Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into bp_test

revert-3824-remove_grad_op_type
dangqingqing 8 years ago
commit 18dcc1c0bc

@ -0,0 +1,15 @@
#!/bin/bash
set -e
readonly VERSION="3.8"
version=$(clang-format -version)
if ! [[ $version == *"$VERSION"* ]]; then
echo "clang-format version check failed."
echo "a version contains '$VERSION' is needed, but get '$version'"
echo "you can install the right version, and make an soft-link to '\$PATH' env"
exit -1
fi
clang-format $@

@ -19,10 +19,10 @@
- id: end-of-file-fixer
- repo: local
hooks:
- id: clang-format
- id: clang-format-with-version-check
name: clang-format
description: Format files with ClangFormat.
entry: clang-format -i
entry: ./.clang_format.hook -i
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
- repo: https://github.com/PaddlePaddle/pre-commit-golang

@ -36,8 +36,8 @@ include(simd)
################################ Configurations #######################################
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." OFF)
option(WITH_MKLML "Compile PaddlePaddle with mklml package." OFF)
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND})
option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND})
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)

@ -34,9 +34,6 @@ RUN apt-get update && \
net-tools && \
apt-get clean -y
# paddle is using numpy.flip, which is introduced since 1.12.0
RUN pip --no-cache-dir install 'numpy>=1.12.0'
# Install Go and glide
RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
tar -xz -C /usr/local && \
@ -58,33 +55,22 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
# FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter
# version util jupyter fixes this issue.
RUN pip install --upgrade pip && \
pip install -U 'protobuf==3.1.0' && \
pip install -U wheel pillow BeautifulSoup && \
pip install -U wheel && \
pip install -U docopt PyYAML sphinx && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark
RUN pip install pre-commit 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2'
pip install opencv-python
COPY ./python/requirements.txt /root/
RUN pip install -r /root/requirements.txt
# To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use
# the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2
RUN apt-get install -y libssl-dev libffi-dev
RUN pip install certifi urllib3[secure]
# TODO(qijun) The template library Eigen doesn't work well with GCC 5
# coming with the default Docker image, so we switch to use GCC 4.8
# by default. And I will check Eigen library later.
RUN ln -sf gcc-4.8 /usr/bin/gcc && \
ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \
ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \
ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \
ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \
ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \
ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \
ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \
ln -sf g++-4.8 /usr/bin/g++ && \
ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++
# Install woboq_codebrowser to /woboq
RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \

@ -73,10 +73,18 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
ADD_LIBRARY(cblas STATIC ${dummyfile})
IF(${CBLAS_PROVIDER} MATCHES MKL)
ADD_LIBRARY(cblas SHARED ${dummyfile})
ELSE()
ADD_LIBRARY(cblas STATIC ${dummyfile})
ENDIF()
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
LIST(APPEND external_project_dependencies cblas)
ELSE()
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_DEPENDENCIES(cblas mklml)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})

@ -9,13 +9,6 @@ function(CheckCompilerCXX11Flag)
if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
endif()
if(NOT ANDROID)
# TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
# Use Debug mode instead for now.
if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
endif()
endif()
elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
# cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
# Apple Clang is a different compiler than upstream Clang which havs different version numbers.
@ -160,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
# So, don't set these flags here.
LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread)
LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")

