Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into build_manually
Merge branch developwangkuiyi-patch-2
weixing02
7 years ago
commit
53a4a583e0
@ -1 +1,2 @@
|
||||
add_subdirectory(v2)
|
||||
add_subdirectory(fluid)
|
||||
|
||||
@ -0,0 +1,49 @@
|
||||
if(NOT DEFINED SPHINX_THEME)
|
||||
set(SPHINX_THEME default)
|
||||
endif()
|
||||
|
||||
if(NOT DEFINED SPHINX_THEME_DIR)
|
||||
set(SPHINX_THEME_DIR)
|
||||
endif()
|
||||
|
||||
# configured documentation tools and intermediate build results
|
||||
set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build")
|
||||
|
||||
# Sphinx cache with pickled ReST documents
|
||||
set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
|
||||
|
||||
# HTML output director
|
||||
set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
|
||||
|
||||
configure_file(
|
||||
"${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.en.in"
|
||||
"${BINARY_BUILD_DIR_EN}/conf.py"
|
||||
@ONLY)
|
||||
|
||||
sphinx_add_target(paddle_fluid_docs
|
||||
html
|
||||
${BINARY_BUILD_DIR_EN}
|
||||
${SPHINX_CACHE_DIR_EN}
|
||||
${CMAKE_CURRENT_SOURCE_DIR}
|
||||
${SPHINX_HTML_DIR_EN})
|
||||
|
||||
# configured documentation tools and intermediate build results
|
||||
set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build")
|
||||
|
||||
# Sphinx cache with pickled ReST documents
|
||||
set(SPHINX_CACHE_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_doctrees")
|
||||
|
||||
# HTML output directory
|
||||
set(SPHINX_HTML_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/html")
|
||||
|
||||
configure_file(
|
||||
"${CMAKE_CURRENT_SOURCE_DIR}/../templates/conf.py.cn.in"
|
||||
"${BINARY_BUILD_DIR_CN}/conf.py"
|
||||
@ONLY)
|
||||
|
||||
sphinx_add_target(paddle_fluid_docs_cn
|
||||
html
|
||||
${BINARY_BUILD_DIR_CN}
|
||||
${SPHINX_CACHE_DIR_CN}
|
||||
${CMAKE_CURRENT_SOURCE_DIR}
|
||||
${SPHINX_HTML_DIR_CN})
|
||||
@ -0,0 +1,2 @@
|
||||
安装与使用
|
||||
------------
|
||||
@ -0,0 +1,2 @@
|
||||
Build and Install
|
||||
------------
|
||||
@ -0,0 +1,2 @@
|
||||
设计思想
|
||||
------------
|
||||
@ -0,0 +1,2 @@
|
||||
Design
|
||||
------------
|
||||
@ -0,0 +1,2 @@
|
||||
开发标准
|
||||
------------
|
||||
@ -0,0 +1,4 @@
|
||||
Development
|
||||
------------
|
||||
|
||||
This is Development page
|
||||
@ -0,0 +1,2 @@
|
||||
FAQ
|
||||
------------
|
||||
@ -0,0 +1,2 @@
|
||||
FAQ
|
||||
------------
|
||||
@ -0,0 +1,4 @@
|
||||
新手入门
|
||||
------------
|
||||
|
||||
新手入门
|
||||
@ -0,0 +1,4 @@
|
||||
GET STARTED
|
||||
------------
|
||||
|
||||
This is get started page
|
||||
@ -0,0 +1,2 @@
|
||||
进阶使用
|
||||
------------
|
||||
@ -0,0 +1,4 @@
|
||||
HOW TO
|
||||
------------
|
||||
|
||||
This is how to page
|
||||
@ -0,0 +1,12 @@
|
||||
PaddlePaddle Fluid
|
||||
==========================
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 1
|
||||
|
||||
getstarted/index_cn.rst
|
||||
design/index_cn.rst
|
||||
build_and_install/index_cn.rst
|
||||
howto/index_cn.rst
|
||||
dev/index_cn.rst
|
||||
faq/index_cn.rst
|
||||
@ -0,0 +1,12 @@
|
||||
PaddlePaddle Fluid
|
||||
==========================
|
||||
|
||||
.. toctree::
|
||||
:maxdepth: 1
|
||||
|
||||
getstarted/index_en.rst
|
||||
design/index_en.rst
|
||||
build_and_install/index_en.rst
|
||||
howto/index_en.rst
|
||||
dev/index_en.rst
|
||||
faq/index_en.rst
|
||||
@ -0,0 +1,216 @@
|
||||
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/fluid/operators/average_accumulates_op.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
template <>
