Merge branch 'develop' of github.com:PaddlePaddle/Paddle into feature/recurrent_op_backward_fix

update-doc-pybind
superjom 8 years ago
commit b545b5b86b

@ -106,22 +106,22 @@ function(merge_static_libs TARGET_NAME)
endforeach()
list(REMOVE_DUPLICATES libs_deps)
if(APPLE) # Use OSX's libtool to merge archives
# To produce a library we need at least one source file.
# It is created by add_custom_command below and will helps
# also help to track dependencies.
set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
# To produce a library we need at least one source file.
# It is created by add_custom_command below and will helps
# also help to track dependencies.
set(target_SRCS ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}_dummy.c)
if(APPLE) # Use OSX's libtool to merge archives
# Make the generated dummy source file depended on all static input
# libs. If input lib changes,the source file is touched
# which causes the desired effect (relink).
add_custom_command(OUTPUT ${dummyfile}
COMMAND ${CMAKE_COMMAND} -E touch ${dummyfile}
add_custom_command(OUTPUT ${target_SRCS}
COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
DEPENDS ${libs})
# Generate dummy staic lib
file(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
add_library(${TARGET_NAME} STATIC ${dummyfile})
file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";")
add_library(${TARGET_NAME} STATIC ${target_SRCS})
target_link_libraries(${TARGET_NAME} ${libs_deps})
foreach(lib ${libs})
@ -130,11 +130,14 @@ function(merge_static_libs TARGET_NAME)
endforeach()
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles})
COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles}
)
else() # general UNIX: use "ar" to extract objects and re-add to a common lib
set(target_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}.dir)
foreach(lib ${libs})
set(objlistfile ${lib}.objlist) # list of objects in the input library
set(objdir ${lib}.objdir)
set(objlistfile ${target_DIR}/${lib}.objlist) # list of objects in the input library
set(objdir ${target_DIR}/${lib}.objdir)
add_custom_command(OUTPUT ${objdir}
COMMAND ${CMAKE_COMMAND} -E make_directory ${objdir}
@ -142,31 +145,32 @@ function(merge_static_libs TARGET_NAME)
add_custom_command(OUTPUT ${objlistfile}
COMMAND ${CMAKE_AR} -x "$<TARGET_FILE:${lib}>"
COMMAND ${CMAKE_AR} -t "$<TARGET_FILE:${lib}>" > ../${objlistfile}
COMMAND ${CMAKE_AR} -t "$<TARGET_FILE:${lib}>" > ${objlistfile}
DEPENDS ${lib} ${objdir}
WORKING_DIRECTORY ${objdir})
# Empty dummy source file that goes into merged library
set(mergebase ${lib}.mergebase.c)
add_custom_command(OUTPUT ${mergebase}
COMMAND ${CMAKE_COMMAND} -E touch ${mergebase}
DEPENDS ${objlistfile})
list(APPEND mergebases "${mergebase}")
list(APPEND target_OBJS "${objlistfile}")
endforeach()
add_library(${TARGET_NAME} STATIC ${mergebases})
# Make the generated dummy source file depended on all static input
# libs. If input lib changes,the source file is touched
# which causes the desired effect (relink).
add_custom_command(OUTPUT ${target_SRCS}
COMMAND ${CMAKE_COMMAND} -E touch ${target_SRCS}
DEPENDS ${libs} ${target_OBJS})
# Generate dummy staic lib
file(WRITE ${target_SRCS} "const char *dummy = \"${target_SRCS}\";")
add_library(${TARGET_NAME} STATIC ${target_SRCS})
target_link_libraries(${TARGET_NAME} ${libs_deps})
# Get the file name of the generated library
set(outlibfile "$<TARGET_FILE:${TARGET_NAME}>")
set(target_LIBNAME "$<TARGET_FILE:${TARGET_NAME}>")
foreach(lib ${libs})
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND ${CMAKE_AR} cr ${outlibfile} *.o
COMMAND ${CMAKE_RANLIB} ${outlibfile}
WORKING_DIRECTORY ${lib}.objdir)
endforeach()
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND ${CMAKE_AR} crs ${target_LIBNAME} `find ${target_DIR} -name '*.o'`
COMMAND ${CMAKE_RANLIB} ${target_LIBNAME}
WORKING_DIRECTORY ${target_DIR})
endif()
endfunction(merge_static_libs)
@ -196,7 +200,7 @@ function(cc_library TARGET_NAME)
add_style_check_target(${TARGET_NAME} ${cc_library_SRCS} ${cc_library_HEADERS})
else(cc_library_SRCS)
if (cc_library_DEPS)
if(cc_library_DEPS)
merge_static_libs(${TARGET_NAME} ${cc_library_DEPS})
else()
message(FATAL "Please specify source file or library in cc_library.")