@ -0,0 +1,101 @@
# Alalysis of large model distributed training in Paddle
***NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.***
## What is it
We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer's memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters.
## How to use
Specify command-line argument like `--loadsave_parameters_in_pserver=true --ports_num_for_sparse=1 --use_old_updater=1` when starting the paddle trainer. And also add something like `--ports_num_for_sparse=1 --pserver_num_threads=5` when starting pserver processes.
Accrodingly, configure your embedding layers like:
```python
SPARSE_REMOTE=True
w1 = data_layer(name="w1", size=dict_size)
emb1 = embedding_layer(input=w1, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
w2 = data_layer(name="w2", size=dict_size)
emb2 = embedding_layer(input=w2, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
...
```
## Implementation details
```c++
enum MatType {
MAT_NORMAL,
MAT_NORMAL_SHARED,
MAT_VALUE_SHARED,
MAT_SPARSE_ROW_IDS,
MAT_SPARSE_ROW_AUTO_GROW,
MAT_CACHE_ROW,
MAT_SPARSE_ROW,
MAT_SPARSE_ROW_PREFETCH,
MAT_SPARSE_ROW_PREFETCH_FULL_SIZE,
};
```
`MAT_SPARSE_ROW_PREFETCH` is what we use when configured to fetch only row of matrix when training.
In `trainer_internal.cpp:L93 trainOneBatch`:
```c++
if (config_->getOptConfig().use_sparse_remote_updater()) {
REGISTER_TIMER("prefetch");
gradientMachine_->prefetch(inArgs);
parameterUpdater_->getParametersRemote();
}
```
When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver.
In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`:
```c++
if (fullSize) {
...
} else {
getParams = [&] {
parameterClient_->getParameterSparse(
/* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType);
};
applyL1 = [](Parameter& para, real decayRate) {
para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate);
};
}
```
Calling `parameterClient_->getParameterSparse` will do remote call to pserver's `getParameterSparse`:
```c++
void ParameterServer2::getParameterSparse(const SendParameterRequest& request,
std::vector<Buffer>& inputBuffers,
SendParameterResponse* response,
std::vector<Buffer>* outputBuffers) {
(void)inputBuffers;
auto& buffer = *readWriteBuffer_;
size_t numReals = 0;
for (const auto& block : request.blocks()) {
numReals += getParameterConfig(block).dims(1);
}
buffer.resize(numReals);
VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals;
ReadLockGuard guard(parameterMutex_);
size_t offset = 0;
for (const auto& block : request.blocks()) {
size_t width = getParameterConfig(block).dims(1);
Buffer buf = {buffer.data() + offset, width};
int type = request.send_back_parameter_type();
sendBackParameterSparse(block, type, response, &buf, width, outputBuffers);
offset += width;
}
}
```
`getParameterConfig(block).dims(1)` returns the width of the current "parameter block"(a shard of parameter object),
then `getParameterSparse` remote call returns only one row of data to the client.

@ -101,6 +101,7 @@ if use_mkldnn
5. 在**Argument**里添加两个`MkldnnMatrixPtr`,取名为`mkldnnValue`和`mkldnnGrad`,用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名),用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑,用于判断`deviceId`并针对device在MKL-DNN和CPU之间不统一的情况做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag用于选择是否使用MKL-DNN的相关功能。
8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面一直保存的是0所以可以充分利用这个信息定义一个枚举处理所有MKLDNN的参数格式从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
## References

@ -68,7 +68,7 @@ As a simple example, consider the following:
1. **BLAS Dependencies(optional)**
CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
To utilize preinstalled BLAS you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
```bash
@ -131,9 +131,9 @@ As a simple example, consider the following:
To build GPU version, you will need the following installed:
1. a CUDA-capable GPU
2. A supported version of Linux with a gcc compiler and toolchain
2. A supported version of Linux with a GCC compiler and toolchain
3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
The CUDA development environment relies on tight integration with the host development environment,
including the host compiler and C runtime libraries, and is therefore only supported on
@ -172,6 +172,7 @@ export PATH=<path to install>/bin:$PATH
# install PaddlePaddle Python modules.
sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
```
## <span id="centos">Build on Centos 7</span>
### Install Dependencies
@ -192,9 +193,9 @@ sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
To build GPU version, you will need the following installed:
1. a CUDA-capable GPU
2. A supported version of Linux with a gcc compiler and toolchain
2. A supported version of Linux with a GCC compiler and toolchain
3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
The CUDA development environment relies on tight integration with the host development environment,
including the host compiler and C runtime libraries, and is therefore only supported on
@ -222,7 +223,7 @@ mkdir build && cd build
```
Finally, you can build and install PaddlePaddle:
```bash
# you can add build option here, such as:
cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>

@ -146,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param(
m->machine->randParameters();
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_get_layer_output(
paddle_gradient_machine machine,
const char* layerName,
paddle_arguments args) {
auto m = cast(machine);
auto out = paddle::capi::cast<paddle::capi::CArguments>(args);
if (m == nullptr || layerName == nullptr || out == nullptr ||
m->machine == nullptr) {
return kPD_NULLPTR;
}
auto layerOutput = m->machine->getLayerOutput(layerName);
out->args.push_back(layerOutput);
return kPD_NO_ERROR;
}

@ -39,7 +39,11 @@ PD_API paddle_error paddle_gradient_machine_create_for_inference(
/**
* @brief Create a gradient machine used for model inference, using config with
* parameters which is generated by `paddle merge_model`.
* @param [out] machine that used for model inference.
* Example:
* paddle merge_model \
* --model_dir="pass-00000" \
* --model_file="merged_model.paddle"
* @param [out] machine that used for model inference
* @param [in] mergedModel
* @param [in] size
* @return paddle_error
@ -97,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine);
PD_API paddle_error
paddle_gradient_machine_destroy(paddle_gradient_machine machine);
/**
* @brief Get the output of the layer named `layerName`.
* @param [in] gradient machine that have run a inference
* @param [in] layerName name of specified layer
* @param [out] args output of the specified layer
* @return paddle_error
*/
PD_API paddle_error
paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
const char* layerName,
paddle_arguments args);
#ifdef __cplusplus
}
#endif