|
||||
void GetAccumulators<paddle::platform::CPUDeviceContext>(
|
||||
const framework::ExecutionContext& ctx, int64_t& num_updates_,
|
||||
int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
|
||||
auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
|
||||
auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
|
||||
auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
|
||||
|
||||
old_num_accumulates_ = in_old_num_accumulates->data<int64_t>()[0];
|
||||
num_accumulates_ = in_num_accumulates->data<int64_t>()[0];
|
||||
num_updates_ = in_num_updates->data<int64_t>()[0];
|
||||
}
|
||||
|
||||
template <>
|
||||
void SetAccumulators<paddle::platform::CPUDeviceContext>(
|
||||
const framework::ExecutionContext& ctx, int64_t num_updates_,
|
||||
int64_t num_accumulates_, int64_t old_num_accumulates_) {
|
||||
auto* out_old_num_accumulates = ctx.Output<Tensor>("out_old_num_accumulates");
|
||||
auto* out_num_accumulates = ctx.Output<Tensor>("out_num_accumulates");
|
||||
auto* out_num_updates = ctx.Output<Tensor>("out_num_updates");
|
||||
|
||||
out_old_num_accumulates->data<int64_t>()[0] = old_num_accumulates_;
|
||||
out_num_accumulates->data<int64_t>()[0] = num_accumulates_;
|
||||
out_num_updates->data<int64_t>()[0] = num_updates_;
|
||||
}
|
||||
|
||||
class AverageAccumulatesOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasInput("param"),
|
||||
"Input (param) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasInput("in_sum_1"),
|
||||
"Input (sum_1) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasInput("in_sum_2"),
|
||||
"Input (sum_2) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasInput("in_sum_3"),
|
||||
"Input (sum_3) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasInput("in_num_accumulates"),
|
||||
"Input (in_num_accumulates) of average_accumulates op should "
|
||||
"not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("in_old_num_accumulates"),
|
||||
"Input (old_num_accumulates) of average_accumulates op "
|
||||
"should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasInput("in_num_updates"),
|
||||
"Input (num_updates) of average_accumulates op should not be null.");
|
||||
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasOutput("out_sum_1"),
|
||||
"Output (sum_1) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasOutput("out_sum_2"),
|
||||
"Output (sum_2) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasOutput("out_sum_3"),
|
||||
"Output (sum_3) of average_accumulates op should not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("out_num_accumulates"),
|
||||
"Output (num_accumulates) of average_accumulates op should "
|
||||
"not be null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("out_old_num_accumulates"),
|
||||
"Output (old_num_accumulates) of average_accumulates op "
|
||||
"should not be null.");
|
||||
PADDLE_ENFORCE(
|
||||
ctx->HasOutput("out_num_updates"),
|
||||
"Output (num_updates) of average_accumulates op should not be null.");
|
||||
|
||||
auto in_dim = ctx->GetInputDim("param");
|
||||
|
||||
ctx->SetOutputDim("out_sum_1", in_dim);
|
||||
ctx->SetOutputDim("out_sum_2", in_dim);
|
||||
ctx->SetOutputDim("out_sum_3", in_dim);
|
||||
ctx->SetOutputDim("out_num_accumulates", {1});
|
||||
ctx->SetOutputDim("out_old_num_accumulates", {1});
|
||||
ctx->SetOutputDim("out_num_updates", {1});
|
||||
}
|
||||
|
||||
protected:
|
||||
framework::OpKernelType GetExpectedKernelType(
|
||||
const framework::ExecutionContext& ctx) const override {
