", + inputs={"Input": in_var, + "Filter": filter_var, + "Bias": bias_var}, + outputs={"Output": out_var}, + attrs=attrs) + def _adjust_input(self): for i in range(len(self.block.ops)): current_op = self.block.ops[i] diff --git a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py index 20ba7ed2b0..3e58e125de 100644 --- a/python/paddle/fluid/transpiler/memory_optimization_transpiler.py +++ b/python/paddle/fluid/transpiler/memory_optimization_transpiler.py @@ -12,8 +12,11 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + from collections import defaultdict from .. import core +from ... import compat as cpt from ..framework import Program, default_main_program, Parameter from ..backward import _rename_arg_ from functools import reduce @@ -125,15 +128,15 @@ class ControlFlowGraph(object): def _has_var(self, block_desc, var_name, is_forward): if is_forward: - return block_desc.has_var(str(var_name)) + return block_desc.has_var(cpt.to_bytes(var_name)) else: - return block_desc.has_var_recursive(str(var_name)) + return block_desc.has_var_recursive(cpt.to_bytes(var_name)) def _find_var(self, block_desc, var_name, is_forward): if is_forward: - return block_desc.find_var(str(var_name)) + return block_desc.find_var(cpt.to_bytes(var_name)) else: - return block_desc.find_var_recursive(str(var_name)) + return block_desc.find_var_recursive(cpt.to_bytes(var_name)) def _check_var_validity(self, block_desc, x, is_forward): if str(x) == "@EMPTY@": @@ -258,7 +261,7 @@ class ControlFlowGraph(object): # Rename the var to the cache var already with # memory allocated in order to reuse the memory. _rename_arg_(self._ops, x, cache_var, begin_idx=i) - self._program.block(block_desc.id).var(str( + self._program.block(block_desc.id).var(cpt.to_text( x)).desc = self._find_var(block_desc, cache_var, is_forward) self._update_graph(x, cache_var, begin_idx=i) diff --git a/python/paddle/fluid/transpiler/ps_dispatcher.py b/python/paddle/fluid/transpiler/ps_dispatcher.py index dcffadd531..6a6d14a69b 100644 --- a/python/paddle/fluid/transpiler/ps_dispatcher.py +++ b/python/paddle/fluid/transpiler/ps_dispatcher.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + class PSDispatcher(object): """ diff --git a/python/paddle/fluid/unique_name.py b/python/paddle/fluid/unique_name.py index b125eba4f8..b9957a699e 100644 --- a/python/paddle/fluid/unique_name.py +++ b/python/paddle/fluid/unique_name.py @@ -12,6 +12,8 @@ # See the License for the specific language governing permissions and # limitations under the License. +from __future__ import print_function + import collections import contextlib import six diff --git a/python/paddle/reader/decorator.py b/python/paddle/reader/decorator.py index ce410e61b9..6d7ac876fd 100644 --- a/python/paddle/reader/decorator.py +++ b/python/paddle/reader/decorator.py @@ -27,6 +27,7 @@ from six.moves import zip import itertools import random import zlib +import paddle.compat as cpt def map_readers(func, *readers): @@ -390,9 +391,9 @@ class PipeReader: buff = self.process.stdout.read(self.bufsize) if buff: if self.file_type == "gzip": - decomp_buff = self.dec.decompress(buff) + decomp_buff = cpt.to_text(self.dec.decompress(buff)) elif self.file_type == "plain": - decomp_buff = buff + decomp_buff = cpt.to_text(buff) else: raise TypeError("file_type %s is not allowed" % self.file_type) diff --git a/python/paddle/reader/tests/creator_test.py b/python/paddle/reader/tests/creator_test.py index c4238c12a7..d7107610a5 100644 --- a/python/paddle/reader/tests/creator_test.py +++ b/python/paddle/reader/tests/creator_test.py @@ -29,6 +29,7 @@ import os import unittest import numpy as np import paddle.reader.creator +import six class TestNumpyArray(unittest.TestCase): @@ -37,7 +38,7 @@ class TestNumpyArray(unittest.TestCase): x = np.array(l, np.int32) reader = paddle.reader.creator.np_array(x) for idx, e in