@ -38,7 +38,7 @@ add_custom_command(TARGET framework_py_proto POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
cc_library(backward SRCS backward.cc DEPS net_op)
cc_test(backward_test SRCS backward_test.cc DEPS backward)
cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
if(WITH_PYTHON)
cc_library(paddle_pybind SHARED

@ -17,12 +17,13 @@
#include <list>
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
namespace paddle {
namespace framework {
template <typename Map, typename T>
static void ForEachVarName(Map& names, T callback) {
static void ForEachVarName(const Map& names, T callback) {
for (auto& name : names) {
for (auto& n : name.second) {
if (callback(n)) return;
@ -44,7 +45,7 @@ static bool AllInSet(
static std::shared_ptr<OperatorBase> NOP() {
auto net_op = std::make_shared<operators::NetOp>();
net_op->type_ = "@NOP@";
net_op->SetType("@NOP@");
net_op->CompleteAddOp();
return net_op;
}
@ -70,8 +71,8 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
// If all input gradients of forwarding operator do not need to calculate,
// just return an NOP. Not return null ptr because NOP does not take
// much time for calculation, but it is useful for simplifying logic.
if (AllInSet(forwardOp.inputs_ /*names*/, kGradVarSuffix /*suffix*/,
// too much time for calculation, but it is useful for simplifying logic.
if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/,
no_grad_names /*set*/)) {
return NOP();
}
@ -79,9 +80,9 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
// All output gradients of forwarding operator do not need to calculate.
// Then all input gradients cannot be computed at all, and we put them into
// `no_grad_names` set. Return an NOP.
if (AllInSet(forwardOp.outputs_ /*names*/, kGradVarSuffix /*suffix*/,
if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/,
no_grad_names /*set*/)) {
ForEachVarName(forwardOp.inputs_,
ForEachVarName(forwardOp.Inputs(),
[&no_grad_names](const std::string& name) -> bool {
no_grad_names.insert(GradVarName(name));
return false;
@ -107,7 +108,7 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
auto fwd = *it;
auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
net->AddOp(bwd);
ForEachVarName(bwd->outputs_,
ForEachVarName(bwd->Outputs(),
[&dup_output_ops, local_op_id](const std::string& out) {
dup_output_ops[out].emplace_back(local_op_id);
return false;
@ -154,13 +155,13 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
} else {
std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
ForEachVarName(grad_op->inputs_, [&no_grad_names,
&net](std::string& grad_input) {
ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net,
grad_op](const std::string& grad_input) {
if (no_grad_names.count(grad_input)) {
// +1 for \0
std::string prefix = grad_input.substr(
0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
grad_input = prefix + kZeroVarSuffix;
grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
// If part of input gradient of that operator is not calculated, fill
// zero variables to that input gradient.
@ -170,20 +171,36 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
return false;
});
ForEachVarName(grad_op->outputs_,
[&no_grad_names](std::string& grad_output) {
ForEachVarName(grad_op->Outputs(),
[&no_grad_names, &grad_op](const std::string& grad_output) {
if (no_grad_names.count(grad_output)) {
grad_output = kEmptyVarName;
grad_op->Rename(grad_output, kEmptyVarName);
}
return false;
});
// process recurrent gradient op as a special operator.
if (forwardOp.Type() == "recurrent_op") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
// this will result in infinite loop.
const auto& rnnop =
*static_cast<const operators::RecurrentOp*>(&forwardOp);
auto rnn_grad_op =
static_cast<operators::RecurrentGradientOp*>(grad_op.get());
const auto& stepnet_op =
*static_cast<const OperatorBase*>(&rnnop.stepnet());
// create stepnet's gradient op
auto grad_stepnet = BackwardRecursive(stepnet_op, no_grad_names, uniq_id);
rnn_grad_op->set_stepnet(
std::static_pointer_cast<operators::NetOp>(grad_stepnet));
}
if (net->ops_.empty()) { // Current no aux op is added to network
return grad_op;
}
net->AddOp(grad_op);
}
net->type_ = "@GENERATED_BACKWARD@";
net->SetType("@GENERATED_BACKWARD@");
net->CompleteAddOp();
return net;
} // namespace framework