|
||||
return framework::OpKernelType(
|
||||
framework::ToDataType(ctx.Input<Tensor>("param")->type()),
|
||||
ctx.GetPlace());
|
||||
}
|
||||
};
|
||||
|
||||
class AverageAccumulatesOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
AverageAccumulatesOpMaker(OpProto* proto, OpAttrChecker* op_checker)
|
||||
: OpProtoAndCheckerMaker(proto, op_checker) {
|
||||
AddInput("param", "(Tensor), The parameter to be accumulated.");
|
||||
AddInput("in_sum_1",
|
||||
"(Tensor), A tensor used to store the parameter "
|
||||
"sums with the same shape as input(param).");
|
||||
AddInput("in_sum_2",
|
||||
"(Tensor), A auxiliary tensor to help "
|
||||
"accumulating sums of parameter values with the same shape as "
|
||||
"input(param). It is used to avoid loss of precision due to too "
|
||||
"many sums.");
|
||||
AddInput("in_sum_3",
|
||||
"(Tensor), A auxiliary tensor to help "
|
||||
"accumulating sums of parameter values with the same shape as "
|
||||
"input(param).");
|
||||
AddInput("in_num_accumulates",
|
||||
"(Tensor<int64_t>), The accumulating times of current window with "
|
||||
"shape [1].");
|
||||
AddInput(
|
||||
"in_old_num_accumulates",
|
||||
"(Tensor<int64_t>), The accumulating times of previous window with "
|
||||
"shape [1].");
|
||||
AddInput("in_num_updates",
|
||||
"(Tensor<int64_t>), The total number of batches used by trainning "
|
||||
"before this batch with shape [1].");
|
||||
|
||||
AddOutput("out_sum_1",
|
||||
"(Tensor), A tensor used to store the "
|
||||
"parameter sums with the same shape as input(param).");
|
||||
AddOutput("out_sum_2",
|
||||
"(Tensor), A auxiliary tensor to help "
|
||||
"accumulating sums of parameter values with the same shape as "
|
||||
"input(param). It is used to avoid loss of precision due to too "
|
||||
"many sums.");
|
||||
AddOutput("out_sum_3",
|
||||
"(Tensor), A auxiliary tensor to help "
|
||||
"accumulating sums of parameter values with the same shape as "
|
||||
"input(param).");
|
||||
AddOutput(
|
||||
"out_num_accumulates",
|
||||
"(Tensor<int64_t>), The accumulating times of current window with "
|
||||
"shape [1].");
|
||||
AddOutput(
|
||||
"out_old_num_accumulates",
|
||||
"(Tensor<int64_t>) The accumulating times of previous window with "
|
||||
"shape [1].");
|
||||
AddOutput(
|
||||
"out_num_updates",
|
||||
"(Tensor<int64_t>), The total number of batches used by trainning "
|
||||
"before this batch with shape [1].");
|
||||
|
||||
AddAttr<float>("average_window",
|
||||
"(float, default 0) "
|
||||
"The rate of average window size relative to num_updates.")
|
||||
.SetDefault(0);
|
||||
AddAttr<int64_t>("max_average_window",
|
||||
"(int64_t) "
|
||||
"Maximum size of average window. It suggests that the "
|
||||
"number of mini-batches "
|
||||
"in one pass is appropriate value to set.");
|
||||
AddAttr<int64_t>("min_average_window",
|
||||
"(int64_t, default 10000L) "
|
||||
"Minimu size of average window.")
|
||||
.SetDefault(10000L);
|
||||
|
||||
AddComment(R"DOC(
|
||||
AverageAccumulates Operator.
|
||||
Accumulate the sum of parameter whtin sliding window. The size of sliding window is
|
||||
determined by 'average_window', 'max_average_window' and 'min_average_window'.
|
||||
Memory was shared by Input(in_sum_1) and Output(out_sum_1) which acts as an accumulator 'sum_1'.
|
||||
'sum_2', 'sum_3', 'num_accumulates', 'old_num_accumulates' and 'num_updates' were the same as 'sum_1'.
|
||||
|
||||
All the accumulators were inited to zero before training.