@ -25,7 +25,7 @@ function(target_circle_link_libraries TARGET_NAME)
endif()
endforeach()
if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang" OR "${CMAKE_CXX_COMPILER_ID}" STREQUAL "AppleClang")
if(IOS AND NOT IOS_ENABLE_BITCODE)
if(NOT IOS_ENABLE_BITCODE)
list(APPEND LIBS "-undefined dynamic_lookup")
endif()
endif()

@ -158,17 +158,23 @@ PaddlePaddle的参数使用名字 :code:`name` 作为参数的ID相同名字
这里 :code:`hidden_a`:code:`hidden_b` 使用了同样的parameter和bias。并且softmax层的两个输入也使用了同样的参数 :code:`softmax_param`
7. \*-cp27mu-linux_x86_64.whl is not a supported wheel on this platform.
7. paddlepaddle\*.whl is not a supported wheel on this platform.
------------------------------------------------------------------------
出现这个问题的主要原因是,系统编译wheel包的时候使用的 :code:`wheel` 包是最新的,
而系统中的 :code:`pip` 包比较老。具体的解决方法是,更新 :code:`pip` 包并重新编译PaddlePaddle。
出现这个问题的主要原因是,没有找到和当前系统匹配的paddlepaddle安装包。最新的paddlepaddle python安装包支持Linux x86_64和MacOS 10.12操作系统并安装了python 2.7和pip 9.0.1。
更新 :code:`pip` 包的方法是\:
.. code-block:: bash
pip install --upgrade pip
如果还不行,可以执行 :code:`python -c "import pip; print(pip.pep425tags.get_supported())"` 获取当前系统支持的python包的后缀
并对比是否和正在安装的后缀一致。
如果系统支持的是 :code:`linux_x86_64` 而安装包是 :code:`manylinux1_x86_64` 需要升级pip版本到最新
如果系统支持 :code:`manylinux1_x86_64` 而安装包(本地)是 :code:`linux_x86_64` 可以重命名这个whl包为 :code:`manylinux1_x86_64` 再安装。
8. python相关的单元测试都过不了
--------------------------------
@ -310,7 +316,7 @@ Paddle二进制在运行时捕获了浮点数异常只要出现浮点数异
* 模型一直不收敛,发散到了一个数值特别大的地方。
* 训练数据有问题,导致参数收敛到了一些奇异的情况。或者输入数据尺度过大,有些特征的取值达到数百万,这时进行矩阵乘法运算就可能导致浮点数溢出。
主要的解决办法是减小学习或者对数据进行归一化处理。
主要的解决办法是减小学习或者对数据进行归一化处理。
15. 编译安装后执行 import paddle.v2 as paddle 报ImportError: No module named v2
------------------------------------------------------------------------
@ -373,3 +379,15 @@ PaddlePaddle保存的模型参数文件内容由16字节头信息和网络参数
parameters = paddle.parameters.create(my_cost)
parameters.set('emb', load_parameter(emb_param_file, 30000, 256))
18. 集群多节点训练,日志中保存均为网络通信类错误
------------------------------
集群多节点训练,日志报错为网络通信类错误,比如 :code:`Connection reset by peer` 等。
此类报错通常是由于某一个节点的错误导致这个节点的训练进程退出,从而引发其他节点无法连接导致,可以参考下面的步骤排查:
* 从 :code:`train.log` :code:`server.log` 找到最早报错的地方查看是否是其他错误引发的报错比如FPE内存不足磁盘空间不足等
* 如果发现最早的报错就是网络通信的问题很有可能是非独占方式执行导致的端口冲突可以联系OP看当前MPI集群是否支持resource=full参数提交如果支持增加此参数提交并更换job 端口。
* 如果当前MPI集群并不支持任务独占模式可以联系OP是否可以更换集群或升级当前集群。

