提交 6e291b34 编写于 作者: C caoying03

Merge branch 'develop' into add_nest_sequence_select

......@@ -27,7 +27,7 @@ RUN apt-get update && \
git python-pip python-dev openssh-server bison \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
curl sed grep graphviz libjpeg-dev zlib1g-dev \
python-numpy python-matplotlib gcc g++ \
python-numpy python-matplotlib gcc-4.8 g++-4.8 \
automake locales clang-format-3.8 swig doxygen cmake \
liblapack-dev liblapacke-dev libboost-dev \
clang-3.8 llvm-3.8 libclang-3.8-dev \
......
......@@ -9,6 +9,11 @@ function(CheckCompilerCXX11Flag)
if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
endif()
# 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()
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.
......
......@@ -104,6 +104,11 @@ cross_channel_norm
------------------
.. autoclass:: paddle.v2.layer.cross_channel_norm
:noindex:
row_l2_norm
-----------
.. autoclass:: paddle.v2.layer.row_l2_norm
:noindex:
Recurrent Layers
================
......@@ -252,6 +257,11 @@ seq_concat
.. autoclass:: paddle.v2.layer.seq_concat
:noindex:
sub_nested_seq
--------------
.. autoclass:: paddle.v2.layer.sub_nested_seq
:noindex:
Reshaping Layers
================
......@@ -320,6 +330,11 @@ scaling
.. autoclass:: paddle.v2.layer.scaling
:noindex:
clip
----
.. autoclass:: paddle.v2.layer.clip
:noindex:
slope_intercept
---------------
.. autoclass:: paddle.v2.layer.slope_intercept
......
......@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/memory/memcpy.h"
namespace paddle {
......@@ -62,9 +61,11 @@ inline T* Tensor::mutable_data(platform::Place place) {
if (platform::is_cpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), size));
} else if (platform::is_gpu_place(place)) {
#ifdef PADDLE_ONLY_CPU
PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
}
#ifndef PADDLE_ONLY_CPU
else if (platform::is_gpu_place(place)) {
#else
holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
boost::get<platform::GPUPlace>(place), size));
}
......
......@@ -20,16 +20,16 @@ namespace paddle {
namespace framework {
template <>
Eigen::DefaultDevice* ExecutionContext::GetEigenDevice<
Eigen::DefaultDevice& ExecutionContext::GetEigenDevice<
platform::CPUPlace, Eigen::DefaultDevice>() const {
return device_context_.get_eigen_device<Eigen::DefaultDevice>();
return *device_context_.get_eigen_device<Eigen::DefaultDevice>();
}
#ifndef PADDLE_ONLY_CPU
template <>
Eigen::GpuDevice*
Eigen::GpuDevice&
ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
return device_context_.get_eigen_device<Eigen::GpuDevice>();
return *device_context_.get_eigen_device<Eigen::GpuDevice>();
}
#endif
......
......@@ -253,7 +253,7 @@ class ExecutionContext : public OperatorContext {
template <typename PlaceType,
typename DeviceType =
typename EigenDeviceConverter<PlaceType>::EigenDeviceType>
DeviceType* GetEigenDevice() const;
DeviceType& GetEigenDevice() const;
platform::Place GetPlace() const { return device_context_.GetPlace(); }
......
/* 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 "Layer.h"
namespace paddle {
/**
* A layer for clipping the input value by the threshold.
* \f[
* out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right)
* \f]
*/
class ClipLayer : public Layer {
protected:
double min_;
double max_;
public:
explicit ClipLayer(const LayerConfig& config) : Layer(config) {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void forward(PassType passType) override;
void backward(const UpdateCallback& callback = nullptr) override;
};
REGISTER_LAYER(clip, ClipLayer);
bool ClipLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
Layer::init(layerMap, parameterMap);
CHECK_EQ(inputLayers_.size(), 1U);
auto layerConf = config_.inputs(0).clip_conf();
min_ = layerConf.min();
max_ = layerConf.max();
CHECK_LT(min_, max_);
return true;
}
void ClipLayer::forward(PassType passType) {
Layer::forward(passType);
MatrixPtr inV = getInputValue(0);
resetOutput(inV->getHeight(), inV->getWidth());
MatrixPtr outV = getOutputValue();
outV->copyFrom(*inV);
outV->clip(min_, max_);
}
void ClipLayer::backward(const UpdateCallback& callback) {
MatrixPtr inV = getInputValue(0);
MatrixPtr inG = getInputGrad(0);
if (inG) {
MatrixPtr outV = getOutputValue();
MatrixPtr outG = getOutputGrad();
MatrixPtr tmpMtx;
Matrix::resizeOrCreate(
tmpMtx, outG->getHeight(), outG->getWidth(), false, useGpu_);
tmpMtx->clipDerivative(*inV, min_, max_);
inG->addDotMul(*outG, *tmpMtx, 1, 1);
}
}
} // namespace paddle
/* 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 "Layer.h"
namespace paddle {
/**
* A layer for L2 normalization in each row,
* \f[
* out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}}
* \f]
* where the size of \f$in\f$ is (batchSize x dataDim),
* and the size of \f$out\f$ is (batchSize x dataDim).