|
||||
|
||||
And for a mini-batch in training, accumulators were computed as below steps:
|
||||
num_updates += 1
|
||||
num_accumulates += 1
|
||||
sum_1 += param
|
||||
if num_updates % kMaxNumAccumulates == 0:
|
||||
sum_2 += sum_1
|
||||
sum_1 = 0
|
||||
if num_accumulates >= min_average_window && num_accumulates >= min(max_average_window, num_updates * average_window):
|
||||
sum_3 = sum_1 + sum_2
|
||||
sum_1 = 0
|
||||
sum_2 = 0
|
||||
old_num_accumulates = num_accumulates
|
||||
num_accumulates = 0
|
||||
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OPERATOR(average_accumulates, ops::AverageAccumulatesOp,
|
||||
ops::AverageAccumulatesOpMaker,
|
||||
paddle::framework::EmptyGradOpMaker);
|
||||
REGISTER_OP_CPU_KERNEL(
|
||||
average_accumulates,
|
||||
ops::AverageAccumulatesKernel<paddle::platform::CPUDeviceContext, float>,
|
||||
ops::AverageAccumulatesKernel<paddle::platform::CPUDeviceContext, double>);
|
||||
@ -0,0 +1,63 @@
|
||||
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#include "paddle/fluid/operators/average_accumulates_op.h"
|
||||
#include "paddle/fluid/platform/gpu_info.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
template <>
|
||||
void GetAccumulators<paddle::platform::CUDADeviceContext>(
|
||||
const framework::ExecutionContext& ctx, int64_t& num_updates_,
|
||||
int64_t& num_accumulates_, int64_t& old_num_accumulates_) {
|
||||
auto* in_old_num_accumulates = ctx.Input<Tensor>("in_old_num_accumulates");
|
||||
auto* in_num_accumulates = ctx.Input<Tensor>("in_num_accumulates");
|
||||
auto* in_num_updates = ctx.Input<Tensor>("in_num_updates");
|
||||
auto stream = ctx.cuda_device_context().stream();
|
||||
memory::Copy(platform::CPUPlace(), &old_num_accumulates_,
|
||||
platform::CUDAPlace(), in_old_num_accumulates->data<int64_t>(),
|
||||
sizeof(int64_t), stream);
|
||||
memory::Copy(platform::CPUPlace(), &num_accumulates_, platform::CUDAPlace(),
|
||||
in_num_accumulates->data<int64_t>(), sizeof(int64_t), stream);
|
||||
memory::Copy(platform::CPUPlace(), &num_updates_, platform::CUDAPlace(),
|
||||
in_num_updates->data<int64_t>(), sizeof(int64_t), stream);
|
||||
}
|
||||
|
||||
template <>
|
||||
void SetAccumulators<paddle::platform::CUDADeviceContext>(
|
||||
const framework::ExecutionContext& ctx, int64_t num_updates_,
|
||||
int64_t num_accumulates_, int64_t old_num_accumulates_) {
|
||||
auto stream = ctx.cuda_device_context().stream();
|
||||
auto* out_old_num_accumulates = ctx.Output<Tensor>("out_old_num_accumulates");
|
||||
auto* out_num_accumulates = ctx.Output<Tensor>("out_num_accumulates");
|
||||
auto* out_num_updates = ctx.Output<Tensor>("out_num_updates");
|
||||
|
||||
memory::Copy(platform::CUDAPlace(), out_old_num_accumulates->data<int64_t>(),
|
||||
platform::CPUPlace(), &old_num_accumulates_, sizeof(int64_t),
|
||||
stream);
|
||||
memory::Copy(platform::CUDAPlace(), out_num_accumulates->data<int64_t>(),
|
||||
platform::CPUPlace(), &num_accumulates_, sizeof(int64_t),
|
||||
stream);
|
||||
memory::Copy(platform::CUDAPlace(), out_num_updates->data<int64_t>(),
|
||||
platform::CPUPlace(), &num_updates_, sizeof(int64_t), stream);
|
||||
}
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OP_CUDA_KERNEL(
|
||||
average_accumulates,
|
||||
ops::AverageAccumulatesKernel<paddle::platform::CUDADeviceContext, float>,
|
||||
ops::AverageAccumulatesKernel<paddle::platform::CUDADeviceContext, double>);
|
||||
@ -0,0 +1,113 @@
|
||||
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License. */
|
||||
|
||||
#pragma once
|
||||
#include <algorithm>
|
||||
#include "paddle/fluid/framework/eigen.h"
|
||||
#include "paddle/fluid/framework/op_registry.h"
|
||||
#include "paddle/fluid/operators/math/math_function.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
|
||||