@ -62,6 +62,7 @@ if(ANDROID)
LIBRARY DESTINATION lib/${ANDROID_ABI})
execute_process(
COMMAND ${GIT_EXECUTABLE} log --pretty=oneline -1
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE GIT_COMMITS_LIST
RESULT_VARIABLE GIT_COMMITS_LIST_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
@ -81,8 +82,7 @@ if(ANDROID)
)"
)
else(ANDROID)
install(TARGETS paddle_capi_whole
ARCHIVE DESTINATION lib)
install(TARGETS paddle_capi_whole ARCHIVE DESTINATION lib)
if(NOT IOS)
install(TARGETS paddle_capi_shared DESTINATION lib)
endif()

@ -19,6 +19,19 @@ limitations under the License. */
namespace paddle {
namespace framework {
static ProgramDesc* g_program_desc = nullptr;
ProgramDesc& GetProgramDesc() {
if (g_program_desc == nullptr) {
g_program_desc = new ProgramDesc();
}
return *g_program_desc;
}
template <>
AttrType AttrTypeID<bool>() {
return BOOLEAN;
}
template <>
AttrType AttrTypeID<int>() {
return INT;
@ -32,6 +45,10 @@ AttrType AttrTypeID<std::string>() {
return STRING;
}
template <>
AttrType AttrTypeID<std::vector<bool>>() {
return BOOLEANS;
}
template <>
AttrType AttrTypeID<std::vector<int>>() {
return INTS;
}
@ -47,40 +64,54 @@ template <>
AttrType AttrTypeID<std::vector<std::pair<int, int>>>() {
return INT_PAIRS;
}
template <>
AttrType AttrTypeID<BlockDesc>() {
return BLOCK;
}
Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
switch (attr_desc.type()) {
case paddle::framework::AttrType::INT: {
case framework::AttrType::BOOLEAN: {
return attr_desc.b();
}
case framework::AttrType::INT: {
return attr_desc.i();
}
case paddle::framework::AttrType::FLOAT: {
case framework::AttrType::FLOAT: {
return attr_desc.f();
}
case paddle::framework::AttrType::STRING: {
case framework::AttrType::STRING: {
return attr_desc.s();
}
case paddle::framework::AttrType::INTS: {
case framework::AttrType::BOOLEANS: {
std::vector<bool> val(attr_desc.bools_size());
for (int i = 0; i < attr_desc.bools_size(); ++i) {
val[i] = attr_desc.bools(i);
}
return val;
}
case framework::AttrType::INTS: {
std::vector<int> val(attr_desc.ints_size());
for (int i = 0; i < attr_desc.ints_size(); ++i) {
val[i] = attr_desc.ints(i);
}
return val;
}
case paddle::framework::AttrType::FLOATS: {
case framework::AttrType::FLOATS: {
std::vector<float> val(attr_desc.floats_size());
for (int i = 0; i < attr_desc.floats_size(); ++i) {
val[i] = attr_desc.floats(i);
}
return val;
}
case paddle::framework::AttrType::STRINGS: {
case framework::AttrType::STRINGS: {
std::vector<std::string> val(attr_desc.strings_size());
for (int i = 0; i < attr_desc.strings_size(); ++i) {
val[i] = attr_desc.strings(i);
}
return val;
}
case paddle::framework::AttrType::INT_PAIRS: {
case framework::AttrType::INT_PAIRS: {
std::vector<std::pair<int, int>> val(attr_desc.int_pairs_size());
for (int i = 0; i < attr_desc.int_pairs_size(); ++i) {
val[i].first = attr_desc.int_pairs(i).first();
@ -88,6 +119,9 @@ Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
}
return val;
}
case framework::AttrType::BLOCK: {
return GetProgramDesc().mutable_blocks(attr_desc.block_idx());
}
}
PADDLE_ENFORCE(false, "Unknown OpDesc::AttrDesc::type !");
return boost::blank();

@ -27,13 +27,16 @@ limitations under the License. */
namespace paddle {
namespace framework {
typedef boost::variant<boost::blank, int, float, std::string, std::vector<int>,
std::vector<float>, std::vector<std::string>,
std::vector<std::pair<int, int>>>
typedef boost::variant<boost::blank, bool, int, float, std::string,
std::vector<bool>, std::vector<int>, std::vector<float>,
std::vector<std::string>,
std::vector<std::pair<int, int>>, BlockDesc*>
Attribute;
typedef std::unordered_map<std::string, Attribute> AttributeMap;
ProgramDesc& GetProgramDesc();
template <typename T>
AttrType AttrTypeID();

@ -166,9 +166,8 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
// If part of input gradient of that operator is not calculated, fill
// zero variables to that input gradient.
net->AppendOp(OpRegistry::CreateOp("fill_zeros_like",
{{"Src", {prefix}}},
{{"Dst", {grad_input}}}, {}));
net->AppendOp(OpRegistry::CreateOp("fill_zeros_like", {{"X", {prefix}}},
{{"Y", {grad_input}}}, {}));
}
return false;
});

@ -127,8 +127,8 @@ class FillZeroOpMaker : public OpProtoAndCheckerMaker {
public:
FillZeroOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("Src", "x");
AddOutput("Dst", "out");
AddInput("X", "x");
AddOutput("Y", "out");
AddComment("");
}
};
@ -325,10 +325,10 @@ TEST(Backward, op_part_of_output_are_not_need) {
auto &fill_zero = *net->ops_[0];
ASSERT_EQ("fill_zeros_like", fill_zero.Type());
ASSERT_EQ(1UL, fill_zero.Inputs("Src").size());
ASSERT_EQ("Z", fill_zero.Input("Src"));
ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size());
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst"));
ASSERT_EQ(1UL, fill_zero.Inputs("X").size());
ASSERT_EQ("Z", fill_zero.Input("X"));
ASSERT_EQ(1UL, fill_zero.Outputs("Y").size());
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Y"));
auto &d_many_out = *net->ops_[1];
ASSERT_EQ("many_output_op_grad", d_many_out.Type());

@ -292,5 +292,13 @@ DDim flatten_to_2d(const DDim& src, int num_col_dims) {
DDim flatten_to_1d(const DDim& src) { return make_ddim({product(src)}); }
DDim stride(const DDim& ddim) {
std::vector<int64_t> strides(ddim.size());
strides[ddim.size() - 1] = 1;
for (int i = ddim.size() - 2; i >= 0; --i) {
strides[i] = strides[i + 1] * ddim[i + 1];
}
return framework::make_ddim(strides);
}
} // namespace framework
} // namespace paddle

@ -121,6 +121,7 @@ DDim flatten_to_2d(const DDim& src, int num_col_dims);
DDim flatten_to_1d(const DDim& src);
DDim stride(const DDim& ddim);
} // namespace framework
} // namespace paddle

@ -23,6 +23,9 @@ enum AttrType {
FLOATS = 4;
STRINGS = 5;
INT_PAIRS = 6;
BOOLEAN = 7;
BOOLEANS = 8;
BLOCK = 9;
}
message IntPair {
@ -44,6 +47,9 @@ message OpDesc {
repeated float floats = 7;
repeated string strings = 8;
repeated IntPair int_pairs = 9;
optional bool b = 10;
repeated bool bools = 11;
optional int32 block_idx = 12;
};
message Var {
@ -108,3 +114,12 @@ message VarDesc {
required string name = 1;
optional LoDTensorDesc lod_tensor = 2;
}
message BlockDesc {
required int32 idx = 1;
required int32 parent_idx = 2;
repeated VarDesc vars = 3;
repeated OpDesc ops = 4;
}
message ProgramDesc { repeated BlockDesc blocks = 1; }

@ -207,23 +207,22 @@ const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
}
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
auto* var = OutputVar(name);
return var == nullptr ? nullptr : const_cast<Tensor*>(GetTensorFromVar(var));
Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const {
auto var = OutputVar(name);
return var == nullptr ? nullptr : var->GetMutable<LoDTensor>();
}
template <>
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
const std::string& name) const {
auto names = op().Outputs(name);
std::vector<Tensor*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
auto var = scope().FindVar(sub_name);
return var == nullptr
? nullptr
: const_cast<Tensor*>(GetTensorFromVar(var));
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr
: var->GetMutable<LoDTensor>();
});
return res;
}

@ -212,9 +212,9 @@ class InferShapeContext {
return res;
}
std::vector<const Variable*> MultiOutputVar(const std::string& name) const {
std::vector<Variable*> MultiOutputVar(const std::string& name) const {
auto names = op_.Outputs(name);
std::vector<const Variable*> res;
std::vector<Variable*> res;
res.reserve(names.size());
std::transform(names.begin(), names.end(), std::back_inserter(res),
[this](const std::string& name) {
@ -271,6 +271,20 @@ class InferShapeContext {
return &var->Get<Tensor>();
}
void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
size_t j = 0) const {
PADDLE_ENFORCE_LT(i, InputSize(in));
PADDLE_ENFORCE_LT(j, OutputSize(out));
auto* in_var = MultiInputVar(in)[i];
auto* out_var = MultiOutputVar(out)[j];
if (!in_var->IsType<LoDTensor>()) return;
PADDLE_ENFORCE(out_var->IsType<LoDTensor>(),
"The %d-th output of Output(%s) must be LoDTensor.", j, out);
auto in_tensor = in_var->Get<LoDTensor>();
auto* out_tensor = out_var->GetMutable<LoDTensor>();
out_tensor->set_lod(in_tensor.lod());
}
private:
const OperatorBase& op_;
const Scope& scope_;
@ -283,6 +297,13 @@ template <>
const std::vector<const Tensor*> InferShapeContext::MultiInput<Tensor>(
const std::string& name) const;
template <>
Tensor* InferShapeContext::Output<Tensor>(const std::string& name) const;
template <>
std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
const std::string& name) const;
template <typename T>
struct EigenDeviceConverter;
@ -315,38 +336,10 @@ class ExecutionContext : public InferShapeContext {
return device_context_;
}
// redefine Output function,
// use Variable::Get instead of Variable::GetMutable
template <typename T>
T* Output(const std::string& name) const {
auto var = OutputVar(name);
return var == nullptr ? nullptr : const_cast<T*>(&var->Get<T>());
}
// redefine MultiOutput function.
// use Variable::Get instead of Variable::GetMutable
template <typename T>
std::vector<T*> MultiOutput(const std::string& name) const {
auto names = op().Outputs(name);
std::vector<T*> res;
res.reserve(names.size());
std::transform(
names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) { return Output<T>(sub_name); });
return res;
}
private:
const platform::DeviceContext& device_context_;
};
template <>
Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
template <>
std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
const std::string& name) const;
class OpKernel {
public:
/**