@ -28,13 +28,6 @@ using OpAttrChecker = framework::OpAttrChecker;
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
class EmptyOp : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
};
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
@ -155,27 +148,24 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
namespace f = paddle::framework;
namespace ops = paddle::operators;
using EnforceNotMet = paddle::platform::EnforceNotMet;
REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp);
REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp);
REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker);
REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker);
REGISTER_OP(add, f::EmptyOp, f::AddOpMaker);
REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp);
REGISTER_OP(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
f::NOP);
REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
f::NOP);
TEST(Backward, simple_op_grad) {
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
ASSERT_NE(fwd, nullptr);
auto gop = f::OpRegistry::CreateGradOp(*fwd);
ASSERT_EQ(1UL, gop->inputs_.size());
ASSERT_EQ("rowwise_add_grad", gop->type_);
ASSERT_EQ(1UL, gop->Inputs().size());
ASSERT_EQ("rowwise_add_grad", gop->Type());
ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
}
@ -211,13 +201,13 @@ TEST(Backward, net_fc_backward_normal) {
ASSERT_EQ(3UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_add = *net->ops_[1];
ASSERT_EQ("rowwise_add_grad", d_add.type_);
ASSERT_EQ("rowwise_add_grad", d_add.Type());
f::OperatorBase &d_mul = *net->ops_[2];
ASSERT_EQ("mul_grad", d_mul.type_);
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_fc_backward_not_have_b) {
@ -237,10 +227,10 @@ TEST(Backward, net_fc_backward_not_have_b) {
ASSERT_EQ(2UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_mul = *net->ops_[1];
ASSERT_EQ("mul_grad", d_mul.type_);
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_input_of_network_not_need_grad) {
@ -294,7 +284,7 @@ TEST(Backward, net_shared_weight) {
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
ASSERT_EQ(3UL, bwd_net->ops_.size());
ASSERT_EQ("add", bwd_net->ops_[2]->type_);
ASSERT_EQ("add", bwd_net->ops_[2]->Type());
}
TEST(Backward, op_register_grad_not_for_network) {
@ -335,15 +325,15 @@ TEST(Backward, op_part_of_output_are_not_need) {
ASSERT_EQ(net->ops_.size(), 2UL);
auto &fill_zero = *net->ops_[0];
ASSERT_EQ("fill_zeros_like", fill_zero.type_);
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"));
auto &d_many_out = *net->ops_[1];
ASSERT_EQ("many_output_op_grad", d_many_out.type_);
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size()); // I/O/OG
ASSERT_EQ("many_output_op_grad", d_many_out.Type());
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size()); // I/O/OG
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
d_many_out.Input(f::GradVarName("z")));
ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
@ -355,9 +345,9 @@ TEST(Backward, op_part_of_input_are_not_need) {
{{"Out", {"out"}}}, {});
auto backward = f::Backward(*fwd, {"a"});
auto &grad_mul = *backward;
ASSERT_EQ(grad_mul.type_, "mul_grad");
ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
ASSERT_EQ(grad_mul.Type(), "mul_grad");
ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.Outputs().size(), 2UL);
ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b"));
ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
@ -395,18 +385,18 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
auto &grad_fc = *bwd_net->ops_[0];
const char *all = paddle::operators::NetOp::kAll;
EXPECT_EQ(grad_fc.inputs_[all].size(),
EXPECT_EQ(grad_fc.Inputs(all).size(),
2UL /* external input number */
+ 1UL /* external output number*/
+ 1UL /* number of gradient of external output*/
+ 2U /* internal variable number*/);
EXPECT_EQ(grad_fc.outputs_[all].size(),
EXPECT_EQ(grad_fc.Outputs(all).size(),
2UL /* input number of mul*/
+ 2UL /* input number of rowwise_add
*/
+ 1UL /* input number of sigmod */);
EXPECT_EQ(bwd_net->ops_[1]->inputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->outputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->inputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->outputs_[all].size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
}

@ -13,23 +13,20 @@ express or implied. See the License for the specific language governing
permissions and limitations under the License. */
#include "paddle/framework/grad_op_builder.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace framework {
enum class OpArgType { IN, OUT };
static void TransOpArg(const OperatorBase* src_op,
OperatorBase::VarNameMap* vars,
const OpArgType& src_type, bool is_grad) {
static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
bool is_grad, OperatorBase::VarNameMap* vars) {
const auto& src_inout =
src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
auto& dst_inout = *vars;
const OpProto& proto = OpProtos().at(src_op->type_);
const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_;
const auto& src_arg_list =
src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
for (const auto& arg : src_arg_list) {
if (arg.no_gradient() && !is_grad) continue;
const std::string src_name = arg.name();
@ -43,22 +40,26 @@ static void TransOpArg(const OperatorBase* src_op,
}
OperatorBase* BuildGradOp(const OperatorBase* op) {
auto gop_type_it = OpRegistry::grad_ops().find(op->type_);
PADDLE_ENFORCE(gop_type_it != OpRegistry::grad_ops().end(),
"Operator %s do not register gradient type", op->type_);
auto& grad_op_type = gop_type_it->second;
auto it = OpRegistry::op_info_map().find(op->Type());
PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
"'%s' has not been registered.", op->Type());
PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.",
op->Type());
std::string grad_op_type = it->second.grad_op_type_;
PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.",
op->Type());
OperatorBase::VarNameMap inputs;
OperatorBase::VarNameMap outputs;
TransOpArg(op, &inputs, OpArgType::IN, false); // I
TransOpArg(op, &inputs, OpArgType::OUT, false); // O
TransOpArg(op, &inputs, OpArgType::OUT, true); // OG
TransOpArg(op, &outputs, OpArgType::IN, true); // IG
auto gop_it = OpRegistry::op_creators().find(grad_op_type);
PADDLE_ENFORCE(gop_it != OpRegistry::op_creators().end(),
"Operator %s 's Gradient %s's creator cannot be found",
op->type_, grad_op_type);
TransOpArg(op, OpArgType::IN, false, &inputs); // I
TransOpArg(op, OpArgType::OUT, false, &inputs); // O
TransOpArg(op, OpArgType::OUT, true, &inputs); // OG
TransOpArg(op, OpArgType::IN, true, &outputs); // IG
return gop_it->second(grad_op_type, inputs, outputs, op->attrs_);
it = OpRegistry::op_info_map().find(grad_op_type);
PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
"'%s' has not been registered.", grad_op_type);
return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs());
}
} // namespace framework

@ -8,14 +8,6 @@ USE_OP(add_two);
namespace paddle {
namespace framework {
class NOP : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope,
const platform::DeviceContext &dev_ctx) const override {}
};
class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
public:
MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
@ -52,8 +44,8 @@ TEST(GradOpBuilder, AddTwo) {
"add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_add_op =
f::OpRegistry::CreateGradOp(*add_op);
EXPECT_EQ(grad_add_op->inputs_.size(), 4UL);
EXPECT_EQ(grad_add_op->outputs_.size(), 2UL);
EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
EXPECT_EQ(grad_add_op->Outputs().size(), 2UL);
EXPECT_EQ(grad_add_op->Input("X"), "x");
EXPECT_EQ(grad_add_op->Input("Y"), "y");
EXPECT_EQ(grad_add_op->Input("Out"), "out");
@ -62,10 +54,8 @@ TEST(GradOpBuilder, AddTwo) {
EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
}
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker);
REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP);
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
TEST(GradOpBuilder, MutiInOut) {
std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
@ -76,7 +66,7 @@ TEST(GradOpBuilder, MutiInOut) {
std::shared_ptr<f::OperatorBase> grad_test_op =
f::OpRegistry::CreateGradOp(*test_op);
ASSERT_EQ(grad_test_op->inputs_.size(), 3UL + 2UL + 2UL);
ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
@ -90,7 +80,7 @@ TEST(GradOpBuilder, MutiInOut) {
std::vector<std::string>(
{f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
ASSERT_EQ(grad_test_op->outputs_.size(), 3UL);
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>({f::GradVarName("in2_1"),
@ -109,7 +99,7 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
f::OpRegistry::CreateGradOp(*test_op);
// 'In2' and 'Out2' are ignored in gradient calculating
ASSERT_EQ(grad_test_op->inputs_.size(), 2UL + 1UL + 2UL);
ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
std::vector<std::string>({"in3_1", "in3_2"}));
@ -121,7 +111,7 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
f::GradVarName("out2"));
ASSERT_EQ(grad_test_op->outputs_.size(), 3UL);
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>(

@ -17,5 +17,49 @@ limitations under the License. */
#include <vector>
namespace paddle {
namespace framework {} // namespace framework
namespace framework {
std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
const VarNameMap& inputs,
const VarNameMap& outputs,
AttributeMap attrs) {
auto it = op_info_map().find(type);
PADDLE_ENFORCE(it != op_info_map().end(),
"Operator '%s' has not been registered.", type);
it->second.checker_->Check(attrs);
auto op = it->second.creator_(type, inputs, outputs, attrs);
return std::shared_ptr<OperatorBase>(op);
}
std::shared_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
AttributeMap attrs;
for (auto& attr : op_desc.attrs()) {
attrs[attr.name()] = GetAttrValue(attr);
}
return CreateOp(op_desc.type(), inputs, outputs, attrs);
}
OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
VarNameMap ret_val;
for (auto& var : op_desc_vars) {
auto& var_names = ret_val[var.parameter()];
auto& var_names_in_proto = var.arguments();
var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
std::back_inserter(var_names));
}
return ret_val;
}
std::shared_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
return grad_op;
}
} // namespace framework
} // namespace paddle

File diff suppressed because it is too large Load Diff

@ -59,11 +59,10 @@ static void BuildVar(const std::string& param_name,
var->add_arguments(arg_name);
}
}
REGISTER_OP(cos_sim, paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker);
REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
paddle::framework::MyTestOpProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(cos_sim, paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp,
paddle::framework::MyTestOpProtoAndCheckerMaker);
TEST(OpRegistry, CreateOp) {
paddle::framework::OpDesc op_desc;

@ -33,14 +33,6 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
}
#endif
static std::unordered_map<std::string, OpProto>* g_op_protos = nullptr;
std::unordered_map<std::string, OpProto>& OpProtos() {
if (g_op_protos == nullptr) {
g_op_protos = new std::unordered_map<std::string, OpProto>();
}
return *g_op_protos;
}
const std::string& OperatorBase::Input(const std::string& name) const {
auto& ins = Inputs(name);
PADDLE_ENFORCE_EQ(ins.size(), 1UL,
@ -149,14 +141,18 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
}
return ret_val;
}
auto it = OpProtos().find(type_);
auto it = OpRegistry::op_info_map().find(type_);
PADDLE_ENFORCE(
it != OpProtos().end(),
it != OpRegistry::op_info_map().end(),
"Operator %s not registered, cannot figure out intermediate outputs",
type_);
PADDLE_ENFORCE(
it->second.proto_ != nullptr,
"Operator %s has no OpProto, cannot figure out intermediate outputs",
type_);
// get all OpProto::Var for outputs
for (auto& o : it->second.outputs()) {
for (auto& o : it->second.proto_->outputs()) {
// ignore all intermediate output
if (o.intermediate()) continue;
auto out = outputs_.find(o.name());
@ -168,5 +164,43 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
return ret_val;
}
void OpProtoAndCheckerMaker::Validate() {
validated_ = true;
CheckNoDuplicatedInOutAttrs();
}
OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddInput(
const std::string& name, const std::string& comment) {
auto* input = proto_->add_inputs();
input->set_name(name);
input->set_comment(comment);
return OpProtoAndCheckerMaker::VariableBuilder{input};
}
OpProtoAndCheckerMaker::VariableBuilder OpProtoAndCheckerMaker::AddOutput(
const std::string& name, const std::string& comment) {
auto* output = proto_->add_outputs();
output->set_name(name);
output->set_comment(comment);
return OpProtoAndCheckerMaker::VariableBuilder{output};
}
void OpProtoAndCheckerMaker::CheckNoDuplicatedInOutAttrs() {
std::unordered_set<std::string> names;
auto checker = [&](const std::string& name) {
PADDLE_ENFORCE(!names.count(name), "[%s] is duplicated", name);
names.insert(name);
};
for (auto& attr : proto_->attrs()) {
checker(attr.name());
}
for (auto& input : proto_->inputs()) {
checker(input.name());
}
for (auto& output : proto_->outputs()) {
checker(output.name());
}
}
} // namespace framework
} // namespace paddle

@ -50,8 +50,6 @@ inline std::string GradVarName(const std::string& var_name) {
return var_name + kGradVarSuffix;
}
extern std::unordered_map<std::string, OpProto>& OpProtos();
class OperatorBase;
class InferShapeContext;
class ExecutionContext;
@ -69,10 +67,6 @@ class OperatorBase {
OperatorBase(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const AttributeMap& attrs);
OperatorBase(const OperatorBase& o) = delete;
OperatorBase& operator=(const OperatorBase& o) = delete;
OperatorBase(OperatorBase&& o) = delete;
virtual ~OperatorBase() {}
template <typename T>
@ -99,6 +93,8 @@ class OperatorBase {
/// rename inputs outputs name
void Rename(const std::string& old_name, const std::string& new_name);
const VarNameMap& Inputs() const { return inputs_; }
const VarNameMap& Outputs() const { return outputs_; }
//! Get a input with argument's name described in `op_proto`
const std::string& Input(const std::string& name) const;
//! Get a input which has multiple variables.
@ -112,13 +108,18 @@ class OperatorBase {
virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
std::string Type() const { return type_; }
const std::string& Type() const { return type_; }
void SetType(const std::string& type) { type_ = type; }
const AttributeMap& Attrs() const { return attrs_; }
public:
// Return a new operator instance, which is as same as this.
// Use unique_ptr to prevent caller forget to delete this pointer.
virtual std::unique_ptr<OperatorBase> Clone() const = 0;
protected:
std::string type_;
// NOTE: in case of OpGrad, inputs_ contains:
// I (Inputs)
// I (Inputs)opear
// O (Outputs)
// OG (Output Gradients)
VarNameMap inputs_;
@ -129,6 +130,102 @@ class OperatorBase {
AttributeMap attrs_;
};
// Macro for define a clone method.
// If you are writing an kernel operator, `Clone` will be defined when you
// register it. i.e. `Clone` method is not needed to define by yourself.
#define DEFINE_OP_CLONE_METHOD(CLS) \
std::unique_ptr<OperatorBase> Clone() const final { \
return std::unique_ptr<OperatorBase>(new CLS(*this)); \
}
// Macro for define a default constructor for Operator.
// You can also use
// using PARENT_CLASS::PARENT_CLASS;
// to use parent's constructor.
#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS) \
CLS(const std::string& type, const VarNameMap& inputs, \
const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
: PARENT_CLS(type, inputs, outputs, attrs) {}
class NOP : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {}
std::unique_ptr<OperatorBase> Clone() const override {
return std::unique_ptr<OperatorBase>(new NOP(*this));
}
};
// this class not only make proto but also init attribute checkers.
class OpProtoAndCheckerMaker {
public:
OpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
: proto_(proto), op_checker_(op_checker) {}
~OpProtoAndCheckerMaker() {
PADDLE_ENFORCE(validated_, "should call Validate after build");
}
void Validate();
protected:
struct VariableBuilder {
OpProto::Var* var_;
VariableBuilder& AsDuplicable() {
var_->set_duplicable(true);
return *this;
}
VariableBuilder& AsIntermediate() {
var_->set_intermediate(true);
return *this;
}
// TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
// means that input/output is not needed when calculate gradient. It does
// not mean no gradient when backward. It should be changed soon.
VariableBuilder& AsNoGradient() {
var_->set_no_gradient(true);
return *this;
}
};
VariableBuilder AddInput(const std::string& name, const std::string& comment);
VariableBuilder AddOutput(const std::string& name,
const std::string& comment);
template <typename T>
TypedAttrChecker<T>& AddAttr(const std::string& name,
const std::string& comment,
bool generated = false) {
auto* attr = proto_->add_attrs();
attr->set_name(name);
attr->set_comment(comment);
attr->set_generated(generated);
attr->set_type(AttrTypeID<T>());
return op_checker_->AddAttrChecker<T>(name);
}
void AddComment(const std::string& comment) { proto_->set_comment(comment); }
private:
void CheckNoDuplicatedInOutAttrs();
OpProto* proto_;
OpAttrChecker* op_checker_;
bool validated_{false};
};
class NOPMaker : public OpProtoAndCheckerMaker {
public:
NOPMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {}
};
class InferShapeContext {
public:
InferShapeContext(const OperatorBase& op, const Scope& scope)
@ -210,7 +307,7 @@ class InferShapeContext {
[&](const std::string& sub_name) {
auto var = scope_.FindVar(sub_name);
PADDLE_ENFORCE_NOT_NULL(
var, "MultiOutput(%s:%s) should not be nullptr", name,
var, "MultiOutput(%s:%s) should not be nullptr.", name,
sub_name);
return var->GetMutable<T>();
});

@ -65,8 +65,9 @@ static void BuildVar(const std::string& param_name,
}
}
REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
TEST(OperatorBase, all) {
paddle::framework::OpDesc op_desc;
@ -184,8 +185,9 @@ class CPUKernalMultiInputsTest : public OpKernel {
} // namespace framework
} // namespace paddle
REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
op_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestProtoAndCheckerMaker);
REGISTER_OP_CPU_KERNEL(op_with_kernel,
paddle::framework::CPUKernelTest<float, float>);
@ -210,8 +212,9 @@ TEST(OpKernel, all) {
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
}
REGISTER_OP(op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
paddle::framework::CPUKernalMultiInputsTest);
@ -242,3 +245,21 @@ TEST(OpKernel, multi_inputs) {
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cpu_device_context);
}
class OperatorClone : public paddle::framework::OperatorBase {
public:
DEFINE_OP_CLONE_METHOD(OperatorClone);
OperatorClone(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs,
const paddle::framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
void InferShape(const paddle::framework::Scope& scope) const override {}
void Run(const paddle::framework::Scope& scope,
const paddle::platform::DeviceContext& dev_ctx) const override {}
};
TEST(Operator, Clone) {
OperatorClone a("ABC", {}, {}, {});
auto b = a.Clone();
ASSERT_EQ(a.Type(), b->Type());
}

@ -20,6 +20,7 @@ limitations under the License. */
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor_py.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "paddle/string/to_string.h"
@ -30,8 +31,8 @@ limitations under the License. */
namespace py = pybind11;
USE_OP(add_two);
USE_CPU_OP(onehot_cross_entropy);
USE_NO_GRAD_OP(sgd);
USE_CPU_ONLY_OP(onehot_cross_entropy);
USE_OP(sgd);
USE_OP(mul);
USE_OP(mean);
USE_OP(sigmoid);
@ -53,15 +54,15 @@ void ExposeOperator(ClassType &m) {
.def("run", &ClassType::type::Run)
.def("type",
[](const typename ClassType::type &op) -> std::string {
return op.type_;
return op.Type();
})
.def("outputs",
[](const typename ClassType::type &op)
-> std::map<std::string, std::vector<std::string>> {
return op.outputs_;
return op.Outputs();
})
.def("inputs",
[](const typename ClassType::type &op) { return op.inputs_; })
[](const typename ClassType::type &op) { return op.Inputs(); })
.def("__str__", &ClassType::type::DebugString)
.def("no_intermediate_outputs",
[](const typename ClassType::type &op) {
@ -160,13 +161,16 @@ All parameter, weight, gradient are variables in Paddle.
//! @note: Be careful! PyBind will return std::string as an unicode, not
//! Python str. If you want a str object, you should cast them in Python.
m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
auto &protos = OpProtos();
auto &op_info_map = OpRegistry::op_info_map();
std::vector<py::bytes> ret_values;
for (auto it = protos.begin(); it != protos.end(); ++it) {
PADDLE_ENFORCE(it->second.IsInitialized(),
"OpProto must all be initialized");
for (auto it = op_info_map.begin(); it != op_info_map.end(); ++it) {
const OpProto *proto = it->second.proto_;
if (proto == nullptr) {
continue;
}
PADDLE_ENFORCE(proto->IsInitialized(), "OpProto must all be initialized");
std::string str;
PADDLE_ENFORCE(it->second.SerializeToString(&str),
PADDLE_ENFORCE(proto->SerializeToString(&str),
"Serialize OpProto Error. This could be a bug of Paddle.");
ret_values.push_back(py::bytes(str));
}
@ -229,7 +233,7 @@ All parameter, weight, gradient are variables in Paddle.
net.def_static("create",
[]() -> std::shared_ptr<operators::NetOp> {
auto retv = std::make_shared<operators::NetOp>();
retv->type_ = "plain_net";
retv->SetType("plain_net");
return retv;
})
.def("add_op", &operators::NetOp::AddOp)
@ -238,6 +242,11 @@ All parameter, weight, gradient are variables in Paddle.
const std::shared_ptr<operators::NetOp> &net) -> void {
self.AddOp(std::static_pointer_cast<OperatorBase>(net));
})
.def("add_op",
[](operators::NetOp &self,
const std::shared_ptr<operators::RecurrentOp> &rnn) -> void {
self.AddOp(std::static_pointer_cast<OperatorBase>(rnn));
})
.def("complete_add_op", &operators::NetOp::CompleteAddOp)
.def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
self->CompleteAddOp();
@ -245,6 +254,29 @@ All parameter, weight, gradient are variables in Paddle.
ExposeOperator(net);
// recurrent_op
py::class_<operators::RecurrentOp, std::shared_ptr<operators::RecurrentOp>>
rnn(m, "RecurrentOp");
rnn.def_static(
"create",
[](py::bytes protobin) -> std::shared_ptr<operators::RecurrentOp> {
OpDesc desc;
PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
"Cannot parse user input to OpDesc");
PADDLE_ENFORCE(desc.IsInitialized(),
"User OpDesc is not initialized, reason %s",
desc.InitializationErrorString());
auto rnn_op = OpRegistry::CreateOp(desc);
return std::dynamic_pointer_cast<operators::RecurrentOp>(rnn_op);
})
.def("set_stepnet",
[](operators::RecurrentOp &self,
const std::shared_ptr<operators::NetOp> &net) -> void {
self.set_stepnet(net);
});
ExposeOperator(rnn);
m.def("unique_integer", UniqueIntegerGenerator);
m.def("is_compile_gpu", IsCompileGPU);

@ -1,7 +1,7 @@
add_subdirectory(detail)
cc_library(memory SRCS memory.cc)
cc_library(memcpy SRCS memcpy.cc DEPS device_context)
cc_library(memcpy SRCS memcpy.cc)
cc_library(paddle_memory
DEPS

@ -16,8 +16,6 @@ limitations under the License. */
#include <cstring> // for memcpy
#include "paddle/platform/device_context.h"
namespace paddle {
namespace memory {

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

Loading…
Cancel
Save