template <typename T, int MajorType = Eigen::RowMajor,
|
||||
typename IndexType = Eigen::DenseIndex>
|
||||
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
|
||||
|
||||
template <typename DeviceContext>
|
||||
void GetAccumulators(const framework::ExecutionContext& ctx,
|
||||
int64_t& num_updates, int64_t& num_accumulates,
|
||||
int64_t& old_num_accumulates);
|
||||
|
||||
template <typename DeviceContext>
|
||||
void SetAccumulators(const framework::ExecutionContext& ctx,
|
||||
int64_t num_updates, int64_t num_accumulates,
|
||||
int64_t old_num_accumulates);
|
||||
|
||||
template <typename DeviceContext, typename T>
|
||||
class AverageAccumulatesKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& ctx) const override {
|
||||
// It is used to avoid loss of precision
|
||||
static const int64_t kMaxNumAccumulates = 16384;
|
||||
// Get accumulators from input
|
||||
int64_t num_updates = 0;
|
||||
int64_t num_accumulates = 0;
|
||||
int64_t old_num_accumulates = 0;
|
||||
GetAccumulators<DeviceContext>(ctx, num_updates, num_accumulates,
|
||||
old_num_accumulates);
|
||||
|
||||
// Get attrs
|
||||
float average_window = ctx.Attr<float>("average_window");
|
||||
int64_t max_average_window = ctx.Attr<int64_t>("max_average_window");
|
||||
int64_t min_average_window = ctx.Attr<int64_t>("min_average_window");
|
||||
min_average_window =
|
||||
std::min<int64_t>(min_average_window, max_average_window);
|
||||
|
||||
// Get inputs
|
||||
auto* param = ctx.Input<Tensor>("param");
|
||||
auto* in_sum_1 = ctx.Input<Tensor>("in_sum_1");
|
||||
auto* in_sum_2 = ctx.Input<Tensor>("in_sum_2");
|
||||
auto* in_sum_3 = ctx.Input<Tensor>("in_sum_3");
|
||||
auto param_tensor = EigenVector<T>::Flatten(*param);
|
||||
auto in_sum_1_tensor = EigenVector<T>::Flatten(*in_sum_1);
|
||||
auto in_sum_2_tensor = EigenVector<T>::Flatten(*in_sum_2);
|
||||
auto in_sum_3_tensor = EigenVector<T>::Flatten(*in_sum_3);
|
||||
|
||||
// Get outputs
|
||||
auto* out_sum_1 = ctx.Output<Tensor>("out_sum_1");
|
||||
auto* out_sum_2 = ctx.Output<Tensor>("out_sum_2");
|
||||
auto* out_sum_3 = ctx.Output<Tensor>("out_sum_3");
|
||||
auto out_sum_1_tensor = EigenVector<T>::Flatten(*out_sum_1);
|
||||
auto out_sum_2_tensor = EigenVector<T>::Flatten(*out_sum_2);
|
||||
auto out_sum_3_tensor = EigenVector<T>::Flatten(*out_sum_3);
|
||||
|
||||
// Compute
|
||||
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
|
||||
math::SetConstant<DeviceContext, T> constant_functor;
|
||||
++num_updates;
|
||||
++num_accumulates;
|
||||
out_sum_1_tensor.device(place) = in_sum_1_tensor + param_tensor;
|
||||
out_sum_2_tensor.device(place) = in_sum_2_tensor;
|
||||
out_sum_3_tensor.device(place) = in_sum_3_tensor;
|
||||
if (num_updates % kMaxNumAccumulates == 0) {
|
||||
// Move the sum to a different buffer to avoid loss of precision due to
|
||||
// too many sums.
|
||||
out_sum_2_tensor.device(place) = in_sum_2_tensor + in_sum_1_tensor;
|
||||
constant_functor(ctx.template device_context<DeviceContext>(), out_sum_1,
|
||||
0.0);
|
||||
}
|
||||
if (num_accumulates >= min_average_window &&
|
||||
num_accumulates >= std::min<int64_t>(max_average_window,
|
||||
num_updates * average_window)) {
|
||||
// Now the average window is too long, discard the old sum.
|
||||
out_sum_3_tensor.device(place) = in_sum_1_tensor + in_sum_2_tensor;
|
||||
constant_functor(ctx.template device_context<DeviceContext>(), out_sum_1,
|
||||
0.0);
|
||||
constant_functor(ctx.template device_context<DeviceContext>(), out_sum_2,
|
||||
0.0);
|
||||
old_num_accumulates = num_accumulates;
|
||||
num_accumulates = 0;
|
||||
}
|
||||
|
||||
// Set accumulators to output
|
||||
SetAccumulators<DeviceContext>(ctx, num_updates, num_accumulates,
|
||||
old_num_accumulates);
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue