提交 c0cef849 编写于 作者: C caoying03

softmax as function.

上级 513bc997
......@@ -55,10 +55,12 @@ set(DEPS_OPS
minus_op
mul_op
recurrent_op
scale_op)
scale_op
softmax_op)
op_library(identity_op DEPS scale_op)
op_library(minus_op DEPS scale_op)
op_library(mul_op DEPS math_function)
op_library(softmax_op DEPS math_function)
op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS framework_proto tensor operator net_op)
op_library(scale_op DEPS net_op)
......
if(WITH_GPU)
nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc
im2col.cu DEPS cblas device_context)
nv_library(math_function SRCS math_function.cc math_function.cu im2col.cc
im2col.cu softmax_function.cc DEPS cblas device_context operator)
else()
cc_library(math_function SRCS math_function.cc im2col.cc DEPS cblas device_context)
cc_library(math_function SRCS math_function.cc im2col.cc
softmax_function.cc DEPS cblas device_context operator)
endif()
nv_test(math_function_test SRCS math_function_test.cc DEPS math_function tensor)
......
/* 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. */
#ifndef PADDLE_ONLY_CPU
#define EIGEN_USE_GPU
#endif
#include "paddle/operators/math/softmax_function.h"
namespace paddle {
namespace operators {
namespace math {
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename Place, typename T>
void softmax(const framework::Tensor* X, framework::Tensor* Y,
const framework::ExecutionContext& context) {
auto logits = EigenMatrix<T>::From(*X);
auto softmax = EigenMatrix<T>::From(*Y);
const int kBatchDim = 0;
const int kClassDim = 1;
const int batch_size = logits.dimension(kBatchDim);
const int num_classes = logits.dimension(kClassDim);
Eigen::DSizes<int, 1> along_class(kClassDim);
Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
Eigen::DSizes<int, 2> one_by_class(1, num_classes);
auto shifted_logits = (logits -
logits.maximum(along_class)
.eval()
.reshape(batch_by_one)
.broadcast(one_by_class));
softmax.device(context.GetEigenDevice<Place>()) = shifted_logits.exp();
softmax.device(context.GetEigenDevice<Place>()) =
(softmax *
softmax.sum(along_class)
.inverse()
.eval()
.reshape(batch_by_one)
.broadcast(one_by_class));
}
} // namespace math
} // namespace operators
} // 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. */
#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/tensor.h"
namespace paddle {
namespace operators {
namespace math {
template <typename Place, typename T>
void softmax(const framework::Tensor* X, framework::Tensor* Y,
const framework::ExecutionContext& context);
} // namespace math
} // namespace operators
} // namespace paddle
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/softmax_function.h"
namespace paddle {
namespace operators {
......@@ -30,36 +31,11 @@ class SoftmaxKernel : public framework::OpKernel {
void Compute(const framework::ExecutionContext& context) const override {
auto X = context.Input<Tensor>("X");
auto Y = context.Output<Tensor>("Y");
Y->mutable_data<T>(context.GetPlace());
auto logits = EigenMatrix<T>::From(*X);
auto softmax = EigenMatrix<T>::From(*Y);
const int kBatchDim = 0;
const int kClassDim = 1;
const int batch_size = logits.dimension(kBatchDim);
const int num_classes = logits.dimension(kClassDim);
// allocate memory on device.
Y->mutable_data<T>(context.GetPlace());
Eigen::DSizes<int, 1> along_class(kClassDim);
Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
Eigen::DSizes<int, 2> one_by_class(1, num_classes);
auto shifted_logits = (logits -
logits.maximum(along_class)
.eval()
.reshape(batch_by_one)
.broadcast(one_by_class));
softmax.device(context.GetEigenDevice<Place>()) = shifted_logits.exp();
softmax.device(context.GetEigenDevice<Place>()) =
(softmax *
softmax.sum(along_class)
.inverse()
.eval()
.reshape(batch_by_one)
.broadcast(one_by_class));
math::softmax<Place, T>(X, Y, context);
}
};
......
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