enumerate(reader()): - self.assertItemsEqual(e, l[idx]) + six.assertCountEqual(self, e, l[idx]) class TestTextFile(unittest.TestCase): diff --git a/python/paddle/utils/dump_config.py b/python/paddle/utils/dump_config.py index d27af7f762..6a96a0a78f 100644 --- a/python/paddle/utils/dump_config.py +++ b/python/paddle/utils/dump_config.py @@ -37,9 +37,9 @@ if __name__ == '__main__': assert isinstance(conf, TrainerConfig_pb2.TrainerConfig) if whole_conf: - print conf + print(conf) else: if binary: sys.stdout.write(conf.model_config.SerializeToString()) else: - print conf.model_config + print(conf.model_config) diff --git a/python/paddle/utils/image_multiproc.py b/python/paddle/utils/image_multiproc.py index 3e3e519f76..d1bbda3fd3 100644 --- a/python/paddle/utils/image_multiproc.py +++ b/python/paddle/utils/image_multiproc.py @@ -15,7 +15,8 @@ import os, sys import numpy as np from PIL import Image -from cStringIO import StringIO +import six +from six.moves import cStringIO as StringIO import multiprocessing import functools import itertools @@ -187,7 +188,8 @@ class PILTransformer(ImageTransformer): return self.transform(im) -def job(is_img_string, transformer, (data, label)): +def job(is_img_string, transformer, data_label_pack): + (data, label) = data_label_pack if is_img_string: return transformer.transform_from_string(data), label else: @@ -208,7 +210,7 @@ class MultiProcessImageTransformer(object): """ Processing image with multi-process. If it is used in PyDataProvider, the simple usage for CNN is as follows: - + .. code-block:: python def hool(settings, is_train, **kwargs): @@ -229,7 +231,7 @@ class MultiProcessImageTransformer(object): @provider(init_hook=hook, pool_size=20480) def process(settings, file_list): with open(file_list, 'r') as fdata: - for line in fdata: + for line in fdata: data_dic = np.load(line.strip()) # load the data batch pickled by Pickle. data = data_dic['data'] labels = data_dic['label'] @@ -249,10 +251,10 @@ class MultiProcessImageTransformer(object): :type channel_swap: tuple or list :param mean: the mean values of image, per-channel mean or element-wise mean. :type mean: array, The dimension is 1 for per-channel mean. - The dimension is 3 for element-wise mean. + The dimension is 3 for element-wise mean. :param is_train: training peroid or testing peroid. :type is_train: bool. - :param is_color: the image is color or gray. + :param is_color: the image is color or gray. :type is_color: bool. :param is_img_string: The input can be the file name of image or image string. :type is_img_string: bool. @@ -273,4 +275,4 @@ class MultiProcessImageTransformer(object): def run(self, data, label): fun = functools.partial(job, self.is_img_string, self.transformer) return self.pool.imap_unordered( - fun, itertools.izip(data, label), chunksize=100 * self.procnum) + fun, six.moves.zip(data, label), chunksize=100 * self.procnum) diff --git a/python/paddle/utils/image_util.py b/python/paddle/utils/image_util.py index d3d79b1440..a8092349cd 100644 --- a/python/paddle/utils/image_util.py +++ b/python/paddle/utils/image_util.py @@ -14,7 +14,7 @@ import numpy as np from PIL import Image -from cStringIO import StringIO +from six.moves import cStringIO as StringIO def resize_image(img, target_size): @@ -34,7 +34,7 @@ def flip(im): """ Return the flipped image. Flip an image along the horizontal direction. - im: input image, (H x W x K) ndarrays + im: input image, (H x W x K) ndarrays """ if len(im.shape) == 3: return im[:, :, ::-1] @@ -132,7 +132,7 @@ def load_meta(meta_path, mean_img_size, crop_size, color=True): def load_image(img_path, is_color=True): """ - Load image and return. + Load image and return. img_path: image path. is_color: is color image or not. """ @@ -205,7 +205,7 @@ class ImageTransformer: def set_mean(self, mean): if mean is not None: - # mean value, may be one value per channel + # mean value, may be one value per channel if mean.ndim == 1: mean = mean[:, np.newaxis, np.newaxis] else: diff --git a/python/paddle/utils/make_model_diagram.py b/python/paddle/utils/make_model_diagram.py index 40f99075de..52759d3ad2 100644 --- a/python/paddle/utils/make_model_diagram.py +++ b/python/paddle/utils/make_model_diagram.py @@ -15,6 +15,9 @@ # Generate dot diagram file for the given paddle model config # The generated file can be viewed using Graphviz (http://graphviz.org) +from __future__ import print_function + +import six import sys import traceback @@ -61,9 +64,9 @@ def make_diagram_from_proto(model_config, dot_file): name2id[mem.link_name]) return s - print >> f, 'digraph graphname {' - print >> f, 'node [width=0.375,height=0.25];' - for i in xrange(len(model_config.layers)): + print('digraph graphname {', file=f) + print('node [width=0.375,height=0.25];', file=f) + for i in six.moves.xrange(len(model_config.layers)): l = model_config.layers[i] name2id[l.name] = i @@ -71,12 +74,12 @@ def make_diagram_from_proto(model_config, dot_file): for sub_model in model_config.sub_models: if sub_model.name == 'root': continue - print >> f, 'subgraph cluster_%s {' % i - print >> f, 'style=dashed;' + print('subgraph cluster_%s {' % i, file=f) + print('style=dashed;', file=f) label = '%s ' % sub_model.name if sub_model.reversed: label += '<==' - print >> f, 'label = "%s";' % label + print('label = "%s";' % label, file=f) i += 1 submodel_layers.add(sub_model.name) for layer_name in sub_model.layer_names: @@ -84,37 +87,41 @@ def make_diagram_from_proto(model_config, dot_file): lid = name2id[layer_name] layer_config = model_config.layers[lid] label = make_layer_label(layer_config) - print >> f, 'l%s [label="%s", shape=box];' % (lid, label) - print >> f, '}' + print('l%s [label="%s", shape=box];' % (lid, label), file=f) + print('}', file=f) - for i in xrange(len(model_config.layers)): + for i in six.moves.xrange(len(model_config.layers)): l = model_config.layers[i] if l.name not in submodel_layers: label = make_layer_label(l) - print >> f, 'l%s [label="%s", shape=box];' % (i, label) + print('l%s [label="%s", shape=box];' % (i, label), file=f) for sub_model in model_config.sub_models: if sub_model.name == 'root': continue for link in sub_model.in_links: - print >> f, make_link(link) + print(make_link(link), file=f) for link in sub_model.out_links: - print >> f, make_link(link) + print(make_link(link), file=f) for mem in sub_model.memories: - print >> f, make_mem(mem) + print(make_mem(mem), file=f) - for i in xrange(len(model_config.layers)): + for i in six.moves.xrange(len(model_config.layers)): for l in model_config.layers[i].inputs: - print >> f, 'l%s -> l%s [label="%s"];' % ( - name2id[l.input_layer_name], i, l.input_parameter_name) + print( + 'l%s -> l%s [label="%s"];' % (name2id[l.input_layer_name], i, + l.input_parameter_name), + file=f) - print >> f, '}' + print('}', file=f) f.close() def usage(): - print >> sys.stderr, ("Usage: python show_model_diagram.py" + - " CONFIG_FILE DOT_FILE [config_str]") + print( + ("Usage: python show_model_diagram.py" + + " CONFIG_FILE DOT_FILE [config_str]"), + file=sys.stderr) exit(1) diff --git a/python/paddle/utils/merge_model.py b/python/paddle/utils/merge_model.py index 2b10020772..b74649e936 100644 --- a/python/paddle/utils/merge_model.py +++ b/python/paddle/utils/merge_model.py @@ -70,4 +70,4 @@ def merge_v2_model(net, param_file, output_file): for pname in param_names: params.serialize(pname, f) - print 'Generate %s success!' % (output_file) + print('Generate %s success!' % (output_file)) diff --git a/python/paddle/utils/plotcurve.py b/python/paddle/utils/plotcurve.py index 27bd8157d3..a95e5497e2 100644 --- a/python/paddle/utils/plotcurve.py +++ b/python/paddle/utils/plotcurve.py @@ -44,6 +44,7 @@ To use this script to generate plot for AvgCost, error: python plotcurve.py -i paddle.INFO -o figure.png AvgCost error """ +import six import sys import matplotlib # the following line is added immediately after import matplotlib @@ -91,7 +92,7 @@ def plot_paddle_curve(keys, inputfile, outputfile, format='png', sys.stderr.write("No data to plot. Exiting!