*/
class RowL2NormLayer : public Layer {
protected:
MatrixPtr inSquare_;
MatrixPtr l2NormReciprocal_;
MatrixPtr dotSum_;
public:
explicit RowL2NormLayer(const LayerConfig& config) : Layer(config) {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void forward(PassType passType) override;
void backward(const UpdateCallback& callback = nullptr) override;
};
REGISTER_LAYER(row_l2_norm, RowL2NormLayer);
bool RowL2NormLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
Layer::init(layerMap, parameterMap);
CHECK_EQ(inputLayers_.size(), 1U);
return true;
}
void RowL2NormLayer::forward(PassType passType) {
Layer::forward(passType);
MatrixPtr inV = getInputValue(0);
/* malloc memory for the output_ if necessary */
size_t batchSize = inV->getHeight();
size_t dataDim = getSize();
CHECK_EQ(dataDim, inV->getWidth());
resetOutput(batchSize, dataDim);
MatrixPtr outV = getOutputValue();
Matrix::resizeOrCreate(inSquare_, batchSize, dataDim, false, useGpu_);
inV->square2(*inSquare_);
Matrix::resizeOrCreate(l2NormReciprocal_, batchSize, 1, false, useGpu_);
inSquare_->rowSum(*l2NormReciprocal_);
l2NormReciprocal_->sqrt2(*l2NormReciprocal_);
l2NormReciprocal_->scalarDiv(*l2NormReciprocal_, 1.0);
outV->rowScale(0, *inV, *l2NormReciprocal_);
}
void RowL2NormLayer::backward(const UpdateCallback& callback) {
MatrixPtr inV = getInputValue(0);
MatrixPtr inG = getInputGrad(0);
MatrixPtr outV = getOutputValue();
MatrixPtr outG = getOutputGrad();
size_t batchSize = inV->getHeight();
// inG[ij] += outG[ij] / l2NormReciprocal
// inG[ij] += -inV[ij] * l2NormReciprocal * l2NormReciprocal * DotMul(outG[i],
// inV[i])
if (inG) {
Matrix::resizeOrCreate(dotSum_, batchSize, 1, false, useGpu_);
dotSum_->zeroMem();
dotSum_->rowDotMul(0, *outG, *outV);
dotSum_->dotMul(*dotSum_, *l2NormReciprocal_);
dotSum_->dotMul(*dotSum_, *l2NormReciprocal_);
inSquare_->rowScale(0, *inV, *dotSum_);
inG->sub(*inSquare_);
inG->addRowScale(0, *outG, *l2NormReciprocal_);
}
}
} // namespace paddle
......@@ -1961,6 +1961,36 @@ TEST(Layer, SubNestedSequenceLayer) {
}
}
TEST(Layer, ClipLayer) {
const size_t batchSize = 128;
const size_t size = 512;
TestConfig config;
config.layerConfig.set_type("clip");
config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
LayerInputConfig* input = config.layerConfig.add_inputs();
ClipConfig* layerConf = input->mutable_clip_conf();
double p1 = std::rand() / (double)RAND_MAX;
double p2 = std::rand() / (double)RAND_MAX;
layerConf->set_min(std::min(p1, p2));
layerConf->set_max(std::max(p1, p2));
for (auto useGpu : {false, true}) {
testLayerGrad(config, "clip", batchSize, false, useGpu, false);
}
}
TEST(Layer, RowL2NormLayer) {
const size_t batchSize = 128;
const size_t size = 512;
TestConfig config;
config.layerConfig.set_type("row_l2_norm");
config.layerConfig.set_size(size);
config.inputDefs.push_back({INPUT_DATA, "input", size, 0});
config.layerConfig.add_inputs();
for (auto useGpu : {false, true}) {
testLayerGrad(config, "row_l2_norm", batchSize, false, useGpu, false);
}
}
int main(int argc, char** argv) {
testing::InitGoogleTest(&argc, argv);
initMain(argc, argv);
......
......@@ -442,6 +442,12 @@ DEFINE_MATRIX_UNARY_PARAMETER_OP(Clip, TWO_PARAMETER,
template<class T>
void BaseMatrixT<T>::clip(T p1, T p2) { applyUnary(unary::Clip<T>(p1, p2)); }
DEFINE_MATRIX_BINARY_PARAMETER_OP(ClipDerivative, TWO_PARAMETER, a = b < p1 ? 0 : (b > p2 ? 0 : 1));
template<class T>
void BaseMatrixT<T>::clipDerivative(BaseMatrixT& b, T p1, T p2) {
applyBinary(binary::ClipDerivative<T>(p1, p2), b);
}
DEFINE_MATRIX_UNARY_PARAMETER_OP(BiggerThanScalar, ONE_PARAMETER,
a = a > p ? 1.0f : 0.0f);
template<class T>
......
......@@ -488,6 +488,13 @@ public:
*/
void clip(T p1, T p2);
/**
* this = b < low ? 0 : 1
*
* this = b > high ? 0 : 1
*/
void clipDerivative(BaseMatrixT& b, T p1, T p2);
/**
* @code
* a = a > p ? 1.0f : 0.0f
......
......@@ -50,10 +50,6 @@ The equation is: Out = X + Y
class AddOpGrad : public OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {}
std::string DebugString() const override {
LOG(INFO) << "AddOpGrad";
return "";
}
};
} // namespace operators
......
#define EIGEN_USE_GPU
#include "paddle/framework/op_registry.h"
#include "paddle/operators/add_op.h"
......
......@@ -28,8 +28,7 @@ public:
output->mutable_data<T>(context.GetPlace());
EigenVector<T>::Flatten(*output).device(
*(context.GetEigenDevice<Place>())) =
EigenVector<T>::Flatten(*output).device(context.GetEigenDevice<Place>()) =
framework::EigenVector<T>::Flatten(*input0) +
framework::EigenVector<T>::Flatten(*input1);
}
......
#define EIGEN_USE_GPU
#include "paddle/operators/cross_entropy_op.h"
REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
......
......@@ -27,7 +27,7 @@ public:
output->mutable_data<T>(context.GetPlace());
EigenScalar<T>::From(*output).device(*(context.GetEigenDevice<Place>())) =
EigenScalar<T>::From(*output).device(context.GetEigenDevice<Place>()) =
EigenVector<T>::Flatten(*input).mean();
}
};
......
......@@ -12,6 +12,7 @@
See the License for the specific language governing permissions and
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/operators/mul_op.h"
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);
\ No newline at end of file
......@@ -29,7 +29,7 @@ public:
auto output = context.Output<Tensor>(0);
output->mutable_data<T>(context.GetPlace());
EigenMatrix<T>::From(*output).device(*(context.GetEigenDevice<Place>())) =
EigenMatrix<T>::From(*output).device(context.GetEigenDevice<Place>()) =
EigenMatrix<T>::From(*context.Input<Tensor>("X"))
.contract(EigenMatrix<T>::From(*context.Input<Tensor>("Y")),
dim_pair);
......
#define EIGEN_USE_GPU
#include "paddle/operators/rowwise_add_op.h"
REGISTER_OP_GPU_KERNEL(rowwise_add,
......
......@@ -33,7 +33,7 @@ public:
const int rest_size = input.size() / bias_size;
Eigen::DSizes<int, 1> one_d(input.size());
Eigen::DSizes<int, 1> bcast(rest_size);
output.reshape(one_d).device(*(context.GetEigenDevice<Place>())) =
output.reshape(one_d).device(context.GetEigenDevice<Place>()) =
input.reshape(one_d) + bias.broadcast(bcast).reshape(one_d);
}
};
......
#define EIGEN_USE_GPU
#include "paddle/operators/sgd_op.h"
REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);
\ No newline at end of file
......@@ -29,7 +29,7 @@ public:
param_out->mutable_data<T>(ctx.GetPlace());
EigenVector<T>::Flatten(*param_out).device(*(ctx.GetEigenDevice<Place>())) =
EigenVector<T>::Flatten(*param_out).device(ctx.GetEigenDevice<Place>()) =
EigenVector<T>::Flatten(*param) - lr * EigenVector<T>::Flatten(*grad);
}
};
......
#define EIGEN_USE_GPU
#include "paddle/operators/sigmoid_op.h"
REGISTER_OP_GPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::GPUPlace, float>);
......@@ -27,8 +27,7 @@ public:
auto output = context.Output<Tensor>(0);
output->mutable_data<T>(context.GetPlace());
EigenVector<T>::Flatten(*output).device(
*(context.GetEigenDevice<Place>())) =
EigenVector<T>::Flatten(*output).device(context.GetEigenDevice<Place>()) =
1.0 / (1.0 + (-1.0 * EigenVector<T>::Flatten(*input)).exp());
}
};
......
#define EIGEN_USE_GPU
#include "paddle/framework/op_registry.h"
#include "paddle/operators/softmax_op.h"
......
......@@ -46,9 +46,9 @@ public:
.reshape(batch_by_one)
.broadcast(one_by_class));
softmax.device(*(context.GetEigenDevice<Place>())) = shifted_logits.exp();
softmax.device(context.GetEigenDevice<Place>()) = shifted_logits.exp();
softmax.device(*(context.GetEigenDevice<Place>())) =
softmax.device(context.GetEigenDevice<Place>()) =
(softmax *
softmax.sum(along_class)
.inverse()
......
......@@ -144,12 +144,12 @@ inline void throw_on_error(T e) {
throw_on_error(e, "");
}
#define PADDLE_THROW(...) \
do { \
throw ::paddle::platform::EnforceNotMet( \
std::make_exception_ptr( \
std::runtime_error(string::Sprintf(__VA_ARGS__))), \
__FILE__, __LINE__); \
#define PADDLE_THROW(...) \
do { \
throw ::paddle::platform::EnforceNotMet( \
std::make_exception_ptr( \
std::runtime_error(paddle::string::Sprintf(__VA_ARGS__))), \
__FILE__, __LINE__); \
} while (0)
#define PADDLE_ENFORCE(...) \
......
cc_library(paddle_pybind SHARED
SRCS pybind.cc
DEPS pybind python
DEPS pybind python backward
fc_op
sgd_op
add_op
......
......@@ -16,10 +16,13 @@ limitations under the License. */
#include <fstream>
#include <vector>
#include "paddle/framework/backward.h"
#include "paddle/framework/net.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/scope.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "paddle/pybind/tensor_bind.h"
#include "pybind11/numpy.h"
#include "pybind11/pybind11.h"
......@@ -43,6 +46,10 @@ template <typename ClassType>
void ExposeOperator(ClassType& m) {
m.def("infer_shape", &ClassType::type::InferShape)
.def("run", &ClassType::type::Run)
.def("type",
[](const typename ClassType::type& op) -> std::string {
return op.type_;
})
.def("outputs",
[](const typename ClassType::type& op) -> std::vector<std::string> {
return op.outputs_;
......@@ -55,6 +62,14 @@ static size_t UniqueIntegerGenerator() {
return generator.fetch_add(1);
}
bool IsCompileGPU() {
#ifdef PADDLE_ONLY_CPU
return false;
#else
return true;
#endif
}
PYBIND11_PLUGIN(core) {
py::module m("core", "C++ core of PaddlePaddle");
......@@ -69,15 +84,27 @@ PYBIND11_PLUGIN(core) {
self.Resize(pd::make_ddim(dim));
})
.def("alloc_float",
[](pd::Tensor& self) {
self.mutable_data<float>(paddle::platform::CPUPlace());
[](pd::Tensor& self, paddle::platform::GPUPlace& place) {
self.mutable_data<float>(place);
})
.def("alloc_float",
[](pd::Tensor& self, paddle::platform::CPUPlace& place) {
self.mutable_data<float>(place);
})
.def("alloc_int",
[](pd::Tensor& self) {
self.mutable_data<int>(paddle::platform::CPUPlace());
[](pd::Tensor& self, paddle::platform::CPUPlace& place) {
self.mutable_data<int>(place);
})
.def("set", paddle::pybind::PyTensorSetFromArray<float>)
.def("set", paddle::pybind::PyTensorSetFromArray<int>)
.def("alloc_int",
[](pd::Tensor& self, paddle::platform::GPUPlace& place) {
self.mutable_data<int>(place);
})
.def("set", paddle::pybind::PyCPUTensorSetFromArray<float>)
.def("set", paddle::pybind::PyCPUTensorSetFromArray<int>)
#ifndef PADDLE_ONLY_CPU
.def("set", paddle::pybind::PyCUDATensorSetFromArray<float>)
.def("set", paddle::pybind::PyCUDATensorSetFromArray<int>)
#endif
.def("shape",
[](pd::Tensor& self) { return pd::vectorize(self.dims()); });
......@@ -136,11 +163,27 @@ All parameter, weight, gradient are variables in Paddle.
"The module will return special predefined variable name in Paddle")
.def("empty", pd::OperatorBase::EMPTY_VAR_NAME)
.def("temp", pd::OperatorBase::TMP_VAR_NAME);
// clang-format off
py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
.def_static("cpu_context", []() -> paddle::platform::DeviceContext* {
return new paddle::platform::CPUDeviceContext();
});
.def_static("create",
[](paddle::platform::CPUPlace& place)
-> paddle::platform::DeviceContext* {
return new paddle::platform::CPUDeviceContext();
})
.def_static("create",
[](paddle::platform::GPUPlace& place)
-> paddle::platform::DeviceContext* {
#ifdef PADDLE_ONLY_CPU
PADDLE_THROW("GPUPlace is not supported in CPU device.");
#else
return new paddle::platform::CUDADeviceContext(place);
#endif
});
// clang-format on
py::class_<paddle::platform::GPUPlace>(m, "GPUPlace").def(py::init<int>());
py::class_<paddle::platform::CPUPlace>(m, "CPUPlace").def(py::init<>());
py::class_<pd::OperatorBase, std::shared_ptr<pd::OperatorBase>> operator_base(
m, "Operator");
......@@ -154,6 +197,13 @@ All parameter, weight, gradient are variables in Paddle.
desc.InitializationErrorString());
return pd::OpRegistry::CreateOp(desc);
});
operator_base.def("backward",
[](const pd::OperatorBase& forwardOp,
const std::unordered_set<std::string>& no_grad_vars) {
return pd::Backward(forwardOp, no_grad_vars);
});
ExposeOperator(operator_base);
py::class_<pd::NetOp, std::shared_ptr<pd::NetOp>> net(m, "Net");
......@@ -176,5 +226,7 @@ All parameter, weight, gradient are variables in Paddle.
m.def("unique_integer", UniqueIntegerGenerator);
m.def("is_compile_gpu", IsCompileGPU);
return m.ptr();
}
......@@ -13,9 +13,11 @@
limitations under the License. */
#pragma once
#include <paddle/framework/tensor.h>
#include <pybind11/numpy.h>
#include <pybind11/pybind11.h>
#include <string>
#include "paddle/framework/tensor.h"
#include "paddle/memory/memcpy.h"
#include "pybind11/numpy.h"
#include "pybind11/pybind11.h"
namespace py = pybind11;
......@@ -40,9 +42,6 @@ template <size_t I, typename... ARGS>
struct CastToPyBufferImpl<true, I, ARGS...> {
using CUR_TYPE = typename std::tuple_element<I, std::tuple<ARGS...>>::type;
py::buffer_info operator()(framework::Tensor &tensor) {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(tensor.holder_->place()),
"Only CPU tensor can cast to numpy array");
if (std::type_index(typeid(CUR_TYPE)) == tensor.holder_->type()) {
auto dim_vec = framework::vectorize(tensor.dims());
std::vector<size_t> dims_outside;
......@@ -56,12 +55,17 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
strides[i - 1] = sizeof(CUR_TYPE) * prod;
prod *= dims_outside[i - 1];
}
framework::Tensor dst_tensor;
if (paddle::platform::is_gpu_place(tensor.holder_->place())) {
dst_tensor.CopyFrom<CUR_TYPE>(tensor, platform::CPUPlace());
} else if (paddle::platform::is_cpu_place(tensor.holder_->place())) {
dst_tensor = tensor;
}
return py::buffer_info(
tensor.mutable_data<CUR_TYPE>(tensor.holder_->place()),
dst_tensor.mutable_data<CUR_TYPE>(dst_tensor.holder_->place()),
sizeof(CUR_TYPE),
py::format_descriptor<CUR_TYPE>::format(),
(size_t)framework::arity(tensor.dims()),
(size_t)framework::arity(dst_tensor.dims()),
dims_outside,
strides);
} else {
......@@ -77,9 +81,10 @@ inline py::buffer_info CastToPyBuffer(framework::Tensor &tensor) {
}
template <typename T>
void PyTensorSetFromArray(
void PyCPUTensorSetFromArray(
framework::Tensor &self,
py::array_t<T, py::array::c_style | py::array::forcecast> array) {
py::array_t<T, py::array::c_style | py::array::forcecast> array,
paddle::platform::CPUPlace &place) {
std::vector<int> dims;
dims.reserve(array.ndim());
for (size_t i = 0; i < array.ndim(); ++i) {
......@@ -87,9 +92,28 @@ void PyTensorSetFromArray(
}
self.Resize(framework::make_ddim(dims));
auto *dst = self.mutable_data<T>(paddle::platform::CPUPlace());
auto *dst = self.mutable_data<T>(place);
std::memcpy(dst, array.data(), sizeof(T) * array.size());
}
#ifndef PADDLE_ONLY_CPU
template <typename T>
void PyCUDATensorSetFromArray(
framework::Tensor &self,
py::array_t<T, py::array::c_style | py::array::forcecast> array,
paddle::platform::GPUPlace &place) {
std::vector<int> dims;
dims.reserve(array.ndim());
for (size_t i = 0; i < array.ndim(); ++i) {
dims.push_back((int)array.shape()[i]);
}
self.Resize(framework::make_ddim(dims));
auto *dst = self.mutable_data<T>(place);
paddle::platform::GpuMemcpySync(
dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice);
}
#endif
} // namespace pybind
} // namespace paddle
......@@ -148,7 +148,7 @@ cat >> /paddle/build/Dockerfile <<EOF
ADD *.deb /
# run paddle version to install python packages first
RUN apt-get update &&\
apt-get install -y python-pip && pip install -U pip && \
apt-get install -y wget python-pip && pip install -U pip && \
dpkg -i /*.deb ; apt-get install -f -y && \
apt-get clean -y && \
rm -f /*.deb && \
......
......@@ -298,6 +298,11 @@ message DetectionOutputConfig {
optional uint32 width = 9 [default = 1];
}
message ClipConfig {
required double min = 1;
required double max = 2;
}
message LayerInputConfig {
required string input_layer_name = 1;
optional string input_parameter_name = 2;
......@@ -318,6 +323,7 @@ message LayerInputConfig {
optional RowConvConfig row_conv_conf = 15;
optional MultiBoxLossConfig multibox_loss_conf = 16;
optional DetectionOutputConfig detection_output_conf = 17;
optional ClipConfig clip_conf = 18;
}
message LayerConfig {
......