@ -55,6 +55,13 @@ function(op_library TARGET)
set(pybind_flag 1)
endif()
# activation_op contains several operators
if ("${TARGET}" STREQUAL "activation_op")
set(pybind_flag 1)
# It's enough to just adding one operator to pybind
file(APPEND ${pybind_file} "USE_OP(sigmoid);\n")
endif()
# pybind USE_NO_KERNEL_OP
file(READ ${TARGET}.cc TARGET_CONTENT)
string(REGEX MATCH "OperatorWithKernel" regex_result "${TARGET_CONTENT}")

@ -39,7 +39,8 @@ class AccuracyOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
"inference size must be the same as label size");
ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
ctx.Output<framework::Tensor>("Accuracy")->Resize({1});
ctx.ShareLoD("Inference", /*->*/ "Accuracy");
}
};
@ -54,11 +55,15 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
// TODO(typhoonzero): AddInput("Weight", ...
AddOutput("Accuracy", "The accuracy of current batch");
AddComment(
R"DOC(Accuracy. It will print accuracy rate for classification.
AddComment(R"DOC(
Accuracy. It will print accuracy rate for classification.
The accuracy is:
.. math::
accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})
Both the input `Inference` and `Label` can carry the LoD (Level of Details)
information, or not. But the output only shares the LoD with input `Inference`.
)DOC");
}
};

File diff suppressed because it is too large Load Diff

@ -0,0 +1,100 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#define EIGEN_USE_GPU
#include "paddle/operators/activation_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(sigmoid,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::SigmoidFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
sigmoid_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::SigmoidGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
exp,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::ExpFunctor>);
REGISTER_OP_GPU_KERNEL(exp_grad,
ops::ActivationGradKernel<paddle::platform::GPUPlace,
float, ops::ExpGradFunctor>);
REGISTER_OP_GPU_KERNEL(relu,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::ReluFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
relu_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::ReluGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
tanh,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::TanhFunctor>);
REGISTER_OP_GPU_KERNEL(
tanh_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::TanhGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
sqrt,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::SqrtFunctor>);
REGISTER_OP_GPU_KERNEL(
sqrt_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::SqrtGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
abs,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::AbsFunctor>);
REGISTER_OP_GPU_KERNEL(abs_grad,
ops::ActivationGradKernel<paddle::platform::GPUPlace,
float, ops::AbsGradFunctor>);
REGISTER_OP_GPU_KERNEL(reciprocal,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::ReciprocalFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
reciprocal_grad,
ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::ReciprocalGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(
log,
ops::ActivationKernel<paddle::platform::GPUPlace, float, ops::LogFunctor>);
REGISTER_OP_GPU_KERNEL(
log_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::LogGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(square,
ops::ActivationKernel<paddle::platform::GPUPlace, float,
ops::SquareFunctor>);
REGISTER_OP_GPU_KERNEL(
square_grad, ops::ActivationGradKernel<paddle::platform::GPUPlace, float,
ops::SquareGradFunctor<float>>);
REGISTER_OP_GPU_KERNEL(brelu,
ops::BReluKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(brelu_grad,
ops::BReluGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(soft_relu,
ops::SoftReluKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
soft_relu_grad, ops::SoftReluGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(pow, ops::PowKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(pow_grad,
ops::PowGradKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(stanh,
ops::STanhKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(stanh_grad,
ops::STanhGradKernel<paddle::platform::GPUPlace, float>);

File diff suppressed because it is too large Load Diff

@ -33,7 +33,7 @@ class AddOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
ctx.Input<Tensor>("Y")->dims(),
"Two input of Add Op's dimension must be same.");
ctx.Output<framework::LoDTensor>("Out")->Resize(
ctx.Output<framework::Tensor>("Out")->Resize(
ctx.Input<Tensor>("X")->dims());
}
};

@ -17,8 +17,6 @@
namespace paddle {
namespace operators {
using framework::LoDTensor;
class ClipOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
@ -29,11 +27,12 @@ class ClipOp : public framework::OperatorWithKernel {
"Input(X) of ClipOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
"Output(Out) of ClipOp should not be null.");
auto x_dims = ctx.Input<LoDTensor>("X")->dims();
auto x_dims = ctx.Input<Tensor>("X")->dims();
auto max = Attr<float>("max");
auto min = Attr<float>("min");
PADDLE_ENFORCE_LT(min, max, "max should be greater than min.");
ctx.Output<LoDTensor>("Out")->Resize(x_dims);
ctx.Output<Tensor>("Out")->Resize(x_dims);
ctx.ShareLoD("X", /*->*/ "Out");
}
};
@ -66,8 +65,8 @@ class ClipOpGrad : public framework::OperatorWithKernel {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
"Input(Out@GRAD) should not be null");
auto x_dims = ctx.Input<LoDTensor>("X")->dims();
auto *x_grad = ctx.Output<LoDTensor>(framework::GradVarName("X"));
auto x_dims = ctx.Input<Tensor>("X")->dims();
auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
if (x_grad != nullptr) {
x_grad->Resize(x_dims);
}

@ -29,7 +29,7 @@ class ConcatOp : public framework::OperatorWithKernel {
"Output(Out) of ConcatOp should not be null.");
auto ins = ctx.MultiInput<framework::Tensor>("X");
auto *out = ctx.Output<framework::LoDTensor>("Out");
auto *out = ctx.Output<framework::Tensor>("Out");
size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
size_t n = ins.size();

@ -37,7 +37,7 @@ class Conv2DOp : public framework::OperatorWithKernel {
auto in = ctx.Input<Tensor>("Input");
auto filter = ctx.Input<Tensor>("Filter");
auto out = ctx.Output<framework::LoDTensor>("Output");
auto out = ctx.Output<framework::Tensor>("Output");
std::vector<int> strides = Attr<std::vector<int>>("strides");
std::vector<int> paddings = Attr<std::vector<int>>("paddings");
int groups = Attr<int>("groups");
@ -111,10 +111,9 @@ class Conv2DOpGrad : public framework::OperatorWithKernel {
void InferShape(const framework::InferShapeContext &ctx) const override {
auto in = ctx.Input<Tensor>("Input");
auto filter = ctx.Input<Tensor>("Filter");
auto d_in =
ctx.Output<framework::LoDTensor>(framework::GradVarName("Input"));
auto d_in = ctx.Output<framework::Tensor>(framework::GradVarName("Input"));
auto d_filter =
ctx.Output<framework::LoDTensor>(framework::GradVarName("Filter"));
ctx.Output<framework::Tensor>(framework::GradVarName("Filter"));
if (d_in) d_in->Resize(in->dims());
if (d_filter) d_filter->Resize(filter->dims());
}

@ -54,9 +54,10 @@ class CosSimOp : public framework::OperatorWithKernel {
" just 1 (which will be broadcasted to match Input(X)).");
// resize tensor
ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
ctx.Output<framework::Tensor>("Out")->Resize({x_dims[0], 1});
ctx.Output<framework::Tensor>("XNorm")->Resize({x_dims[0], 1});
ctx.Output<framework::Tensor>("YNorm")->Resize({y_dims[0], 1});
ctx.ShareLoD("X", /*->*/ "Out");
}
};
@ -81,10 +82,13 @@ Cosine Similarity Operator.
The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
Input(X) and Input(Y) must have the same shape, except that the 1st dimension
of Input(Y) could be just 1 (different from Input(X)), which will be
broadcasted to match the shape of Input(X) before computing their cosine
The input `X` and `Y` must have the same shape, except that the 1st dimension
of input `Y` could be just 1 (different from input `X`), which will be
broadcasted to match the shape of input `X` before computing their cosine
similarity.
Both the input `X` and `Y` can carry the LoD (Level of Details) information,
or not. But the output only shares the LoD with input `X`.
)DOC");
}
};
@ -139,10 +143,8 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
"Shape of Input(Out@Grad) must be [X.Dim(0), 1].");
// resize tensor
auto *x_grad =
ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
auto *y_grad =
ctx.Output<framework::LoDTensor>(framework::GradVarName("Y"));
auto *x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
auto *y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
if (x_grad) x_grad->Resize(x_dims);
if (y_grad) y_grad->Resize(y_dims);
}

@ -0,0 +1,138 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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/operators/crop_op.h"
#include <boost/lexical_cast.hpp>
namespace paddle {
namespace operators {
using framework::Tensor;
class CropOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
"Input(X) of CropOp should not be null.");
PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
"Output(Out) of CropOp should not be null.");
auto x_dim = ctx.Input<Tensor>("X")->dims();
auto *y = ctx.Input<Tensor>("Y");
auto *out = ctx.Output<Tensor>("Out");
if (y == nullptr) {
auto shape = Attr<std::vector<int>>("shape");
PADDLE_ENFORCE_EQ(
int64_t(shape.size()), x_dim.size(),
"Shape size should be equal to dimention size of input tensor.");
std::vector<int64_t> tensor_shape(shape.size());
for (size_t i = 0; i < shape.size(); ++i) {
tensor_shape[i] = static_cast<int64_t>(shape[i]);
}
out->Resize(framework::make_ddim(tensor_shape));
} else {
PADDLE_ENFORCE_EQ(framework::arity(x_dim), framework::arity(y->dims()),
"Tensor rank of both CropOp's "
"inputs must be same.");
out->Resize(y->dims());
}
}
};
class CropOpMaker : public framework::OpProtoAndCheckerMaker {
public:
CropOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X",
"The input of pad op. "
"The input should be a k-D tensor(k > 0 and k < 7)");
AddInput("Y",
"The input used as reference for cropping"
" with the same dimension as X. ");
AddOutput("Out",
"The output of crop op "
"with the same dimension as X.");
AddAttr<std::vector<int>>("offsets",
"A list<int> describing offsets to be cropped."
"The size of offsets list should be as same as "
"dimension size of input X.");
AddAttr<std::vector<int>>("shape",
"A list<int> describing the shape of output."
"The size of shape list should be as same as "
"dimension size of input X.")
.SetDefault(std::vector<int>());
AddComment(R"DOC(
Crop Operator.
Crop input into output, as specified by offsets and shape.
There are two ways to set shape:
1. referenc input: crop input X as shape as reference input.
The dimension of reference input should
be as same as input X.
2. shape list: crop input X by shape described by a list<int>.
The size of shape list should be as same as
dimension size of input X.
The input should be a k-D tensor(k > 0 and k < 7). As an example:
Given:
X = [[0, 1, 2, 0, 0]
[0, 3, 4, 0, 0]
[0, 0, 0, 0, 0]]
and
offsets = [0, 1]
and
shape = [2, 2]
then we get
Out = [[1, 2],
[3, 4]]
)DOC");
}
};
class CropOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
"Input(Out@GRAD) should not be null");
auto x_dims = ctx.Input<Tensor>("X")->dims();
auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
if (x_grad != nullptr) {
x_grad->Resize(x_dims);
}
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(crop, ops::CropOp, ops::CropOpMaker, crop_grad, ops::CropOpGrad);
REGISTER_OP_CPU_KERNEL(crop, ops::CropKernel<float>);
REGISTER_OP_CPU_KERNEL(crop_grad,
ops::CropGradKernel<paddle::platform::CPUPlace, float>);

@ -0,0 +1,21 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#define EIGEN_USE_GPU
#include "paddle/operators/crop_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(crop, ops::CropKernel<float>);
REGISTER_OP_GPU_KERNEL(crop_grad,
ops::CropGradKernel<paddle::platform::GPUPlace, float>);

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save