\n") return m = len(keys) + 1 - for i in xrange(1, m): + for i in six.moves.xrange(1, m): pyplot.plot( x[:, 0], x[:, i], diff --git a/python/paddle/utils/predefined_net.py b/python/paddle/utils/predefined_net.py index fa05f981f2..2801f4877c 100644 --- a/python/paddle/utils/predefined_net.py +++ b/python/paddle/utils/predefined_net.py @@ -13,6 +13,7 @@ # limitations under the License. import numpy as np +import six import os from paddle.trainer.config_parser import * from paddle.utils.preprocess_img import \ @@ -112,7 +113,7 @@ def simple_conv_net(data_conf, is_color=False): num_classes: num of classes. is_color: whether the input images are color. """ - for k, v in data_conf.iteritems(): + for k, v in six.iteritems(data_conf): globals()[k] = v data_input, label_input, num_image_channels = \ image_data_layers(image_size, num_classes, is_color, is_predict) @@ -340,7 +341,7 @@ def small_vgg(data_conf, is_predict=False): num_classes: num of classes. is_color: whether the input images are color. """ - for k, v in data_conf.iteritems(): + for k, v in six.iteritems(data_conf): globals()[k] = v vgg_conv_net(image_size, num_classes, num_layers=[2, 2, 3, 3], diff --git a/python/paddle/utils/preprocess_img.py b/python/paddle/utils/preprocess_img.py index 975f1e9ede..a322f7b769 100644 --- a/python/paddle/utils/preprocess_img.py +++ b/python/paddle/utils/preprocess_img.py @@ -17,9 +17,9 @@ import os import random import numpy as np import PIL.Image as Image -import StringIO -import preprocess_util -from image_util import crop_img +from six.moves import cStringIO as StringIO +from . import preprocess_util +from .image_util import crop_img def resize_image(img, target_size): @@ -52,7 +52,7 @@ class DiskImage: def read_image(self): if self.img is None: - print "reading: " + self.path + print("reading: " + self.path) image = resize_image(Image.open(self.path), self.target_size) self.img = image @@ -69,7 +69,7 @@ class DiskImage: convert the image into the paddle batch format. """ self.read_image() - output = StringIO.StringIO() + output = StringIO() self.img.save(output, "jpeg") contents = output.getvalue() return contents @@ -127,7 +127,7 @@ class ImageClassificationDatasetCreater(preprocess_util.DatasetCreater): image_path = items[0] label_name = items[1] if not label_name in label_set: - label_set[label_name] = len(label_set.keys()) + label_set[label_name] = len(list(label_set.keys())) img = DiskImage(path=image_path, target_size=self.target_size) label = preprocess_util.Lablel( label=label_set[label_name], name=label_name) @@ -144,7 +144,7 @@ class ImageClassificationDatasetCreater(preprocess_util.DatasetCreater): return create_dataset_from_list(path) label_set = preprocess_util.get_label_set_from_dir(path) data = [] - for l_name in label_set.keys(): + for l_name in list(label_set.keys()): image_paths = preprocess_util.list_images( os.path.join(path, l_name)) for p in image_paths: diff --git a/python/paddle/utils/preprocess_util.py b/python/paddle/utils/preprocess_util.py index 1d17a48824..05b2067d01 100644 --- a/python/paddle/utils/preprocess_util.py +++ b/python/paddle/utils/preprocess_util.py @@ -14,7 +14,7 @@ import os import math -import cPickle as pickle +import six.moves.cPickle as pickle import random import collections @@ -169,7 +169,7 @@ class Dataset: random.shuffle(keyvalue_indices[k]) num_data_per_key_batch = \ - math.ceil(num_per_batch / float(len(keyvalue_indices.keys()))) + math.ceil(num_per_batch / float(len(list(keyvalue_indices.keys())))) if num_data_per_key_batch < 2: raise Exception("The number of data in a batch is too small") @@ -182,8 +182,8 @@ class Dataset: end_idx = int( min(begin_idx + num_data_per_key_batch, len(keyvalue_indices[k]))) - print "begin_idx, end_idx" - print begin_idx, end_idx + print("begin_idx, end_idx") + print(begin_idx, end_idx) for idx in range(begin_idx, end_idx): permuted_data.append(self.data[keyvalue_indices[k][idx]]) keyvalue_readpointer[k] = end_idx @@ -357,6 +357,6 @@ class DatasetCreater(object): data_batcher.create_batches_and_list( self.output_path, self.train_list_name, self.test_list_name, self.label_set_name) - self.num_classes = len(train_label_set.keys()) + self.num_classes = len(list(train_label_set.keys())) self.create_meta_file(train_data) return out_path diff --git a/python/paddle/utils/show_pb.py b/python/paddle/utils/show_pb.py index 20614826d1..da7a71a665 100644 --- a/python/paddle/utils/show_pb.py +++ b/python/paddle/utils/show_pb.py @@ -15,6 +15,8 @@ Show the content of proto buffer data file of PADDLE """ +from __future__ import print_function + import os import sys from google.protobuf.internal.decoder import _DecodeVarint @@ -39,7 +41,7 @@ def read_proto(file, message): def usage(): - print >> sys.stderr, "Usage: python show_pb.py PROTO_DATA_FILE" + print("Usage: python show_pb.py PROTO_DATA_FILE", file=sys.stderr) exit(1) @@ -50,8 +52,8 @@ if __name__ == '__main__': f = open(sys.argv[1]) header = DataFormat.DataHeader() read_proto(f, header) - print header + print(header) sample = DataFormat.DataSample() while read_proto(f, sample): - print sample + print(sample) diff --git a/python/paddle/utils/torch2paddle.py b/python/paddle/utils/torch2paddle.py index 91490111a1..398d3aa4e0 100644 --- a/python/paddle/utils/torch2paddle.py +++ b/python/paddle/utils/torch2paddle.py @@ -24,7 +24,7 @@ import sys import struct import numpy as np import torchfile -import cPickle as pickle +import six.moves.cPickle as pickle import argparse @@ -48,7 +48,7 @@ def save_net_parameters(layers, params, output_path): biases = params[i * 2 + 1] weight_file = os.path.join(output_path, '_%s.w0' % layers[i]) biases_file = os.path.join(output_path, '_%s.wbias' % layers[i]) - print "Saving for layer %s." % layers[i] + print("Saving for layer %s." % layers[i]) save_layer_parameters(weight_file, [weight]) save_layer_parameters(biases_file, biases) diff --git a/python/setup.py.in b/python/setup.py.in index 4a6cddbbea..786c9f2e39 100644 --- a/python/setup.py.in +++ b/python/setup.py.in @@ -159,18 +159,20 @@ if '${WITH_MKL}' == 'ON': shutil.copy('${MKLML_LIB}', libs_path) shutil.copy('${MKLML_IOMP_LIB}', libs_path) package_data['paddle.libs']+=['libmklml_intel.so','libiomp5.so'] -if '${WITH_MKLDNN}' == 'ON': - # TODO(typhoonzero): use install_name_tool to patch mkl libs once - # we can support mkl on mac. - # - # change rpath of libmkldnn.so.0, add $ORIGIN/ to it. - # The reason is that all thirdparty libraries in the same directory, - # thus, libmkldnn.so.0 will find libmklml_intel.so and libiomp5.so. - command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}" - if os.system(command) != 0: - raise Exception("patch libmkldnn.so failed, command: %s" % command) - package_data['paddle.libs']+=['libmkldnn.so.0'] - shutil.copy('${MKLDNN_SHARED_LIB}', libs_path) +if '${CMAKE_BUILD_TYPE}' == 'Release': + # only change rpath in Release mode. + if '${WITH_MKLDNN}' == 'ON': + # TODO(typhoonzero): use install_name_tool to patch mkl libs once + # we can support mkl on mac. + # + # change rpath of libmkldnn.so.0, add $ORIGIN/ to it. + # The reason is that all thirdparty libraries in the same directory, + # thus, libmkldnn.so.0 will find libmklml_intel.so and libiomp5.so. + command = "patchelf --set-rpath '$ORIGIN/' ${MKLDNN_SHARED_LIB}" + if os.system(command) != 0: + raise Exception("patch libmkldnn.so failed, command: %s" % command) + package_data['paddle.libs']+=['libmkldnn.so.0'] + shutil.copy('${MKLDNN_SHARED_LIB}', libs_path) # remove unused paddle/libs/__init__.py os.remove(libs_path+'/__init__.py') package_dir['paddle.libs']=libs_path @@ -179,20 +181,22 @@ package_dir['paddle.libs']=libs_path # The reason is that libwarpctc.so, libiomp5.so etc are in paddle.libs, and # core.so is in paddle.fluid, thus paddle/fluid/../libs will pointer to above libraries. # This operation will fix https://github.com/PaddlePaddle/Paddle/issues/3213 -if "@APPLE@" == "1": - command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so" -else: - command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so" -if os.system(command) != 0: - raise Exception("patch core.so failed, command: %s" % command) -if '${WITH_FLUID_ONLY}'== 'OFF': - # change rpath of _swig_paddle.so. +if '${CMAKE_BUILD_TYPE}' == 'Release': + # only change rpath in Release mode, since in Debug mode, core.so is too large to be changed. if "@APPLE@" == "1": - command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so" + command = "install_name_tool -id \"@loader_path/../libs/\" ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so" else: - command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so" + command = "patchelf --set-rpath '$ORIGIN/../libs/' ${PADDLE_BINARY_DIR}/python/paddle/fluid/core.so" if os.system(command) != 0: - raise Exception("patch _swig_paddle.so failed, command: %s" % command) + raise Exception("patch core.so failed, command: %s" % command) + if '${WITH_FLUID_ONLY}'== 'OFF': + # change rpath of _swig_paddle.so. + if "@APPLE@" == "1": + command = "install_name_tool -id \"@loader_path/../paddle/libs/\" ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so" + else: + command = "patchelf --set-rpath '$ORIGIN/../paddle/libs/' ${PADDLE_BINARY_DIR}/python/py_paddle/_swig_paddle.so" + if os.system(command) != 0: + raise Exception("patch _swig_paddle.so failed, command: %s" % command) setup(name='${PACKAGE_NAME}', version='${PADDLE_VERSION}', diff --git a/tools/manylinux1/build_scripts/build.sh b/tools/manylinux1/build_scripts/build.sh index 93591fa9dd..eb4b477dcb 100644 --- a/tools/manylinux1/build_scripts/build.sh +++ b/tools/manylinux1/build_scripts/build.sh @@ -28,7 +28,7 @@ AUTOCONF_HASH=954bd69b391edc12d6a4a51a2dd1476543da5c6bbf05a95b59dc0dd6fd4c2969 PYTHON_COMPILE_DEPS="zlib-devel bzip2-devel ncurses-devel sqlite-devel readline-devel tk-devel gdbm-devel db4-devel libpcap-devel xz-devel" # Libraries that are allowed as part of the manylinux1 profile -MANYLINUX1_DEPS="glibc-devel libstdc++-devel glib2-devel libX11-devel libXext-devel libXrender-devel mesa-libGL-devel libICE-devel libSM-devel ncurses-devel" +MANYLINUX1_DEPS="glibc-devel libstdc++-devel glib2-devel libX11-devel libXext-devel libXrender-devel mesa-libGL-devel libICE-devel libSM-devel ncurses-devel freetype-devel libpng-devel" # Get build utilities MY_DIR=$(dirname "${BASH_SOURCE[0]}") @@ -105,7 +105,7 @@ curl-config --features rm -rf /usr/local/ssl # Install patchelf (latest with unreleased bug fixes) -curl -sLO https://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz +curl -sLO http://nipy.bic.berkeley.edu/manylinux/patchelf-0.9njs2.tar.gz check_sha256sum patchelf-0.9njs2.tar.gz $PATCHELF_HASH tar -xzf patchelf-0.9njs2.tar.gz (cd patchelf-0.9njs2 && ./configure && make && make install) diff --git a/tools/manylinux1/build_scripts/install_nccl2.sh b/tools/manylinux1/build_scripts/install_nccl2.sh index 282c5c290d..43a99d8287 100644 --- a/tools/manylinux1/build_scripts/install_nccl2.sh +++ b/tools/manylinux1/build_scripts/install_nccl2.sh @@ -21,5 +21,5 @@ for sub_deb in $DEBS; do ar x $sub_deb && tar xf data.tar.xz done mv -f usr/include/nccl.h /usr/local/include/ -mv -f usr/lib/libnccl* /usr/local/lib/ +mv -f usr/lib/x86_64-linux-gnu/libnccl* /usr/local/lib/ rm -rf $DIR diff --git a/tools/test_runner.py b/tools/test_runner.py index 2d6a3cf8a9..9b9f165e73 100644 --- a/tools/test_runner.py +++ b/tools/test_runner.py @@ -13,6 +13,7 @@ # limitations under the License. from __future__ import print_function + import unittest import os import sys