......@@ -2198,6 +2198,20 @@ class RowConvLayer(LayerBase):
self.create_input_parameter(0, psize, dims)
@config_layer('clip')
class ClipLayer(LayerBase):
def __init__(self, name, inputs, min, max, **xargs):
super(ClipLayer, self).__init__(name, 'clip', 0, inputs=inputs, **xargs)
config_assert(
len(self.inputs) == 1,
'ClipLayer must have one and only one input.')
config_assert(min < max, 'min must be less than max.')
input_layer = self.get_input_layer(0)
self.set_layer_size(input_layer.size)
self.config.inputs[0].clip_conf.min = min
self.config.inputs[0].clip_conf.max = max
# key: cost type
# value: cost class
g_cost_map = {}
......@@ -2773,6 +2787,16 @@ class SumToOneNormLayer(LayerBase):
self.set_layer_size(input_layer0.size)
@config_layer('row_l2_norm')
class RowL2NormLayer(LayerBase):
def __init__(self, name, inputs, **xargs):
super(RowL2NormLayer, self).__init__(
name, 'row_l2_norm', 0, inputs=inputs, **xargs)
config_assert(len(self.inputs) == 1, 'RowL2NormLayer must have 1 input')
input_layer = self.get_input_layer(0)
self.set_layer_size(input_layer.size)
@config_layer('cos_vm')
class CosSimVecMatLayer(LayerBase):
def __init__(self, name, size, inputs, cos_scale=1.0, device=None):
......
......@@ -76,6 +76,7 @@ __all__ = [
'trans_layer',
'rotate_layer',
'sum_to_one_norm_layer',
'row_l2_norm_layer',
'get_output_layer',
'LayerType',
'context_projection',
......@@ -129,6 +130,7 @@ __all__ = [
'gated_unit_layer',
'crop_layer',
'sub_nested_seq_layer',
'clip_layer',
'slice_projection',
]
......@@ -161,6 +163,7 @@ class LayerType(object):
BATCH_NORM_LAYER = 'batch_norm'
NORM_LAYER = 'norm'
SUM_TO_ONE_NORM_LAYER = 'sum_to_one_norm'
ROW_L2_NORM_LAYER = 'row_l2_norm'
ADDTO_LAYER = 'addto'
CONCAT_LAYER = 'concat'
......@@ -223,6 +226,7 @@ class LayerType(object):
PRELU = 'prelu'
CROP_LAYER = 'crop'
SUB_NESTED_SEQ = 'sub_nested_seq'
CLIP_LAYER = 'clip'
@staticmethod
def is_layer_type(type_name):
......@@ -2891,6 +2895,42 @@ def sum_to_one_norm_layer(input, name=None, layer_attr=None):
name, LayerType.SUM_TO_ONE_NORM_LAYER, parents=[input], size=input.size)
@wrap_name_default()
@layer_support()
def row_l2_norm_layer(input, name=None, layer_attr=None):
"""
A layer for L2-normalization in each row.
.. math::
out[i] = \frac{in[i]}{\sqrt{\sum_{k=1}^N in[k]^{2}}}
where the size of :math:`in` is (batchSize x dataDim) ,
and the size of :math:`out` is a (batchSize x dataDim) .
The example usage is:
.. code-block:: python
row_l2_norm_layer = row_l2_norm_layer(input=layer)
:param input: Input layer.
:type input: LayerOutput
:param name: Layer name.
:type name: basestring
:param layer_attr: extra layer attributes.
:type layer_attr: ExtraLayerAttribute.
:return: LayerOutput object.
:rtype: LayerOutput
"""
Layer(
name=name,
type=LayerType.ROW_L2_NORM_LAYER,
inputs=[input.name],
**ExtraAttr.to_kwargs(layer_attr))
return LayerOutput(
name, LayerType.ROW_L2_NORM_LAYER, parents=[input], size=input.size)
@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
......@@ -6089,3 +6129,36 @@ def sub_nested_seq_layer(input, name=None, top_k=1):
layer_type=LayerType.SUB_NESTED_SEQ,
parents=input,
size=l.config.size)
@wrap_name_default("clip")
def clip_layer(input, min, max, name=None):
"""
A layer for clipping the input value by the threshold.
.. math::
out[i] = \min\left(\max\left(in[i],p_{1}\right),p_{2}\right)
.. code-block:: python
clip = clip_layer(input=input_layer, min=-10, max=10)
:param name: The Layer Name.
:type name: basestring
:param input: The input layer.
:type input: LayerOutput.
:param min: The lower threshold for clipping.
:type min: double
:param max: The upper threshold for clipping.
:type max: double
:return: LayerOutput
"""
Layer(
name=name,
type=LayerType.CLIP_LAYER,
inputs=[input.name],
min=min,
max=max)
return LayerOutput(
name, LayerType.CLIP_LAYER, parents=[input], size=input.size)
......@@ -7,6 +7,7 @@ test_rnn_group shared_fc shared_lstm shared_gru test_cost_layers_with_weight
test_spp_layer test_bilinear_interp test_maxout test_bi_grumemory math_ops
test_seq_concat_reshape test_pad test_smooth_l1 test_multiplex_layer
test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_layer
test_recursive_topology test_gated_unit_layer test_seq_select_layers)
test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
test_seq_select_layers)
export whole_configs=(test_split_datasource)
type: "nn"
layers {
name: "input"
type: "data"
size: 300
active_type: ""
}
layers {
name: "__clip_0__"
type: "clip"
size: 300
active_type: ""
inputs {
input_layer_name: "input"
clip_conf {
min: -10
max: 10
}
}
}
input_layer_names: "input"
output_layer_names: "__clip_0__"
sub_models {
name: "root"
layer_names: "input"
layer_names: "__clip_0__"
input_layer_names: "input"
output_layer_names: "__clip_0__"
is_recurrent_layer_group: false
}
type: "nn"
layers {
name: "input"
type: "data"
size: 300
active_type: ""
}
layers {
name: "__row_l2_norm_layer_0__"
type: "row_l2_norm"
size: 300
active_type: ""
inputs {
input_layer_name: "input"
}
}
input_layer_names: "input"
output_layer_names: "__row_l2_norm_layer_0__"
sub_models {
name: "root"
layer_names: "input"
layer_names: "__row_l2_norm_layer_0__"
input_layer_names: "input"
output_layer_names: "__row_l2_norm_layer_0__"
is_recurrent_layer_group: false
}
from paddle.trainer_config_helpers import *
data = data_layer(name='input', size=300)
clip = clip_layer(input=data, min=-10, max=10)
outputs(clip)
from paddle.trainer_config_helpers import *
data = data_layer(name='input', size=300)
row_l2_norm = row_l2_norm_layer(input=data)
outputs(row_l2_norm)
......@@ -8,7 +8,6 @@ add_python_test(test_framework
test_fc_op.py
test_add_two_op.py
test_sgd_op.py
test_cross_entropy_op.py
test_mul_op.py
test_mean_op.py
test_sigmoid_op.py
......
......@@ -26,40 +26,45 @@ class OpTestMeta(type):
scope = core.Scope()
kwargs = dict()
places = []
places.append(core.CPUPlace())
if core.is_compile_gpu():
places.append(core.GPUPlace(0))
for in_name in func.all_input_args:
if hasattr(self, in_name):
kwargs[in_name] = in_name
var = scope.new_var(in_name).get_tensor()
arr = getattr(self, in_name)
var.set_dims(arr.shape)
var.set(arr)
else:
kwargs[in_name] = "@EMPTY@"
for place in places:
for in_name in func.all_input_args:
if hasattr(self, in_name):
kwargs[in_name] = in_name
var = scope.new_var(in_name).get_tensor()
arr = getattr(self, in_name)
var.set_dims(arr.shape)
var.set(arr, place)
else:
kwargs[in_name] = "@EMPTY@"
for out_name in func.all_output_args:
if hasattr(self, out_name):
kwargs[out_name] = out_name
scope.new_var(out_name).get_tensor()
for out_name in func.all_output_args:
if hasattr(self, out_name):
kwargs[out_name] = out_name
scope.new_var(out_name).get_tensor()
for attr_name in func.all_attr_args:
if hasattr(self, attr_name):
kwargs[attr_name] = getattr(self, attr_name)
for attr_name in func.all_attr_args:
if hasattr(self, attr_name):
kwargs[attr_name] = getattr(self, attr_name)
op = func(**kwargs)
op = func(**kwargs)
op.infer_shape(scope)
op.infer_shape(scope)
ctx = core.DeviceContext.cpu_context()
op.run(scope, ctx)
ctx = core.DeviceContext.create(place)
op.run(scope, ctx)
for out_name in func.all_output_args:
actual = numpy.array(scope.find_var(out_name).get_tensor())
expect = getattr(self, out_name)
# TODO(qijun) The default decimal is 7, but numpy.dot and eigen.mul
# has some diff, and could not pass unittest. So I set decimal 3 here.
# And I will check this in future.
numpy.testing.assert_almost_equal(actual, expect, decimal=3)
for out_name in func.all_output_args:
actual = numpy.array(scope.find_var(out_name).get_tensor())
expect = getattr(self, out_name)
# TODO(qijun) The default decimal is 7, but numpy.dot and eigen.mul
# has some diff, and could not pass unittest. So I set decimal 3 here.
# And I will check this in future.
numpy.testing.assert_almost_equal(actual, expect, decimal=3)
obj.test_all = test_all
return obj
import unittest
from op_test_util import OpTestMeta
import numpy
import paddle.v2.framework.core as core
import paddle.v2.framework.create_op_creation_methods as creation
from op_test_util import OpTestMeta
class TestAddOp(unittest.TestCase):
......@@ -8,10 +12,19 @@ class TestAddOp(unittest.TestCase):
def setUp(self):
self.type = "add_two"
self.X = numpy.random.random((342, 345)).astype("float32")
self.Y = numpy.random.random((342, 345)).astype("float32")
self.X = numpy.random.random((102, 105)).astype("float32")
self.Y = numpy.random.random((102, 105)).astype("float32")
self.Out = self.X + self.Y
class TestAddGradOp(unittest.TestCase):
def test_add_grad(self):
op = creation.op_creations.add_two(X="X", Y="Y", Out="Out")
backward_op = core.Operator.backward(op, set())
self.assertEqual(backward_op.type(), "add_two_grad")
expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).'''
self.assertEqual(expected, str(backward_op))
if __name__ == '__main__':
unittest.main()
......@@ -7,17 +7,19 @@ import paddle.v2.framework.create_op_creation_methods as creation
class TestFc(unittest.TestCase):
def test_fc(self):
scope = core.Scope()
place = core.CPUPlace()
x = scope.new_var("X")
x_tensor = x.get_tensor()
x_tensor.set_dims([1000, 784])
x_tensor.alloc_float()
x_tensor.alloc_float(place)
w = scope.new_var("W")
w_tensor = w.get_tensor()
w_tensor.set_dims([784, 100])
w_tensor.alloc_float()
w_tensor.alloc_float(place)
w_tensor.set(numpy.random.random((784, 100)).astype("float32"))
w_tensor.set(numpy.random.random((784, 100)).astype("float32"), place)
# Set a real numpy array here.
# x_tensor.set(numpy.array([]))
......@@ -32,7 +34,7 @@ class TestFc(unittest.TestCase):
op.infer_shape(scope)
self.assertEqual([1000, 100], tensor.shape())
ctx = core.DeviceContext.cpu_context()
ctx = core.DeviceContext.create(place)
op.run(scope, ctx)
......
......@@ -8,8 +8,8 @@ class TestMulOp(unittest.TestCase):
def setUp(self):
self.type = "mul"
self.X = np.random.random((32, 784)).astype("float32")
self.Y = np.random.random((784, 100)).astype("float32")
self.X = np.random.random((32, 84)).astype("float32")
self.Y = np.random.random((84, 100)).astype("float32")
self.Out = np.dot(self.X, self.Y)
......
......@@ -8,8 +8,8 @@ class TestRowwiseAddOp(unittest.TestCase):
def setUp(self):
self.type = "rowwise_add"
self.X = np.random.random((32, 784)).astype("float32")
self.b = np.random.random(784).astype("float32")
self.X = np.random.random((32, 84)).astype("float32")
self.b = np.random.random(84).astype("float32")
self.Out = np.add(self.X, self.b)
......
......@@ -8,8 +8,8 @@ class TestSGD(unittest.TestCase):
def setUp(self):
self.type = "sgd"
self.param = numpy.random.random((342, 345)).astype("float32")
self.grad = numpy.random.random((342, 345)).astype("float32")
self.param = numpy.random.random((102, 105)).astype("float32")
self.grad = numpy.random.random((102, 105)).astype("float32")
self.learning_rate = 0.1
self.param_out = self.param - self.learning_rate * self.grad
......
......@@ -7,16 +7,17 @@ class TestScope(unittest.TestCase):
def test_int_tensor(self):
scope = core.Scope()
var = scope.new_var("test_tensor")
place = core.CPUPlace()
tensor = var.get_tensor()
tensor.set_dims([1000, 784])
tensor.alloc_int()
tensor.alloc_int(place)
tensor_array = numpy.array(tensor)
self.assertEqual((1000, 784), tensor_array.shape)
tensor_array[3, 9] = 1
tensor_array[19, 11] = 2
tensor.set(tensor_array)
tensor.set(tensor_array, place)
tensor_array_2 = numpy.array(tensor)
self.assertEqual(1.0, tensor_array_2[3, 9])
......@@ -25,16 +26,18 @@ class TestScope(unittest.TestCase):
def test_float_tensor(self):
scope = core.Scope()
var = scope.new_var("test_tensor")
place = core.CPUPlace()
tensor = var.get_tensor()
tensor.set_dims([1000, 784])
tensor.alloc_float()
tensor.alloc_float(place)
tensor_array = numpy.array(tensor)
self.assertEqual((1000, 784), tensor_array.shape)
tensor_array[3, 9] = 1.0
tensor_array[19, 11] = 2.0
tensor.set(tensor_array)
tensor.set(tensor_array, place)
tensor_array_2 = numpy.array(tensor)
self.assertAlmostEqual(1.0, tensor_array_2[3, 9])
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册