Skip to content

P
Paddle

Crayon鑫 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 cf1944af 编写于 作者: S sneaxiy
浏览文件
操作

test=develop

上级 2494ca83 2eaa291e
隐藏空白更改
内联 并排
Showing with 1452 addition and 608 deletion
CMakeLists.txt
浏览文件 @ cf1944af
......@@ -62,13 +62,11 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded Eigen" OFF)
option(WITH_ARM_FP16 "Use half precision support on armv8.2-a cpu" OFF)
option(WITH_FAST_BUNDLE_TEST "Bundle tests that can be run in a single process together to reduce launch overhead" OFF)
option(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." OFF)
option(WITH_ANAKIN "Compile with Anakin library" OFF)
option(WITH_GRPC "Use grpc as the default rpc framework" ${WITH_DISTRIBUTE})
option(WITH_BRPC_RDMA "Use brpc rdma as the rpc protocal" OFF)
option(WITH_INFERENCE "Compile fluid inference library" ON)
option(ON_INFER "Turn on inference optimization." OFF)
option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF)
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
......
README.md
浏览文件 @ cf1944af
......@@ -2,8 +2,8 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.0/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......@@ -19,7 +19,7 @@ Our vision is to enable deep learning for everyone via PaddlePaddle.
Please refer to our [release announcement](https://github.com/PaddlePaddle/Paddle/releases) to track the latest feature of PaddlePaddle.
### Latest PaddlePaddle Release: [Fluid 1.0.1](https://github.com/PaddlePaddle/Paddle/tree/release/1.0.0)
### Latest PaddlePaddle Release: [Fluid 1.1.0](https://github.com/PaddlePaddle/Paddle/tree/release/1.1)
### Install Latest Stable Release:
```
# Linux CPU
......@@ -27,9 +27,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.0.1.post87
pip install paddlepaddle-gpu==1.1.0.post87
# Linux GPU cuda8cudnn5
pip install paddlepaddle-gpu==1.0.1.post85
pip install paddlepaddle-gpu==1.1.0.post85
# For installation on other platform, refer to http://paddlepaddle.org/
```
......@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.0.1.post85
## Installation
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) on our website.
It is recommended to read [this doc](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) on our website.
## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation.
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book)
You might want to start from this online interactive book that can run in a Jupyter Notebook.
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.0/user_guides/howto/training/cluster_howto.html)
- [Distributed Training](http://paddlepaddle.org/documentation/docs/zh/1.1/user_guides/howto/training/cluster_howto.html)
You can run distributed training jobs on MPI clusters.
- [Python API](http://paddlepaddle.org/documentation/api/zh/1.0/fluid.html)
- [Python API](http://paddlepaddle.org/documentation/api/zh/1.1/fluid.html)
Our new API enables much shorter programs.
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.0/advanced_usage/development/contribute_to_paddle.html)
- [How to Contribute](http://paddlepaddle.org/documentation/docs/zh/1.1/advanced_usage/development/contribute_to_paddle.html)
We appreciate your contributions!
......
paddle/CMakeLists.txt
浏览文件 @ cf1944af
......@@ -24,6 +24,7 @@ if(NOT WITH_FLUID_ONLY)
endif()
add_subdirectory(testing)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory")
if(NOT MOBILE_INFERENCE AND NOT RPI AND NOT WITH_C_API)
add_subdirectory(fluid)
endif()
paddle/fluid/API.spec
浏览文件 @ cf1944af
......@@ -64,7 +64,7 @@ paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', '
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type', 'is_test'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
......@@ -177,6 +177,8 @@ paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, k
paddle.fluid.layers.sequence_reverse ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.affine_channel ArgSpec(args=['x', 'scale', 'bias', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(None, None, 'NCHW', None))
paddle.fluid.layers.hash ArgSpec(args=['input', 'hash_size', 'num_hash', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.log_loss ArgSpec(args=['input', 'label', 'epsilon', 'name'], varargs=None, keywords=None, defaults=(0.0001, None))
paddle.fluid.layers.add_position_encoding ArgSpec(args=['input', 'alpha', 'beta', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
paddle.fluid.layers.read_file ArgSpec(args=['reader'], varargs=None, keywords=None, defaults=None)
......
paddle/fluid/CMakeLists.txt
浏览文件 @ cf1944af
......@@ -9,8 +9,6 @@ add_subdirectory(pybind)
add_subdirectory(recordio)
endif(NOT WIN32)
if(WITH_INFERENCE)
# NOTE: please add subdirectory inference at last.
add_subdirectory(inference)
add_subdirectory(train)
endif()
# NOTE: please add subdirectory inference at last.
add_subdirectory(inference)
add_subdirectory(train)
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ cf1944af
......@@ -58,6 +58,7 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu
# device_context reduce_op_handle )
cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc
DEPS fetch_op_handle ssa_graph_executor scope simple_threadpool device_context)
cc_test(fused_broadcast_op_test SRCS fused_broadcast_op_handle_test.cc DEPS fused_broadcast_op_handle)
cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass
......
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ cf1944af
......@@ -34,7 +34,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
nccl_ctxs_(ctxs) {
if (nccl_ctxs_) {
for (auto &p : places_) {
this->dev_ctxes_[p] = nccl_ctxs_->DevCtx(p);
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
}
}
}
......@@ -46,7 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
#endif
void AllReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
if (NoDummyInputSize() == 1) {
return; // No need to all reduce when GPU count = 1;
......@@ -127,7 +127,7 @@ void AllReduceOpHandle::RunImpl() {
*local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto &p = places_[i];
auto *var = scope.FindVar(out_var_handles[i]->name_);
auto *dev_ctx = dev_ctxes_[p];
auto *dev_ctx = dev_ctxes_.at(p);
RunAndRecordEvent(p, [&trg, var, dev_ctx, p] {
auto &tensor_gpu = *var->GetMutable<framework::LoDTensor>();
......
paddle/fluid/framework/details/broadcast_op_handle.h
浏览文件 @ cf1944af
......@@ -44,7 +44,8 @@ struct BroadcastOpHandle : public OpHandleBase {
nccl_ctxs_(nccl_ctxs) {
if (nccl_ctxs_) {
for (auto &p_ctx : nccl_ctxs_->contexts_) {
dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
this->SetDeviceContext(platform::CUDAPlace(p_ctx.first),
p_ctx.second.ctx_.get());
}
}
}
......
paddle/fluid/framework/details/broadcast_op_handle_test.cc
浏览文件 @ cf1944af
......@@ -12,232 +12,12 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "gtest/gtest.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
namespace paddle {
namespace framework {
namespace details {
namespace f = paddle::framework;
namespace p = paddle::platform;
// test data amount
const f::DDim kDims = {20, 20};
struct TestBroadcastOpHandle {
std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
std::vector<Scope*> local_scopes_;
std::vector<Scope*> param_scopes_;
Scope g_scope_;
std::unique_ptr<OpHandleBase> op_handle_;
std::vector<std::unique_ptr<VarHandleBase>> vars_;
std::vector<p::Place> gpu_list_;
bool use_gpu_;
#ifdef PADDLE_WITH_CUDA
std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
#endif
void WaitAll() {
for (size_t j = 0; j < ctxs_.size(); ++j) {
ctxs_[j]->Wait();
}
#ifdef PADDLE_WITH_CUDA
if (nccl_ctxs_) {
nccl_ctxs_->WaitAll();
}
#endif
}
void InitCtxOnGpu(bool use_gpu) {
use_gpu_ = use_gpu;
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
int count = p::GetCUDADeviceCount();
if (count <= 1) {
LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
"device count is "
<< count;
exit(0);
}
for (int i = 0; i < count; ++i) {
auto p = p::CUDAPlace(i);
gpu_list_.push_back(p);
ctxs_.emplace_back(new p::CUDADeviceContext(p));
}
nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
int count = 8;
for (int i = 0; i < count; ++i) {
auto p = p::CPUPlace();
gpu_list_.push_back(p);
ctxs_.emplace_back(new p::CPUDeviceContext(p));
}
#ifdef PADDLE_WITH_CUDA
nccl_ctxs_.reset(nullptr);
#endif
}
}
void InitBroadcastOp(size_t input_scope_idx) {
for (size_t j = 0; j < gpu_list_.size(); ++j) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
*local_scopes_.back()
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
local_scope.Var("out");
param_scopes_.emplace_back(&local_scope);
}
param_scopes_[input_scope_idx]->Var("input");
std::unique_ptr<ir::Node> n =
ir::CreateNodeForTest("node0", ir::Node::Type::kOperation);
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_,
nccl_ctxs_.get()));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_,
nccl_ctxs_.get()));
#else
op_handle_.reset(
new BroadcastOpHandle(n.get(), local_scopes_, gpu_list_));
#endif
}
std::unique_ptr<ir::Node> v =
ir::CreateNodeForTest("node1", ir::Node::Type::kVariable);
auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input",
gpu_list_[input_scope_idx]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
// add dummy var
std::unique_ptr<ir::Node> v2 =
ir::CreateNodeForTest("node2", ir::Node::Type::kVariable);
vars_.emplace_back(new DummyVarHandle(v2.get()));
DummyVarHandle* dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
dummy_var_handle->ClearGeneratedOp();
op_handle_->AddInput(dummy_var_handle);
for (size_t j = 0; j < gpu_list_.size(); ++j) {
if (!use_gpu_) {
op_handle_->SetDeviceContext(gpu_list_[j], ctxs_[j].get());
}
std::unique_ptr<ir::Node> v3 =
ir::CreateNodeForTest("node3", ir::Node::Type::kVariable);
VarHandle* out_var_handle =
new VarHandle(v3.get(), 2, j, "out", gpu_list_[j]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
// add dummy var
std::unique_ptr<ir::Node> v4 =
ir::CreateNodeForTest("node4", ir::Node::Type::kVariable);
vars_.emplace_back(new DummyVarHandle(v4.get()));
DummyVarHandle* out_dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
out_dummy_var_handle->ClearGeneratedOp();
op_handle_->AddOutput(out_dummy_var_handle);
}
void TestBroadcastLodTensor(size_t input_scope_idx) {
auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
PADDLE_ENFORCE_NOT_NULL(in_var);
auto in_lod_tensor = in_var->GetMutable<f::LoDTensor>();
in_lod_tensor->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k;
}
f::LoD lod{{0, 10, 20}};
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), in_lod_tensor);
in_lod_tensor->set_lod(lod);
in_lod_tensor->Resize(kDims);
op_handle_->Run(false);
WaitAll();
p::CPUPlace cpu_place;
for (size_t j = 0; j < gpu_list_.size(); ++j) {
auto out_var = param_scopes_[j]->FindVar("out");
PADDLE_ENFORCE_NOT_NULL(out_var);
auto out_tensor = out_var->Get<f::LoDTensor>();
PADDLE_ENFORCE_EQ(out_tensor.lod(), lod, "lod is not equal.");
f::Tensor result_tensor;
f::TensorCopySync(out_tensor, cpu_place, &result_tensor);
float* ct = result_tensor.mutable_data<float>(cpu_place);
for (int64_t i = 0; i < f::product(kDims); ++i) {
ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
}
}
}
void TestBroadcastSelectedRows(size_t input_scope_idx) {
auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
PADDLE_ENFORCE_NOT_NULL(in_var);
auto in_selected_rows = in_var->GetMutable<f::SelectedRows>();
auto value = in_selected_rows->mutable_value();
value->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
int height = static_cast<int>(kDims[0]) * 2;
std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
in_selected_rows->set_height(height);
in_selected_rows->set_rows(rows);
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k;
}
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), value);
op_handle_->Run(false);
WaitAll();
p::CPUPlace cpu_place;
for (size_t j = 0; j < gpu_list_.size(); ++j) {
auto out_var = param_scopes_[j]->FindVar("out");
PADDLE_ENFORCE_NOT_NULL(out_var);
auto& out_select_rows = out_var->Get<f::SelectedRows>();
auto rt = out_select_rows.value();
PADDLE_ENFORCE_EQ(out_select_rows.height(), height,
"height is not equal.");
for (size_t k = 0; k < out_select_rows.rows().size(); ++k) {
PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k]);
}
f::Tensor result_tensor;
f::TensorCopySync(rt, cpu_place, &result_tensor);
float* ct = result_tensor.data<float>();
for (int64_t i = 0; i < f::product(kDims); ++i) {
ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
}
}
}
};
TEST(BroadcastTester, TestCPUBroadcastTestLodTensor) {
TestBroadcastOpHandle test_op;
size_t input_scope_idx = 0;
......
paddle/fluid/framework/details/broadcast_op_handle_test.h 0 → 100644
浏览文件 @ cf1944af
// 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 <string>
#include <vector>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/details/broadcast_op_handle.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle {
namespace framework {
namespace details {
namespace f = paddle::framework;
namespace p = paddle::platform;
// test data amount
const f::DDim kDims = {20, 20};
struct TestBroadcastOpHandle {
std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
std::vector<Scope*> local_scopes_;
std::vector<Scope*> param_scopes_;
Scope g_scope_;
std::unique_ptr<OpHandleBase> op_handle_;
std::vector<std::unique_ptr<VarHandleBase>> vars_;
std::vector<p::Place> place_list_;
bool use_gpu_;
#ifdef PADDLE_WITH_CUDA
std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
#endif
void WaitAll() {
for (size_t j = 0; j < ctxs_.size(); ++j) {
ctxs_[j]->Wait();
}
#ifdef PADDLE_WITH_CUDA
if (nccl_ctxs_) {
nccl_ctxs_->WaitAll();
}
#endif
}
void InitCtxOnGpu(bool use_gpu) {
use_gpu_ = use_gpu;
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
int count = p::GetCUDADeviceCount();
if (count <= 1) {
LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
"device count is "
<< count;
exit(0);
}
for (int i = 0; i < count; ++i) {
auto p = p::CUDAPlace(i);
place_list_.push_back(p);
ctxs_.emplace_back(new p::CUDADeviceContext(p));
}
nccl_ctxs_.reset(new platform::NCCLContextMap(place_list_));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
int count = 8;
for (int i = 0; i < count; ++i) {
auto p = p::CPUPlace();
place_list_.push_back(p);
ctxs_.emplace_back(new p::CPUDeviceContext(p));
}
#ifdef PADDLE_WITH_CUDA
nccl_ctxs_.reset(nullptr);
#endif
}
}
void InitBroadcastOp(size_t input_scope_idx) {
for (size_t j = 0; j < place_list_.size(); ++j) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
*local_scopes_.back()
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
local_scope.Var("out");
param_scopes_.emplace_back(&local_scope);
}
param_scopes_[input_scope_idx]->Var("input");
std::unique_ptr<ir::Node> n =
ir::CreateNodeForTest("node0", ir::Node::Type::kOperation);
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_,
place_list_, nccl_ctxs_.get()));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new BroadcastOpHandle(n.get(), local_scopes_,
place_list_, nccl_ctxs_.get()));
#else
op_handle_.reset(
new BroadcastOpHandle(n.get(), local_scopes_, place_list_));
#endif
}
std::unique_ptr<ir::Node> v =
ir::CreateNodeForTest("node1", ir::Node::Type::kVariable);
auto* in_var_handle = new VarHandle(v.get(), 1, input_scope_idx, "input",
place_list_[input_scope_idx]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
// add dummy var
std::unique_ptr<ir::Node> v2 =
ir::CreateNodeForTest("node2", ir::Node::Type::kVariable);
vars_.emplace_back(new DummyVarHandle(v2.get()));
DummyVarHandle* dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
dummy_var_handle->ClearGeneratedOp();
op_handle_->AddInput(dummy_var_handle);
for (size_t j = 0; j < place_list_.size(); ++j) {
if (!use_gpu_) {
op_handle_->SetDeviceContext(place_list_[j], ctxs_[j].get());
}
std::unique_ptr<ir::Node> v3 =
ir::CreateNodeForTest("node3", ir::Node::Type::kVariable);
VarHandle* out_var_handle =
new VarHandle(v3.get(), 2, j, "out", place_list_[j]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
// add dummy var
std::unique_ptr<ir::Node> v4 =
ir::CreateNodeForTest("node4", ir::Node::Type::kVariable);
vars_.emplace_back(new DummyVarHandle(v4.get()));
DummyVarHandle* out_dummy_var_handle =
static_cast<DummyVarHandle*>(vars_.back().get());
out_dummy_var_handle->ClearGeneratedOp();
op_handle_->AddOutput(out_dummy_var_handle);
}
std::vector<float> InitLoDTensor(const std::string& varname,
size_t input_scope_idx, const f::LoD& lod,
float val_scalar = 0.0) {
auto var = param_scopes_[input_scope_idx]->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto lod_tensor = var->GetMutable<f::LoDTensor>();
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k + val_scalar;
}
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), lod_tensor);
lod_tensor->set_lod(lod);
lod_tensor->Resize(kDims);
return send_vector;
}
std::vector<float> InitSelectedRows(const std::string& varname,
size_t input_scope_idx,
const std::vector<int64_t>& rows,
int height, float value_scalar = 0.0) {
std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
for (size_t k = 0; k < send_vector.size(); ++k) {
send_vector[k] = k + value_scalar;
}
auto var = param_scopes_[input_scope_idx]->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto selected_rows = var->GetMutable<f::SelectedRows>();
auto value = selected_rows->mutable_value();
value->mutable_data<float>(kDims, place_list_[input_scope_idx]);
selected_rows->set_height(height);
selected_rows->set_rows(rows);
paddle::framework::TensorFromVector<float>(
send_vector, *(ctxs_[input_scope_idx]), value);
return send_vector;
}
void SelectedRowsEqual(const std::string& varname, int input_scope_idx,
const std::vector<float>& send_vector,
const std::vector<int64_t>& rows, int height) {
auto var = param_scopes_[input_scope_idx]->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto& selected_rows = var->Get<f::SelectedRows>();
auto rt = selected_rows.value();
PADDLE_ENFORCE_EQ(selected_rows.height(), height, "height is not equal.");
for (size_t k = 0; k < selected_rows.rows().size(); ++k) {
PADDLE_ENFORCE_EQ(selected_rows.rows()[k], rows[k]);
}
p::CPUPlace cpu_place;
f::Tensor result_tensor;
f::TensorCopySync(rt, cpu_place, &result_tensor);
float* ct = result_tensor.data<float>();
for (int64_t i = 0; i < f::product(kDims); ++i) {
ASSERT_NEAR(ct[i], send_vector[i], 1e-5);
}
}
void LoDTensorEqual(const std::string& varname,
const std::vector<float>& send_vec, const f::LoD& lod,
framework::Scope* scope) {
p::CPUPlace cpu_place;
auto var = scope->FindVar(varname);
PADDLE_ENFORCE_NOT_NULL(var);
auto tensor = var->Get<f::LoDTensor>();
PADDLE_ENFORCE_EQ(tensor.lod(), lod, "lod is not equal.");
f::Tensor result_tensor;
f::TensorCopySync(tensor, cpu_place, &result_tensor);
float* ct = result_tensor.mutable_data<float>(cpu_place);
for (int64_t k = 0; k < f::product(kDims); ++k) {
ASSERT_NEAR(ct[k], send_vec[k], 1e-5);
}
}
void TestBroadcastLodTensor(size_t input_scope_idx) {
f::LoD lod{{0, 10, 20}};
auto send_vector = InitLoDTensor("input", input_scope_idx, lod);
op_handle_->Run(false);
WaitAll();
for (size_t j = 0; j < place_list_.size(); ++j) {
LoDTensorEqual("out", send_vector, lod, param_scopes_[j]);
}
}
void TestBroadcastSelectedRows(size_t input_scope_idx) {
std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
int height = static_cast<int>(kDims[0] * 2);
auto send_vector = InitSelectedRows("input", input_scope_idx, rows, height);
op_handle_->Run(false);
WaitAll();
for (size_t j = 0; j < place_list_.size(); ++j) {
SelectedRowsEqual("out", input_scope_idx, send_vector, rows, height);
}
}
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/computation_op_handle.cc
浏览文件 @ cf1944af
......@@ -37,7 +37,7 @@ void ComputationOpHandle::RunImpl() {
bool ComputationOpHandle::NeedWait(VarHandleBase *in_var) {
bool need_wait =
in_var && in_var->GeneratedOp() &&
in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_[place_];
in_var->GeneratedOp()->DeviceContext(place_) != dev_ctxes_.at(place_);
return need_wait;
}
......
paddle/fluid/framework/details/data_balance_op_handle.cc
浏览文件 @ cf1944af
......@@ -28,7 +28,7 @@ DataBalanceOpHandle::DataBalanceOpHandle(
: OpHandleBase(node), local_scopes_(local_scopes), places_(places) {
if (ctxs) {
for (auto &p : places_) {
this->dev_ctxes_[p] = ctxs->DevCtx(p);
this->SetDeviceContext(p, ctxs->DevCtx(p));
}
}
}
......@@ -89,8 +89,8 @@ void DataBalanceOpHandle::RunImpl() {
PADDLE_ENFORCE_GT(places_.size(), 1,
"Data balance can only be enabled when the number of "
"places to run larger than 1.");
auto in_var_handles = DynamicCast<VarHandle>(inputs_);
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE(in_var_handles.size() % places_.size() == 0);
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.cc
浏览文件 @ cf1944af
......@@ -92,13 +92,13 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
size_t num_complete = 0;
remaining_ = 0;
BlockingQueue<size_t> complete_q;
auto complete_q = std::make_shared<BlockingQueue<size_t>>();
for (auto op : bootstrap_ops_) {
RunOpAsync(op_deps.get(), op, &complete_q);
RunOpAsync(op_deps.get(), op, complete_q);
}
while (num_complete != op_deps->size()) {
size_t num_comp = complete_q.Pop();
size_t num_comp = complete_q->Pop();
if (num_comp == -1UL) {
int remaining = 0;
while (true) {
......@@ -107,7 +107,7 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
break;
}
for (int i = 0; i < remaining; ++i) {
complete_q.Pop();
complete_q->Pop();
}
}
exception_.ReThrow();
......@@ -120,7 +120,8 @@ FeedFetchList FastThreadedSSAGraphExecutor::Run(
}
void FastThreadedSSAGraphExecutor::RunOpAsync(
std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
OpHandleBase *op, BlockingQueue<size_t> *complete_q) {
OpHandleBase *op,
const std::shared_ptr<BlockingQueue<size_t>> &complete_q) {
++remaining_;
this->pool_.enqueue([=] {
OpHandleBase *op_to_run = op;
......@@ -144,7 +145,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
if (op_to_run == nullptr) {
op_to_run = pending_op;
} else {
this->RunOpAsync(op_deps, pending_op, complete_q);
RunOpAsync(op_deps, pending_op, complete_q);
}
}
}
......@@ -156,8 +157,7 @@ void FastThreadedSSAGraphExecutor::RunOpAsync(
}
void FastThreadedSSAGraphExecutor::PrepareAtomicOpDeps() {
atomic_op_deps_ = pool_.enqueue([&] {
std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps =
new std::unordered_map<OpHandleBase *, std::atomic<int>>;
auto *op_deps = new std::unordered_map<OpHandleBase *, std::atomic<int>>;
for (auto &pair : op_deps_) {
(*op_deps)[pair.first] = pair.second;
}
......
paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h
浏览文件 @ cf1944af
......@@ -50,7 +50,8 @@ class FastThreadedSSAGraphExecutor : public SSAGraphExecutor {
std::atomic<int> remaining_;
void RunOpAsync(std::unordered_map<OpHandleBase *, std::atomic<int>> *op_deps,
OpHandleBase *op, BlockingQueue<size_t> *complete_q);
OpHandleBase *op,
const std::shared_ptr<BlockingQueue<size_t>> &complete_q);
void PrepareAtomicOpDeps();
......
paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc 0 → 100644
浏览文件 @ cf1944af
// 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/framework/details/fused_broadcast_op_handle.h"
#include "gtest/gtest.h"
#include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
namespace paddle {
namespace framework {
namespace details {
struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
std::vector<std::string> out_varnames_;
void InitFusedBroadcastOp(std::vector<size_t> input_scope_idxes) {
// initialize scope and var
for (size_t i = 0; i < place_list_.size(); ++i) {
local_scopes_.push_back(&(g_scope_.NewScope()));
Scope& local_scope = local_scopes_.back()->NewScope();
*local_scopes_.back()
->Var(details::kLocalExecScopeName)
->GetMutable<Scope*>() = &local_scope;
for (size_t j = 0; j < input_scope_idxes.size(); ++j) {
local_scope.Var("out_var" + j);
if (i == j) local_scope.Var("in_var" + j);
}
param_scopes_.emplace_back(&local_scope);
}
// create op handle node
std::unique_ptr<ir::Node> n =
ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation);
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new FusedBroadcastOpHandle(
n.get(), local_scopes_, place_list_, nccl_ctxs_.get()));
#else
PADDLE_THROW("CUDA is not supported.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
op_handle_.reset(new FusedBroadcastOpHandle(
n.get(), local_scopes_, place_list_, nccl_ctxs_.get()));
#else
op_handle_.reset(
new FusedBroadcastOpHandle(n.get(), local_scopes_, place_list_));
#endif
}
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
// add input var handle
std::unique_ptr<ir::Node> in_node =
ir::CreateNodeForTest("in_node" + i, ir::Node::Type::kVariable);
VarHandle* in_var_handle =
new VarHandle(in_node.get(), 1, input_scope_idxes[i], "in_var" + i,
place_list_[input_scope_idxes[i]]);
vars_.emplace_back(in_var_handle);
op_handle_->AddInput(in_var_handle);
// add output var handle
for (size_t j = 0; j < place_list_.size(); ++j) {
std::unique_ptr<ir::Node> out_node =
ir::CreateNodeForTest("out_node" + i, ir::Node::Type::kVariable);
VarHandle* out_var_handle =
new VarHandle(out_node.get(), 2, j, "out_var" + i, place_list_[j]);
vars_.emplace_back(out_var_handle);
op_handle_->AddOutput(out_var_handle);
}
}
}
void TestFusedBroadcastLoDTensor(std::vector<size_t> input_scope_idxes) {
std::vector<std::vector<float>> send_vec;
f::LoD lod{{0, 10, 20}};
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i);
float val_scalar = static_cast<float>(i);
send_vec.push_back(
InitLoDTensor(varname, input_scope_idxes[i], lod, val_scalar));
}
op_handle_->Run(false);
WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i);
for (size_t j = 0; j < place_list_.size(); ++j) {
LoDTensorEqual(varname, send_vec[i], lod, param_scopes_[j]);
}
}
}
void TestFusedBroadcastSelectedRows(std::vector<size_t> input_scope_idxes) {
std::vector<std::vector<float>> send_vector;
std::vector<int64_t> rows{0, 1, 2, 3, 3, 0, 14, 7, 3, 1,
2, 4, 6, 3, 1, 1, 1, 1, 3, 7};
int height = static_cast<int>(kDims[0] * 2);
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string varname("in_var" + i);
float val_scalar = static_cast<float>(i);
send_vector.push_back(InitSelectedRows(varname, input_scope_idxes[i],
rows, height, val_scalar));
}
op_handle_->Run(false);
WaitAll();
for (size_t i = 0; i < input_scope_idxes.size(); ++i) {
const std::string& varname("out_var" + i);
for (size_t j = 0; j < place_list_.size(); ++j) {
SelectedRowsEqual(varname, input_scope_idxes[i], send_vector[i], rows,
height);
}
}
}
};
TEST(FusedBroadcastTester, CPULodTensor) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(false);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastLoDTensor(input_scope_idxes);
}
TEST(FusedBroadcastTester, CPUSelectedRows) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(false);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastSelectedRows(input_scope_idxes);
}
#ifdef PADDLE_WITH_CUDA
TEST(FusedBroadcastTester, GPULodTensor) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(true);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastLoDTensor(input_scope_idxes);
}
TEST(FusedBroadcastTester, GPUSelectedRows) {
TestFusedBroadcastOpHandle test_op;
std::vector<size_t> input_scope_idxes = {0, 1};
test_op.InitCtxOnGpu(true);
test_op.InitFusedBroadcastOp(input_scope_idxes);
test_op.TestFusedBroadcastSelectedRows(input_scope_idxes);
}
#endif
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/gather_op_handle.cc
浏览文件 @ cf1944af
......@@ -36,7 +36,7 @@ void GatherOpHandle::RunImpl() {
VarHandle *out_var_handle;
{
auto out_var_handles = DynamicCast<VarHandle>(outputs_);
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(out_var_handles.size(), 1,
"The number of output should be one.");
out_var_handle = out_var_handles.front();
......@@ -99,7 +99,7 @@ void GatherOpHandle::RunImpl() {
Tensor *out_tensor = out_value->mutable_value();
// copy
auto dev_ctx = dev_ctxes_[out_var_handle->place_];
auto dev_ctx = dev_ctxes_.at(out_var_handle->place_);
RunAndRecordEvent(out_var_handle->place_, [in_tensors, out_tensor, &dev_ctx,
t_out_p] {
int s = 0, e = 0;
......
paddle/fluid/framework/details/op_handle_base.cc
浏览文件 @ cf1944af
......@@ -103,7 +103,7 @@ void OpHandleBase::WaitInputVarGenerated() {
void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
for (auto *in : inputs_) {
if (NeedWait(in)) {
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[place]);
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(place));
}
}
}
......
paddle/fluid/framework/details/reduce_op_handle.cc
浏览文件 @ cf1944af
......@@ -27,7 +27,7 @@ namespace framework {
namespace details {
void ReduceOpHandle::RunImpl() {
platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
platform::RecordEvent record_event(Name(), dev_ctxes_.cbegin()->second);
if (places_.size() == 1) return;
// the input and output may have dummy var.
......
paddle/fluid/framework/details/reduce_op_handle.h
浏览文件 @ cf1944af
......@@ -46,7 +46,8 @@ struct ReduceOpHandle : public OpHandleBase {
nccl_ctxs_(nccl_ctxs) {
if (nccl_ctxs_) {
for (auto &p_ctx : nccl_ctxs_->contexts_) {
dev_ctxes_[platform::CUDAPlace(p_ctx.first)] = p_ctx.second.ctx_.get();
this->SetDeviceContext(platform::CUDAPlace(p_ctx.first),
p_ctx.second.ctx_.get());
}
}
}
......
paddle/fluid/framework/details/rpc_op_handle.cc
浏览文件 @ cf1944af
......@@ -38,7 +38,7 @@ void RPCOpHandle::RunImpl() {
continue;
}
if (in->GeneratedOp()) {
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_[p]);
in->GeneratedOp()->RecordWaitEventOnCtx(dev_ctxes_.at(p));
}
}
auto &tmp_scope = local_scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
......
paddle/fluid/framework/details/scale_loss_grad_op_handle.cc
浏览文件 @ cf1944af
......@@ -27,7 +27,7 @@ ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev,
coeff_(static_cast<float>(1.0 / num_dev)),
scope_(scope),
place_(place) {
dev_ctxes_[place_] = dev_ctx;
this->SetDeviceContext(place_, dev_ctx);
}
ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
......@@ -46,9 +46,9 @@ void ScaleLossGradOpHandle::RunImpl() {
} else {
#ifdef PADDLE_WITH_CUDA
this->RunAndRecordEvent([&] {
auto stream =
static_cast<platform::CUDADeviceContext *>(this->dev_ctxes_[place_])
->stream();
auto stream = static_cast<platform::CUDADeviceContext *>(
this->dev_ctxes_.at(place_))
->stream();
memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
platform::CPUPlace(), &coeff_, sizeof(float), stream);
VLOG(10) << place_ << "RUN Scale loss grad op";
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ cf1944af
......@@ -39,7 +39,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
new platform::RecordEvent("ThreadedSSAGraphExecutorPrepare", nullptr));
std::unordered_map<OpHandleBase *, size_t> pending_ops;
std::unordered_set<VarHandleBase *> pending_vars;
BlockingQueue<VarHandleBase *> ready_vars;
auto ready_vars = std::make_shared<BlockingQueue<VarHandleBase *>>();
std::unordered_set<OpHandleBase *> ready_ops;
// For ops (e.g. nccl_all_reduce) that need to coordinate multiple
// streams from multiple GPUs, it's faster to buffer them and schedule
......@@ -51,12 +51,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
for (auto &name_pair : var_map) {
for (auto &version_pair : name_pair.second) {
InsertPendingVar(&pending_vars, &ready_vars, version_pair.get());
InsertPendingVar(&pending_vars, ready_vars.get(), version_pair.get());
}
}
}
for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
InsertPendingVar(&pending_vars, &ready_vars, var.get());
InsertPendingVar(&pending_vars, ready_vars.get(), var.get());
}
for (auto &op : graph_->Get<details::GraphOps>(details::kGraphOps)) {
......@@ -73,12 +73,12 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
FeedFetchList fetch_data(fetch_tensors.size());
InsertFetchOps(fetch_tensors, &fetch_ops, &fetch_dependencies, &pending_ops,
&pending_vars, &ready_vars, &fetch_data);
&pending_vars, ready_vars.get(), &fetch_data);
auto run_all_ops = [&](std::unordered_set<OpHandleBase *> &set) {
for (auto *op : set) {
running_ops_++;
RunOp(&ready_vars, op);
RunOp(ready_vars, op);
}
set.clear();
};
......@@ -87,7 +87,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
run_op_futures_.clear();
exception_holder_.Clear();
event.reset(nullptr);
// Step 3. Execution
while (!pending_vars.empty()) {
// 1. Run All Ready ops
......@@ -103,7 +102,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
// 2. Find ready variable
bool timeout;
auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
auto cur_ready_vars = ready_vars->PopAll(1, &timeout);
if (timeout) {
if (exception_holder_.IsCaught()) {
......@@ -133,7 +132,6 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
}
}
PADDLE_ENFORCE(ready_ops.empty());
// Wait FetchOps.
ClearFetchOp(graph_.get(), &fetch_ops);
......@@ -206,7 +204,8 @@ void ThreadedSSAGraphExecutor::InsertPendingVar(
}
void ThreadedSSAGraphExecutor::RunOp(
BlockingQueue<VarHandleBase *> *ready_var_q, details::OpHandleBase *op) {
const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
details::OpHandleBase *op) {
auto op_run = [ready_var_q, op, this] {
try {
if (VLOG_IS_ON(10)) {
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @ cf1944af
......@@ -51,7 +51,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
~ThreadedSSAGraphExecutor() {}
private:
void RunOp(BlockingQueue<VarHandleBase *> *ready_var_q,
void RunOp(const std::shared_ptr<BlockingQueue<VarHandleBase *>> &ready_var_q,
details::OpHandleBase *op);
private:
......
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @ cf1944af
......@@ -259,6 +259,15 @@ GraphPatternDetector::DetectPatterns() {
return result;
}
bool GraphItemCMP(const std::pair<PDNode *, Node *> &a,
const std::pair<PDNode *, Node *> &b) {
if (a.first != b.first) {
return a.first < b.first;
} else {
return a.second < b.second;
}
}
// TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550
void GraphPatternDetector::UniquePatterns(
......@@ -267,12 +276,16 @@ void GraphPatternDetector::UniquePatterns(
std::vector<GraphPatternDetector::subgraph_t> result;
std::unordered_set<size_t> set;
std::hash<std::string> hasher;
for (auto &g : *subgraphs) {
size_t key = 0;
for (auto &item : g) {
key ^= std::hash<void *>{}(item.first);
key ^= std::hash<void *>{}(item.second);
// Sort the items in the sub-graph, and transform to a string key.
std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
std::stringstream ss;
for (auto &item : sorted_keys) {
ss << item.first << ":" << item.second;
}
auto key = hasher(ss.str());
if (!set.count(key)) {
result.emplace_back(g);
set.insert(key);
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ cf1944af
......@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor(
PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
for (size_t j = 0; j < lod.size(); ++j) {
auto &sub_lod = new_lod[j];
auto &offset = sub_lod.back();
size_t offset = sub_lod.back();
for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset);
}
......
paddle/fluid/framework/lod_tensor_array.h
浏览文件 @ cf1944af
......@@ -19,81 +19,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
// NOTE The vector<LoDTensor> can't be replaced with the class LoDTensorArray
// directly, because there are many vector<LoDTensor> used accross the project,
// and some of them are treated as LoDTensorArray.
#if !defined(PADDLE_ON_INFERENCE)
using LoDTensorArray = std::vector<LoDTensor>;
#else // !PADDLE_ON_INFERENCE
#pragma message "LoDTensorArray is replaced with the inference one."
/*
* A LoDTensorArray which will not deallocate buffer when resized, fix the data
* diff in inference, and more performance friendly in the concurrency
* scenerios.
*/
class LoDTensorArray {
public:
LoDTensorArray() = default;
using iterator = std::vector<LoDTensor>::iterator;
using const_iterator = std::vector<LoDTensor>::const_iterator;
const_iterator begin() const { return array_.begin(); }
const_iterator end() const { return array_.begin() + size_; }
iterator begin() { return array_.begin(); }
iterator end() { return array_.begin() + size_; }
void push_back(const LoDTensor& x) {
if (size_ < array_.size()) {
array_[size_++] = x;
} else {
array_.push_back(x);
++size_;
}
}
void resize(size_t size) {
if (array_.size() < size) {
array_.resize(size);
}
size_ = size;
}
void emplace_back() { array_.emplace_back(); }
void emplace_back(LoDTensor&& x) { array_.emplace_back(std::move(x)); }
LoDTensor& back() { return array_.back(); }
size_t space() const { return array_.size(); }
void reserve(size_t size) {
// Naive warning to tell user this array might be to large. The memory and
// buffer used by this TensorArray will not be deleted during the training
// and inference phase, so attention not to make it expand too long.
if (size > 800UL) {
LOG(WARNING) << "TensorArray has more than 800 items";
}
array_.reserve(size);
}
bool empty() const { return size_ == 0UL; }
void clear() { size_ = 0UL; }
LoDTensor& operator[](size_t id) { return array_[id]; }
const LoDTensor& operator[](size_t id) const { return array_[id]; }
LoDTensor& at(size_t id) { return array_.at(id); }
const LoDTensor& at(size_t id) const { return array_.at(id); }
size_t size() const { return size_; }
private:
size_t size_{0};
std::vector<LoDTensor> array_;
};
#endif // !PADDLE_ON_INFERENCE
} // namespace framework
} // namespace paddle
paddle/fluid/framework/operator.cc
浏览文件 @ cf1944af
......@@ -354,18 +354,18 @@ void OperatorBase::GenerateTemporaryNames() {
}
}
static bool VarIsTensor(const Variable* var) {
return var->IsType<LoDTensor>() || var->IsType<SelectedRows>();
static bool VarIsTensor(const Variable& var) {
return var.IsType<LoDTensor>() || var.IsType<SelectedRows>();
}
static const Tensor* GetTensorFromVar(Variable* var) {
if (var->IsType<LoDTensor>()) {
return var->GetMutable<LoDTensor>();
} else if (var->IsType<SelectedRows>()) {
return var->GetMutable<SelectedRows>()->mutable_value();
const Tensor* GetTensorFromVar(const Variable& var) {
if (var.IsType<LoDTensor>()) {
return static_cast<const Tensor*>(&(var.Get<LoDTensor>()));
} else if (var.IsType<SelectedRows>()) {
return &(var.Get<SelectedRows>().value());
} else {
PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.",
var->Type().name());
var.Type().name());
}
}
......@@ -415,8 +415,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const {
template <>
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
auto* var = InputVar(name);
return var == nullptr ? nullptr
: GetTensorFromVar(const_cast<Variable*>(var));
return var == nullptr ? nullptr : GetTensorFromVar(*var);
}
template <>
......@@ -428,7 +427,7 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) {
auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr : GetTensorFromVar(var);
return var == nullptr ? nullptr : GetTensorFromVar(*var);
});
return res;
}
......@@ -770,8 +769,10 @@ void OperatorWithKernel::TransferInplaceVarsBack(
for (auto& var_name : inplace_vars) {
VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
auto* transformed_tensor =
GetTensorFromVar(transfer_scope.FindVar(var_name));
auto* var = transfer_scope.FindVar(var_name);
PADDLE_ENFORCE(var != nullptr, "The var[%s] should not be nullptr",
var_name);
auto* transformed_tensor = GetTensorFromVar(*var);
original_tensor->ShareDataWith(*transformed_tensor);
}
}
......@@ -784,11 +785,11 @@ Scope* OperatorWithKernel::TryTransferData(
for (auto& var_name : var_name_item.second) {
auto* var = scope.FindVar(var_name);
// Only tensor can be tranfer to another device.
if (var == nullptr || !VarIsTensor(var)) {
if (var == nullptr || !VarIsTensor(*var)) {
continue;
}
auto* tensor_in = GetTensorFromVar(var);
auto* tensor_in = GetTensorFromVar(*var);
if (!tensor_in->IsInitialized()) {
continue;
}
......
paddle/fluid/framework/operator.h
浏览文件 @ cf1944af
......@@ -63,6 +63,7 @@ inline std::string GradVarName(const std::string& var_name) {
}
proto::VarType::Type GetDataTypeOfVar(const Variable* var);
const Tensor* GetTensorFromVar(const Variable& var);
class OperatorBase;
class ExecutionContext;
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ cf1944af
......@@ -303,10 +303,8 @@ void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
}
ParallelExecutor::~ParallelExecutor() {
const auto dev_ctxs =
platform::DeviceContextPool::Instance().GetAllDeviceContexts();
for (auto &dev_ctx : dev_ctxs) {
dev_ctx->Wait();
for (auto &p : member_->places_) {
platform::DeviceContextPool::Instance().Get(p)->Wait();
}
if (member_->own_local_scope_) {
......
paddle/fluid/framework/tensor_test.cc
浏览文件 @ cf1944af
......@@ -75,6 +75,19 @@ TEST(Tensor, MutableData) {
platform::CPUPlace());
EXPECT_EQ(p1, p2);
}
// Not sure if it's desired, but currently, Tensor type can be changed.
{
framework::Tensor src_tensor;
int8_t* p1 = src_tensor.mutable_data<int8_t>(framework::make_ddim({1}),
platform::CPUPlace());
EXPECT_NE(p1, nullptr);
*p1 = 1;
uint8_t* p2 = src_tensor.mutable_data<uint8_t>(framework::make_ddim({1}),
platform::CPUPlace());
EXPECT_NE(p2, nullptr);
EXPECT_EQ(static_cast<int>(p2[0]), 1);
}
#ifdef PADDLE_WITH_CUDA
{
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ cf1944af
if(WITH_TESTING)
include(test.cmake) # some generic cmake funtion for inference
endif()
# analysis and tensorrt must be added before creating static library,
# otherwise, there would be undefined reference to them in static library.
add_subdirectory(analysis)
......
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @ cf1944af
......@@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis)
cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
function (inference_analysis_test TARGET)
if(WITH_TESTING)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS ARGS EXTRA_DEPS)
cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(mem_opt "")
if(WITH_GPU)
set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
endif()
cc_test(${TARGET}
SRCS "${analysis_test_SRCS}"
DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS})
set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
endif(WITH_TESTING)
function(inference_analysis_test TARGET)
if(WITH_TESTING)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS ARGS EXTRA_DEPS)
cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
inference_base_test(${TARGET}
SRCS ${analysis_test_SRCS}
DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS})
endif()
endfunction(inference_analysis_test)
inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)
......
paddle/fluid/inference/api/CMakeLists.txt
浏览文件 @ cf1944af
......@@ -17,39 +17,12 @@ if(APPLE)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
endif(APPLE)
set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB}
)
set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager naive_executor ${GLOB_PASS_LIB})
if(WITH_GPU AND TENSORRT_FOUND)
set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor)
endif()
function(inference_api_test TARGET_NAME)
if (WITH_TESTING)
set(options "")
set(oneValueArgs SRC)
set(multiValueArgs ARGS)
cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if (WITH_GPU)
cc_test(${TARGET_NAME}
SRCS ${inference_test_SRC}
DEPS "${inference_deps}"
ARGS --dirname=${PYTHON_TESTS_DIR}/book/ --fraction_of_gpu_memory_to_use=0.15)
else()
cc_test(${TARGET_NAME}
SRCS ${inference_test_SRC}
DEPS "${inference_deps}"
ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
endif()
if(inference_test_ARGS)
set_tests_properties(${TARGET_NAME}
PROPERTIES DEPENDS "${inference_test_ARGS}")
endif()
endif(WITH_TESTING)
endfunction(inference_api_test)
cc_library(reset_tensor_array SRCS details/reset_tensor_array.cc DEPS lod_tensor scope)
cc_library(paddle_inference_api SRCS api.cc api_impl.cc helper.cc DEPS reset_tensor_array lod_tensor scope)
cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor)
......@@ -59,10 +32,11 @@ cc_test(test_paddle_inference_api
SRCS api_tester.cc
DEPS paddle_inference_api)
inference_api_test(test_api_impl SRC api_impl_tester.cc
ARGS test_word2vec test_image_classification)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
if(WITH_TESTING)
inference_base_test(test_api_impl SRCS api_impl_tester.cc DEPS ${inference_deps}
ARGS --word2vec_dirname=${WORD2VEC_MODEL_DIR} --book_dirname=${PYTHON_TESTS_DIR}/book)
set_tests_properties(test_api_impl PROPERTIES DEPENDS test_image_classification)
endif()
cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api
ARGS --dirname=${PYTHON_TESTS_DIR}/book)
......@@ -70,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND)
cc_library(paddle_inference_tensorrt_subgraph_engine
SRCS api_tensorrt_subgraph_engine.cc
DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy)
inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec)
if(WITH_TESTING)
inference_base_test(test_api_tensorrt_subgraph_engine SRCS api_tensorrt_subgraph_engine_tester.cc DEPS ${inference_deps}
ARGS --dirname=${WORD2VEC_MODEL_DIR})
endif()
endif()
if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
......
paddle/fluid/inference/api/api_impl_tester.cc
浏览文件 @ cf1944af
......@@ -27,7 +27,9 @@ limitations under the License. */
#define ACC_DIFF 1e-3
#endif
DEFINE_string(dirname, "", "Directory of the inference model.");
DEFINE_string(word2vec_dirname, "",
"Directory of the word2vec inference model.");
DEFINE_string(book_dirname, "", "Directory of the book inference model.");
namespace paddle {
......@@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
NativeConfig GetConfig() {
NativeConfig config;
config.model_dir = FLAGS_dirname + "/word2vec.inference.model";
config.model_dir = FLAGS_word2vec_dirname;
LOG(INFO) << "dirname " << config.model_dir;
config.fraction_of_gpu_memory = 0.15;
#ifdef PADDLE_WITH_CUDA
......@@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) {
NativeConfig config = GetConfig();
config.use_gpu = use_gpu;
config.model_dir =
FLAGS_dirname + "/image_classification_resnet.inference.model";
FLAGS_book_dirname + "/image_classification_resnet.inference.model";
const bool is_combined = false;
std::vector<std::vector<int64_t>> feed_target_shapes =
......@@ -187,7 +189,7 @@ void MainThreadsWord2Vec(bool use_gpu) {
std::vector<std::thread> threads;
for (int tid = 0; tid < num_jobs; ++tid) {
threads.emplace_back([&, tid]() {
auto predictor = main_predictor->Clone();
auto predictor = CreatePaddlePredictor(config);
auto& local_inputs = paddle_tensor_feeds[tid];
std::vector<PaddleTensor> local_outputs;
ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs));
......@@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) {
NativeConfig config = GetConfig();
config.use_gpu = use_gpu;
config.model_dir =
FLAGS_dirname + "/image_classification_resnet.inference.model";
FLAGS_book_dirname + "/image_classification_resnet.inference.model";
auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
std::vector<framework::LoDTensor> jobs(num_jobs);
......@@ -245,7 +247,7 @@ void MainThreadsImageClassification(bool use_gpu) {
std::vector<std::thread> threads;
for (int tid = 0; tid < num_jobs; ++tid) {
threads.emplace_back([&, tid]() {
auto predictor = main_predictor->Clone();
auto predictor = CreatePaddlePredictor(config);
auto& local_inputs = paddle_tensor_feeds[tid];
std::vector<PaddleTensor> local_outputs;
ASSERT_TRUE(predictor->Run(local_inputs, &local_outputs));
......
paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc
浏览文件 @ cf1944af
......@@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
//# 1. Create PaddlePredictor with a config.
NativeConfig config0;
config0.model_dir = FLAGS_dirname + "word2vec.inference.model";
config0.model_dir = FLAGS_dirname;
config0.use_gpu = true;
config0.fraction_of_gpu_memory = 0.3;
config0.device = 0;
MixedRTConfig config1;
config1.model_dir = FLAGS_dirname + "word2vec.inference.model";
config1.model_dir = FLAGS_dirname;
config1.use_gpu = true;
config1.fraction_of_gpu_memory = 0.3;
config1.device = 0;
......
paddle/fluid/inference/api/demo_ci/run.sh
浏览文件 @ cf1944af
......@@ -62,7 +62,7 @@ for WITH_STATIC_LIB in ON OFF; do
-DWITH_GPU=$TEST_GPU_CPU \
-DWITH_STATIC_LIB=$WITH_STATIC_LIB
make -j
word2vec_model=${PADDLE_ROOT}'/build/python/paddle/fluid/tests/book/word2vec.inference.model'
word2vec_model=$DATA_DIR'/word2vec/word2vec.inference.model'
if [ -d $word2vec_model ]; then
for use_gpu in $use_gpu_list; do
./simple_on_word2vec \
......
paddle/fluid/inference/test.cmake 0 → 100644
浏览文件 @ cf1944af
set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url")
set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
"A path setting inference demo download directories.")
function (inference_download install_dir url filename)
message(STATUS "Download inference test stuff from ${url}/${filename}")
execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
message(STATUS "finish downloading ${filename}")
endfunction()
function (inference_download_and_uncompress install_dir url filename)
inference_download(${install_dir} ${url} ${filename})
execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
endfunction()
set(WORD2VEC_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/word2vec")
if (NOT EXISTS ${WORD2VEC_INSTALL_DIR})
inference_download_and_uncompress(${WORD2VEC_INSTALL_DIR} ${INFERENCE_URL} "word2vec.inference.model.tar.gz")
endif()
set(WORD2VEC_MODEL_DIR "${WORD2VEC_INSTALL_DIR}/word2vec.inference.model")
function (inference_base_test TARGET)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS ARGS DEPS)
cmake_parse_arguments(base_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if(WITH_GPU)
set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
endif()
cc_test(${TARGET} SRCS ${base_test_SRCS} DEPS ${base_test_DEPS} ARGS ${mem_opt} ${base_test_ARGS})
endfunction()
paddle/fluid/inference/tests/api/CMakeLists.txt
浏览文件 @ cf1944af
set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com")
set(INFERENCE_DEMO_INSTALL_DIR "${THIRD_PARTY_PATH}/inference_demo" CACHE STRING
"A path setting inference demo download directories.")
set(INFERENCE_EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor)
function (inference_download install_dir url filename)
message(STATUS "Download inference test stuff from ${url}/${filename}")
execute_process(COMMAND bash -c "mkdir -p ${install_dir}")
execute_process(COMMAND bash -c "cd ${install_dir} && wget -q ${url}/${filename}")
message(STATUS "finish downloading ${filename}")
endfunction()
function (inference_download_and_uncompress install_dir url filename)
inference_download(${install_dir} ${url} ${filename})
execute_process(COMMAND bash -c "cd ${install_dir} && tar xzf ${filename}")
endfunction()
function(download_model_and_data install_dir model_name data_name)
if (NOT EXISTS ${install_dir})
......
paddle/fluid/operators/add_position_encoding_op.cc 0 → 100644
浏览文件 @ cf1944af
/* 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/add_position_encoding_op.h"
namespace paddle {
namespace operators {
class AddPositionEncodingOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"X(Input) of add_position_encoding_op should not be null.");
PADDLE_ENFORCE(
ctx->HasOutput("Out"),
"Out(Output) of add_position_encoding_op should not be null.");
auto x_dims = ctx->GetInputDim("X");
ctx->SetOutputDim("Out", x_dims);
ctx->ShareLoD("X", /*->*/ "Out");
}
};
class AddPositionEncodingOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "X(Input) must not be null.");
PADDLE_ENFORCE(ctx->HasInput("Out"), "Out must not be null.");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
"Out@GRAD must not be null.");
auto out_dims = ctx->GetInputDim("Out");
if (ctx->HasOutput(framework::GradVarName("X"))) {
ctx->SetOutputDim(framework::GradVarName("X"), out_dims);
}
}
};
class AddPositionEncodingOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "Input of AddPositionEncoding operator");
AddOutput("Out", "Output of AddPositionEncoding operator");
AddAttr<float>("alpha", "The scale of Original Embedding.")
.SetDefault(1.0f)
.AddCustomChecker([](const float& alpha) {
PADDLE_ENFORCE(alpha >= 0.0f, "'alpha' must be above 0.0.");
});
AddAttr<float>("beta", "The scale of Position Embedding.")
.SetDefault(1.0f)
.AddCustomChecker([](const float& beta) {
PADDLE_ENFORCE(beta >= 0.0f, "'beta' must be between 0.0.");
});
AddComment(R"DOC(
Add Position Encoding Operator.
The add position encoding calculates the output based on the input, alpha, beta.
The size of each dimension of the parameters checked in the infer-shape.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
namespace plt = paddle::platform;
REGISTER_OPERATOR(add_position_encoding, ops::AddPositionEncodingOp,
ops::AddPositionEncodingOpMaker,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(add_position_encoding_grad, ops::AddPositionEncodingOpGrad);
REGISTER_OP_CPU_KERNEL(
add_position_encoding,
ops::AddPositionEncodingKernel<plt::CPUDeviceContext, float>,
ops::AddPositionEncodingKernel<plt::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
add_position_encoding_grad,
ops::AddPositionEncodingGradKernel<plt::CPUDeviceContext, float>,
ops::AddPositionEncodingGradKernel<plt::CPUDeviceContext, double>);
paddle/fluid/operators/add_position_encoding_op.h 0 → 100644
浏览文件 @ cf1944af
/* 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 "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
namespace paddle {
namespace operators {
template <typename DeviceContext, typename T>
class AddPositionEncodingKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* X = context.Input<framework::LoDTensor>("X");
auto& x_lod = X->lod();
auto* src_ptr = X->data<T>();
auto* Out = context.Output<framework::LoDTensor>("Out");
auto* dst_ptr = Out->mutable_data<T>(context.GetPlace());
float alpha = context.Attr<float>("alpha");
float beta = context.Attr<float>("beta");
auto x_dim = X->dims();
int batch_size = 0;
int max_seq_len = 0;
int enc_size = 0;
if (x_lod.empty()) {
PADDLE_ENFORCE(
x_dim.size() == 3UL,
"The input X of Add Position Encoding should be 3-D Tensor!");
batch_size = x_dim[0];
max_seq_len = x_dim[1];
enc_size = x_dim[2];
} else {
PADDLE_ENFORCE(
x_dim.size() == 2UL,
"The input X of Add Position Encoding should be 2-D LoDTensor!");
PADDLE_ENFORCE(
x_lod.size() == 1UL,
"The Add Position Encoding Op only supports lod_level == 1!");
batch_size = x_lod[0].size() - 1;
max_seq_len = -1;
enc_size = x_dim[1];
}
PADDLE_ENFORCE(enc_size % 2 == 0, "Only support even encode size!");
const int half_size = enc_size / 2;
for (int i = 0; i < batch_size; ++i) {
const int max_length =
x_lod.empty() ? max_seq_len : x_lod[0][i + 1] - x_lod[0][i];
for (int j = 0; j < max_length; ++j) {
for (int k = 0; k < half_size; ++k) {
const double val = (half_size > 1)
? j / pow(10000.0, double(k) / (half_size - 1))
: j / 10000.0;
dst_ptr[k] = src_ptr[k] * alpha + sin(val) * beta;
dst_ptr[half_size + k] =
src_ptr[half_size + k] * alpha + cos(val) * beta;
}
src_ptr += enc_size;
dst_ptr += enc_size;
}
}
}
};
template <typename DeviceContext, typename T>
class AddPositionEncodingGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* dOut =
context.Input<framework::LoDTensor>(framework::GradVarName("Out"));
auto dout = framework::EigenVector<T>::Flatten(*dOut);
auto* dX =
context.Output<framework::LoDTensor>(framework::GradVarName("X"));
dX->mutable_data<T>(context.GetPlace());
auto dx = framework::EigenVector<T>::Flatten(*dX);
float alpha = context.Attr<float>("alpha");
auto* place =
context.template device_context<DeviceContext>().eigen_device();
dx.device(*place) = dout * static_cast<T>(alpha);
}
};
} // namespace operators
} // namespace paddle
paddle/fluid/operators/gather_op.cc
浏览文件 @ cf1944af
......@@ -102,7 +102,9 @@ REGISTER_OPERATOR(gather, ops::GatherOp, ops::GatherOpMaker,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(gather_grad, ops::GatherGradOp);
REGISTER_OP_CPU_KERNEL(gather, ops::GatherOpKernel<float>,
ops::GatherOpKernel<int>, ops::GatherOpKernel<double>);
ops::GatherOpKernel<double>, ops::GatherOpKernel<int>,
ops::GatherOpKernel<int64_t>);
REGISTER_OP_CPU_KERNEL(gather_grad, ops::GatherGradientOpKernel<float>,
ops::GatherGradientOpKernel<double>,
ops::GatherGradientOpKernel<int>,
ops::GatherGradientOpKernel<double>);
ops::GatherGradientOpKernel<int64_t>);
paddle/fluid/operators/gather_op.cu
浏览文件 @ cf1944af
......@@ -61,5 +61,11 @@ class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel<float>);
REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel<float>);
REGISTER_OP_CUDA_KERNEL(gather, ops::GatherOpCUDAKernel<float>,
ops::GatherOpCUDAKernel<double>,
ops::GatherOpCUDAKernel<int64_t>,
ops::GatherOpCUDAKernel<int>);
REGISTER_OP_CUDA_KERNEL(gather_grad, ops::GatherGradOpCUDAKernel<float>,
ops::GatherGradOpCUDAKernel<double>,
ops::GatherGradOpCUDAKernel<int64_t>,
ops::GatherGradOpCUDAKernel<int>);
paddle/fluid/operators/math/sequence_pooling.cc
浏览文件 @ cf1944af
......@@ -31,7 +31,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename T>
template <typename T, bool is_test>
class MaxSeqPoolFunctor {
public:
void operator()(const platform::CPUDeviceContext& context,
......@@ -70,7 +70,41 @@ class MaxSeqPoolFunctor {
}
}
};
// Instantisation of Max Sequence Pooling for test phase eg. no need to fill
// index buffer
template <typename T>
class MaxSeqPoolFunctor<T, true> {
public:
void operator()(const platform::CPUDeviceContext& context,
const framework::LoDTensor& input, framework::Tensor* output,
framework::Tensor* index) {
auto in_dims = input.dims();
auto out_dims = output->dims();
PADDLE_ENFORCE_GT(in_dims.size(), 1);
PADDLE_ENFORCE_GT(out_dims.size(), 1);
for (int64_t i = 1; i < in_dims.size(); ++i) {
PADDLE_ENFORCE_EQ(in_dims[i], out_dims[i]);
}
auto starts = input.lod()[0];
const T* in_data = input.data<T>();
T* out_data = output->data<T>();
int64_t num_seq = out_dims[0];
int64_t dim = output->numel() / num_seq;
for (int64_t i = 0; i < num_seq; ++i) {
std::memcpy(&out_data[i * dim], &in_data[starts[i] * dim],
dim * sizeof(T));
for (size_t j = starts[i] + 1; j < starts[i + 1]; ++j) {
for (int64_t k = 0; k < dim; ++k) {
if (in_data[j * dim + k] > out_data[i * dim + k]) {
out_data[i * dim + k] = in_data[j * dim + k];
}
}
}
}
}
};
template <typename T>
class MaxSeqPoolGradFunctor {
public:
......@@ -188,11 +222,16 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
/* max pool has index output */
void operator()(const platform::CPUDeviceContext& context,
const std::string pooltype, const framework::LoDTensor& input,
framework::Tensor* output,
framework::Tensor* output, bool is_test,
framework::Tensor* index = nullptr) {
if (pooltype == "MAX") {
math::MaxSeqPoolFunctor<T> max_pool;
max_pool(context, input, output, index);
if (is_test) {
math::MaxSeqPoolFunctor<T, true> max_pool;
max_pool(context, input, output, index);
} else {
math::MaxSeqPoolFunctor<T, false> max_pool;
max_pool(context, input, output, index);
}
return;
}
if (pooltype == "LAST") {
......@@ -200,6 +239,7 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
last_pool(context, input, output);
return;
}
if (pooltype == "FIRST") {
math::FirstSeqPoolFunctor<T> first_pool;
first_pool(context, input, output);
......
paddle/fluid/operators/math/sequence_pooling.cu
浏览文件 @ cf1944af
......@@ -133,7 +133,7 @@ class SequencePoolFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const std::string pooltype, const framework::LoDTensor& input,
framework::Tensor* output,
framework::Tensor* output, bool is_test,
framework::Tensor* index = nullptr) {
auto& lod = input.lod()[0];
const size_t item_dim = output->numel() / output->dims()[0];
......
paddle/fluid/operators/math/sequence_pooling.h
浏览文件 @ cf1944af
......@@ -28,7 +28,7 @@ class SequencePoolFunctor {
/* max pool has index output */
void operator()(const DeviceContext& context, const std::string pooltype,
const framework::LoDTensor& input, framework::Tensor* output,
framework::Tensor* index = nullptr);
bool is_test = false, framework::Tensor* index = nullptr);
};
template <typename DeviceContext, typename T>
......
paddle/fluid/operators/sequence_pool_op.cc
浏览文件 @ cf1944af
......@@ -47,6 +47,7 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor<int>) This tensor is used for the sequence max-pooling "
"to record the max indexes.")
.AsIntermediate();
AddAttr<bool>("is_test", "").SetDefault(false);
AddAttr<std::string>(
"pooltype",
"(string, default 'AVERAGE') the pooling pooltype of SequencePoolOp.")
......
paddle/fluid/operators/sequence_pool_op.h
浏览文件 @ cf1944af
......@@ -32,10 +32,6 @@ class SequencePoolKernel : public framework::OpKernel<T> {
auto* in = context.Input<LoDTensor>("X");
auto* out = context.Output<Tensor>("Out");
std::string pooltype = context.Attr<std::string>("pooltype");
Tensor* index = nullptr;
if (pooltype == "MAX") {
index = context.Output<Tensor>("MaxIndex");
}
auto dims = in->dims();
auto lod = in->lod();
......@@ -48,13 +44,22 @@ class SequencePoolKernel : public framework::OpKernel<T> {
dims[0] = lod[0].size() - 1;
out->Resize({dims});
out->mutable_data<T>(context.GetPlace());
if (pooltype == "MAX") {
Tensor* index = nullptr;
const bool is_test = context.Attr<bool>("is_test");
// Do not create index buffer for inference (is_test) mode
// TODO(jczaja): Skip index buffer creation for other devices eg. GPU
if (pooltype == "MAX" &&
(is_test == false ||
platform::is_cpu_place(context.GetPlace()) == false)) {
index = context.Output<Tensor>("MaxIndex");
index->Resize({dims});
index->mutable_data<int>(context.GetPlace());
}
math::SequencePoolFunctor<DeviceContext, T> pool;
pool(context.template device_context<DeviceContext>(), pooltype, *in, out,
index);
is_test, index);
}
};
......
paddle/fluid/operators/sum_op.cc
浏览文件 @ cf1944af
......@@ -67,6 +67,7 @@ class SumOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto x_vars = ctx.MultiInputVar("X");
auto x_vars_name = ctx.Inputs("X");
framework::LibraryType library{framework::LibraryType::kPlain};
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
......@@ -81,15 +82,18 @@ class SumOp : public framework::OperatorWithKernel {
if (x_vars[0]->IsType<framework::LoDTensor>()) {
int dtype = -1;
for (auto& x_var : x_vars) {
auto& lod_tensor = x_var->Get<framework::LoDTensor>();
if (lod_tensor.numel() == 0) {
for (size_t idx = 0; idx < x_vars.size(); ++idx) {
PADDLE_ENFORCE(x_vars[idx] != nullptr,
"Input var[%s] should not be nullptr", x_vars_name[idx]);
// FIXME(zcd): The input x_var may be SelectedRows or LoDTensor.
auto tensor = framework::GetTensorFromVar(*x_vars[idx]);
if (tensor->numel() == 0) {
continue;
}
if (dtype == -1) {
dtype = framework::ToDataType(lod_tensor.type());
dtype = framework::ToDataType(tensor->type());
} else {
PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(lod_tensor.type()));
PADDLE_ENFORCE_EQ(dtype, framework::ToDataType(tensor->type()));
}
}
PADDLE_ENFORCE_NE(dtype, -1,
......
paddle/fluid/platform/device_context.cc
浏览文件 @ cf1944af
......@@ -32,23 +32,25 @@ platform::DeviceContext* DeviceContextPool::Get(const platform::Place& place) {
"'Place' is not supported, Please re-compile with WITH_GPU "
"option");
}
return it->second.get();
return it->second.get().get();
}
const std::vector<const DeviceContext*>
DeviceContextPool::GetAllDeviceContexts() const {
std::vector<const DeviceContext*> all_device_ctx;
all_device_ctx.reserve(device_contexts_.size());
for (auto& dev_ctx : device_contexts_) {
all_device_ctx.emplace_back(dev_ctx.second.get());
}
return all_device_ctx;
template <typename DevCtx, typename PlaceType>
inline void EmplaceDeviceContext(
std::map<Place, std::shared_future<std::unique_ptr<DeviceContext>>>*
map_ptr,
platform::Place p) {
using PtrType = std::unique_ptr<DeviceContext>;
map_ptr->emplace(p, std::async(std::launch::deferred, [=] {
// lazy evaluation. i.e., only create device context at
// first `Get`
return PtrType(new DevCtx(boost::get<PlaceType>(p)));
}));
}
DeviceContextPool::DeviceContextPool(
const std::vector<platform::Place>& places) {
PADDLE_ENFORCE_GT(places.size(), 0);
using PtrType = std::unique_ptr<DeviceContext>;
std::set<Place> set;
for (auto& p : places) {
set.insert(p);
......@@ -57,16 +59,13 @@ DeviceContextPool::DeviceContextPool(
for (auto& p : set) {
if (platform::is_cpu_place(p)) {
#ifdef PADDLE_WITH_MKLDNN
device_contexts_.emplace(
p, PtrType(new MKLDNNDeviceContext(boost::get<CPUPlace>(p))));
EmplaceDeviceContext<MKLDNNDeviceContext, CPUPlace>(&device_contexts_, p);
#else
device_contexts_.emplace(
p, PtrType(new CPUDeviceContext(boost::get<CPUPlace>(p))));
EmplaceDeviceContext<CPUDeviceContext, CPUPlace>(&device_contexts_, p);
#endif
} else if (platform::is_gpu_place(p)) {
#ifdef PADDLE_WITH_CUDA
device_contexts_.emplace(
p, PtrType(new CUDADeviceContext(boost::get<CUDAPlace>(p))));
EmplaceDeviceContext<CUDADeviceContext, CUDAPlace>(&device_contexts_, p);
#else
PADDLE_THROW(
"'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
......@@ -74,9 +73,8 @@ DeviceContextPool::DeviceContextPool(
#endif
} else if (platform::is_cuda_pinned_place(p)) {
#ifdef PADDLE_WITH_CUDA
device_contexts_.emplace(
p,
PtrType(new CUDAPinnedDeviceContext(boost::get<CUDAPinnedPlace>(p))));
EmplaceDeviceContext<CUDAPinnedDeviceContext, CUDAPinnedPlace>(
&device_contexts_, p);
#else
PADDLE_THROW(
"'CUDAPlace' is not supported, Please re-compile with WITH_GPU "
......
paddle/fluid/platform/device_context.h
浏览文件 @ cf1944af
......@@ -10,6 +10,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <future> // NOLINT
#include <memory>
#include <mutex> // NOLINT
#include <string>
......@@ -223,9 +224,6 @@ class DeviceContextPool {
/*! \brief Return handle of single device context. */
platform::DeviceContext* Get(const platform::Place& place);
/*! \brief Return all the device contexts. */
const std::vector<const DeviceContext*> GetAllDeviceContexts() const;
template <typename Place>
const typename DefaultDeviceContextType<Place>::TYPE* GetByPlace(
const Place& place) {
......@@ -237,7 +235,8 @@ class DeviceContextPool {
private:
static DeviceContextPool* pool;
std::map<Place, std::unique_ptr<DeviceContext>> device_contexts_;
std::map<Place, std::shared_future<std::unique_ptr<DeviceContext>>>
device_contexts_;
DISABLE_COPY_AND_ASSIGN(DeviceContextPool);
};
......
paddle/scripts/paddle_build.sh
浏览文件 @ cf1944af
......@@ -147,13 +147,11 @@ function cmake_gen() {
-DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
-DCUDNN_ROOT=/usr/
-DWITH_TESTING=${WITH_TESTING:-ON}
-DWITH_FAST_BUNDLE_TEST=ON
-DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON
-DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF}
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON
-DWITH_CONTRIB=${WITH_CONTRIB:-ON}
-DWITH_INFERENCE=${WITH_INFERENCE:-ON}
-DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON}
-DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR}
-DWITH_ANAKIN=${WITH_ANAKIN:-OFF}
......@@ -181,12 +179,10 @@ EOF
-DWITH_PYTHON=${WITH_PYTHON:-ON} \
-DCUDNN_ROOT=/usr/ \
-DWITH_TESTING=${WITH_TESTING:-ON} \
-DWITH_FAST_BUNDLE_TEST=ON \
-DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \
-DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
-DWITH_CONTRIB=${WITH_CONTRIB:-ON} \
-DWITH_INFERENCE=${WITH_INFERENCE:-ON} \
-DWITH_INFERENCE_API_TEST=${WITH_INFERENCE_API_TEST:-ON} \
-DINFERENCE_DEMO_INSTALL_DIR=${INFERENCE_DEMO_INSTALL_DIR} \
-DWITH_ANAKIN=${WITH_ANAKIN:-OFF} \
......@@ -653,7 +649,7 @@ function gen_capi_package() {
function gen_fluid_lib() {
mkdir -p ${PADDLE_ROOT}/build
cd ${PADDLE_ROOT}/build
if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then
if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then
cat <<EOF
========================================
Generating fluid library for train and inference ...
......@@ -666,7 +662,7 @@ EOF
}
function tar_fluid_lib() {
if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then
if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then
cat <<EOF
========================================
Taring fluid library for train and inference ...
......@@ -681,7 +677,7 @@ EOF
}
function test_fluid_lib() {
if [[ ${WITH_C_API:-OFF} == "OFF" && ${WITH_INFERENCE:-ON} == "ON" ]] ; then
if [[ ${WITH_C_API:-OFF} == "OFF" ]] ; then
cat <<EOF
========================================
Testing fluid library for inference ...
......
python/paddle/fluid/layers/nn.py
浏览文件 @ cf1944af
......@@ -157,6 +157,8 @@ __all__ = [
'sequence_reverse',
'affine_channel',
'hash',
'log_loss',
'add_position_encoding',
]
......@@ -747,7 +749,7 @@ def dynamic_gru(input,
attr=helper.bias_attr, shape=[1, 3 * size], dtype=dtype, is_bias=True)
batch_size = input.shape[0]
inputs = {'Input': input, 'Weight': weight, 'Bias': bias}
if h_0 != None:
if h_0:
assert h_0.shape == (
batch_size, size
), 'The shape of h0 should be(batch_size, %d)' % size
......@@ -1823,7 +1825,7 @@ def conv3d(input,
return helper.append_activation(pre_act)
def sequence_pool(input, pool_type):
def sequence_pool(input, pool_type, is_test=False):
"""
This function add the operator for sequence pooling.
It pools features of all time-steps of each instance, and is applied
......@@ -1860,6 +1862,7 @@ def sequence_pool(input, pool_type):
input(variable): The input variable which is a LoDTensor.
pool_type (string): The pooling type of sequence_pool.
It supports average, sum, sqrt and max.
is_test(bool, Default False): Used distinguish training from scoring mode.
Returns:
The sequence pooling variable which is a Tensor.
......@@ -1887,7 +1890,8 @@ def sequence_pool(input, pool_type):
inputs={"X": input},
outputs={"Out": pool_out,
"MaxIndex": max_index},
attrs={"pooltype": pool_type.upper()})
attrs={"pooltype": pool_type.upper(),
"is_test": is_test})
# when pool_type is max, variable max_index is initialized,
# so we stop the gradient explicitly here
......@@ -3016,7 +3020,8 @@ def sequence_pad(x, pad_value, maxlen=None, name=None):
x = fluid.layers.data(name='y', shape=[10, 5],
dtype='float32', lod_level=1)
pad_value = fluid.layers.assign(input=numpy.array([0]))
pad_value = fluid.layers.assign(
input=numpy.array([0], dtype=numpy.float32))
out = fluid.layers.sequence_pad(x=x, pad_value=pad_value)
"""
......@@ -7580,3 +7585,99 @@ def hash(input, hash_size, num_hash=1, name=None):
attrs={'num_hash': num_hash,
'mod_by': hash_size})
return out
def log_loss(input, label, epsilon=1e-4, name=None):
"""
**Negative Log Loss Layer**
This layer accepts input predictions and target label and returns the
negative log loss.
.. math::
Out = -label * \\log{(input + \\epsilon)}
- (1 - label) * \\log{(1 - input + \\epsilon)}
Args:
input (Variable|list): a 2-D tensor with shape [N x 1], where N is the
batch size. This input is a probability computed
by the previous operator.
label (Variable|list): the ground truth which is a 2-D tensor with
shape [N x 1], where N is the batch size.
epsilon (float): epsilon
name (string): the name of log_loss
Returns:
Variable: A 2-D tensor with shape [N x 1], the negative log loss.
Examples:
.. code-block:: python
prob = fluid.layers.sigmoid(net)
cost = fluid.layers.log_loss(input=prob, label=label)
"""
helper = LayerHelper('log_loss', **locals())
if name is None:
loss = helper.create_variable_for_type_inference(dtype=input.dtype)
else:
loss = helper.create_variable(
name=name, dtype=input.dtype, persistable=False)
helper.append_op(
type='log_loss',
inputs={'Predicted': [input],
'Labels': [label]},
outputs={'Loss': [loss]},
attrs={'epsilon': epsilon})
return loss
def add_position_encoding(input, alpha, beta, name=None):
"""
**Add Position Encoding Layer**
This layer accepts an input 3D-Tensor of shape [N x M x P], and return an
output Tensor of shape [N x M x P] with positional encoding value.
Refer to `Attention Is All You Need<http://arxiv.org/pdf/1706.03762.pdf>`_ .
.. math::
PE(pos, 2i) = \\sin{(pos / 10000^{2i / P})} \\\\
PE(pos, 2i + 1) = \\cos{(pos / 10000^{2i / P})} \\\\
Out(:, pos, i) = \\alpha * input(:, pos, i) + \\beta * PE(pos, i)
Where:
* PE(pos, 2i): the increment for the number at even position
* PE(pos, 2i + 1): the increment for the number at odd position
Args:
input (Variable): 3-D input tensor with shape [N x M x P]
alpha (float): multiple of Input Tensor
beta (float): multiple of Positional Encoding Tensor
name (string): the name of position encoding layer
Returns:
Variable: A 3-D Tensor of shape [N x M x P] with positional encoding.
Examples:
.. code-block:: python
position_tensor = fluid.layers.add_position_encoding(input=tensor)
"""
helper = LayerHelper('add_position_encoding', **locals())
dtype = helper.input_dtype()
if name is None:
out = helper.create_variable_for_type_inference(dtype=dtype)
else:
out = helper.create_variable(name=name, dtype=dtype, persistable=False)
helper.append_op(
type="add_position_encoding",
inputs={"X": input},
outputs={"Out": out},
attrs={"alpha": alpha,
"beta": beta})
return out
python/paddle/fluid/metrics.py
浏览文件 @ cf1944af
......@@ -194,7 +194,7 @@ class CompositeMetric(MetricBase):
or soft-label, should custom the corresponding update rule.
"""
for m in self._metrics:
ans.append(m.update(preds, labels))
m.update(preds, labels)
def eval(self):
"""
......
python/paddle/fluid/tests/CMakeLists.txt
浏览文件 @ cf1944af
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests CACHE INTERNAL "python tests directory")
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
......
python/paddle/fluid/tests/book/high-level-api/image_classification/CMakeLists.txt
浏览文件 @ cf1944af
file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
# default test
foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
endforeach()
if(NOT APPLE)
# default test
foreach(src ${TEST_OPS})
py_test(${src} SRCS ${src}.py)
endforeach()
else()
foreach(src ${TEST_OPS})
if(${src} STREQUAL "test_image_classification_vgg")
message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
elseif(${src} STREQUAL "test_image_classification_resnet")
message(WARNING "These tests has been disabled in OSX for random fail: \n" ${src})
elseif()
py_test(${src} SRCS ${src}.py)
endif()
endforeach()
endif()
python/paddle/fluid/tests/unittests/CMakeLists.txt
浏览文件 @ cf1944af
......@@ -17,6 +17,10 @@ if(NOT WITH_DISTRIBUTE)
list(REMOVE_ITEM TEST_OPS test_listen_and_serv_op)
LIST(REMOVE_ITEM TEST_OPS test_dist_mnist)
LIST(REMOVE_ITEM TEST_OPS test_dist_word2vec)
LIST(REMOVE_ITEM TEST_OPS test_dist_ctr)
LIST(REMOVE_ITEM TEST_OPS test_dist_simnet_bow)
LIST(REMOVE_ITEM TEST_OPS test_dist_mnist_batch_merge)
LIST(REMOVE_ITEM TEST_OPS test_dist_text_classification)
endif(NOT WITH_DISTRIBUTE)
list(REMOVE_ITEM TEST_OPS test_seq_concat_op) # FIXME(helin): https://github.com/PaddlePaddle/Paddle/issues/8290
......@@ -55,6 +59,7 @@ function(py_test_modules TARGET_NAME)
if (py_test_modules_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
endif()
set_tests_properties(${TARGET_NAME} PROPERTIES TIMEOUT 600)
endif()
endfunction()
list(REMOVE_ITEM TEST_OPS test_warpctc_op)
......@@ -88,4 +93,6 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE
py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)
set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 150)
py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL)
py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
if(NOT APPLE)
py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
endif()
python/paddle/fluid/tests/unittests/dist_mnist.py
浏览文件 @ cf1944af
......@@ -90,8 +90,10 @@ class TestDistMnist2x2(TestDistRunnerBase):
inference_program = fluid.default_main_program().clone()
# Optimization
opt = fluid.optimizer.AdamOptimizer(
learning_rate=0.001, beta1=0.9, beta2=0.999)
# TODO(typhoonzero): fix distributed adam optimizer
# opt = fluid.optimizer.AdamOptimizer(
# learning_rate=0.001, beta1=0.9, beta2=0.999)
opt = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
# Reader
train_reader = paddle.batch(
......
python/paddle/fluid/tests/unittests/test_add_position_encoding_op.py 0 → 100644
浏览文件 @ cf1944af
# 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.
import unittest
import numpy as np
import math
import paddle.fluid.core as core
from op_test import OpTest
class TestAddPositionEncodingTensorOp(OpTest):
"""
This class is to test the AddPositionEncodingOp
"""
def setUp(self):
"""
the prepared section for add position encoding op
"""
self.op_type = "add_position_encoding"
self.dtype = np.float32
self.init_input_output()
self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(self.x), }
self.outputs = {'Out': self.out}
self.attrs = {'alpha': self.alpha, 'beta': self.beta}
def test_check_output(self):
"""
check the correctness of output
"""
self.check_output()
def test_check_grad(self):
"""
check the correctness of grad
"""
self.check_grad(['X'], 'Out', max_relative_error=0.005)
def init_input_output(self):
"""
init the input and output for test cases
"""
self.alpha = 0.6
self.beta = 0.5
self.x = np.random.uniform(0.1, 1, [2, 4, 4]).astype(self.dtype)
self.out = np.copy(self.x)
batch_size = self.x.shape[0]
max_length = self.x.shape[1]
enc_size = self.x.shape[2]
half_shape = int(enc_size / 2)
for i in range(batch_size):
for j in range(max_length):
for k in range(half_shape):
val = j / pow(10000.0, k / (
half_shape - 1)) if half_shape > 1 else j / 10000.0
self.out[i, j, k] = \
self.x[i, j, k] * self.alpha + math.sin(val) * self.beta
self.out[i, j, half_shape + k] = \
self.x[i, j, half_shape + k] * self.alpha + math.cos(val) * self.beta
class TestAddPositionEncodingLoDTensorOp(OpTest):
"""
This class is to test the AddPositionEncodingLoDTensorOp
"""
def setUp(self):
"""
the prepared section for add position encoding LoDTensor op
"""
self.op_type = "add_position_encoding"
self.dtype = np.float32
self.init_input_output()
self.inputs = {'X': (self.x, self.lod), }
self.outputs = {'Out': (self.out, self.lod)}
self.attrs = {'alpha': self.alpha, 'beta': self.beta}
def test_check_output(self):
"""
check the correctness of output
"""
self.check_output()
def test_check_grad(self):
"""
check the correctness of grad
"""
self.check_grad(['X'], 'Out', max_relative_error=0.005)
def init_input_output(self):
"""
init the input and output for test cases
"""
self.alpha = 0.6
self.beta = 0.5
self.x = np.random.uniform(0.1, 1, [10, 4]).astype(self.dtype)
self.lod = [[3, 7]]
self.out = np.copy(self.x)
batch_size = len(self.lod[0])
enc_size = self.x.shape[1]
start = 0
half_shape = int(enc_size / 2)
for i in range(batch_size):
max_length = self.lod[0][i]
for j in range(max_length):
for k in range(half_shape):
val = j / pow(10000.0, k / (
half_shape - 1)) if half_shape > 1 else j / 10000.0
pos = start + j
self.out[pos, k] = \
self.x[pos, k] * self.alpha + math.sin(val) * self.beta
self.out[pos, half_shape + k] = \
self.x[pos, half_shape + k] * self.alpha + math.cos(val) * self.beta
start += max_length
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_dist_base.py
浏览文件 @ cf1944af
......@@ -22,6 +22,8 @@ import signal
import subprocess
import six
import argparse
import pickle
import numpy as np
import paddle.fluid as fluid
......@@ -128,10 +130,15 @@ class TestDistRunnerBase(object):
else:
return origin_batch
out_losses = []
for _ in six.moves.xrange(RUN_STEP):
loss, = exe.run(fetch_list=[avg_cost.name],
feed=feeder.feed(get_data()))
print(loss)
out_losses.append(loss[0])
if six.PY2:
print(pickle.dumps(out_losses))
else:
sys.stdout.buffer.write(pickle.dumps(out_losses))
def runtime_main(test_class):
......@@ -149,7 +156,7 @@ def runtime_main(test_class):
parser.add_argument('--use_cuda', action='store_true')
parser.add_argument('--use_reduce', action='store_true')
parser.add_argument(
'--use_reader_alloc', action='store_true', required=False, default=True)
'--use_reader_alloc', action='store_true', required=False)
parser.add_argument('--batch_size', required=False, type=int, default=2)
parser.add_argument(
'--batch_merge_repeat', required=False, type=int, default=1)
......@@ -188,7 +195,7 @@ class TestDistBase(unittest.TestCase):
self._pservers = 2
self._ps_endpoints = "127.0.0.1:%s,127.0.0.1:%s" % (
self._find_free_port(), self._find_free_port())
self._python_interp = "python"
self._python_interp = sys.executable
self._sync_mode = True
self._enforce_place = None
self._mem_opt = False
......@@ -237,21 +244,6 @@ class TestDistBase(unittest.TestCase):
return ps0_proc, ps1_proc, ps0_pipe, ps1_pipe
def _wait_ps_ready(self, pid):
retry_times = 50
while True:
assert retry_times >= 0, "wait ps ready failed"
time.sleep(3)
try:
# the listen_and_serv_op would touch a file which contains the listen port
# on the /tmp directory until it was ready to process all the RPC call.
os.stat("/tmp/paddle.%d.port" % pid)
return
except os.error as e:
sys.stderr.write('waiting for pserver: %s, left retry %d\n' %
(e, retry_times))
retry_times -= 1
def _run_local(self,
model,
envs,
......@@ -288,23 +280,20 @@ class TestDistBase(unittest.TestCase):
env=envs)
local_out, local_err = local_proc.communicate()
local_ret = cpt.to_text(local_out)
if check_error_log:
err_log.close()
sys.stderr.write('local_stdout: %s\n' % local_ret)
sys.stderr.write('local_stdout: %s\n' % pickle.loads(local_out))
sys.stderr.write('local_stderr: %s\n' % local_err)
local_losses = local_ret.split("\n")
return local_losses
return pickle.loads(local_out)
def _run_cluster(self, model, envs, check_error_log):
# Run dist train to compare with local results
ps0, ps1, ps0_pipe, ps1_pipe = self.start_pserver(model,
check_error_log, envs)
self._wait_ps_ready(ps0.pid)
self._wait_ps_ready(ps1.pid)
ps0_ep, ps1_ep = self._ps_endpoints.split(",")
tr_cmd = "%s %s --role trainer --endpoints %s --trainer_id %d --current_endpoint %s --trainers %d --is_dist"
......@@ -339,8 +328,8 @@ class TestDistBase(unittest.TestCase):
env0.update(envs)
env1.update(envs)
print("tr0_cmd:{}, env0: {}".format(tr0_cmd, env0))
print("tr1_cmd:{}, env1: {}".format(tr1_cmd, env1))
print("tr0_cmd:{}".format(tr0_cmd))
print("tr1_cmd:{}".format(tr1_cmd))
tr0_pipe = open("/tmp/tr0_err.log", "wb")
tr1_pipe = open("/tmp/tr1_err.log", "wb")
......@@ -356,9 +345,7 @@ class TestDistBase(unittest.TestCase):
env=env1)
tr0_out, tr0_err = tr0_proc.communicate()
tr0_loss_text = cpt.to_text(tr0_out)
tr1_out, tr1_err = tr1_proc.communicate()
tr1_loss_text = cpt.to_text(tr1_out)
# close trainer file
tr0_pipe.close()
......@@ -373,15 +360,13 @@ class TestDistBase(unittest.TestCase):
ps1.terminate()
# print log
sys.stderr.write('trainer 0 stdout:\n %s\n' % tr0_loss_text)
sys.stderr.write('trainer 0 stderr:\n %s\n' % tr0_err)
sys.stderr.write('trainer 1 stdout: %s\n' % tr1_loss_text)
sys.stderr.write('trainer 0 stdout: %s\n' % pickle.loads(tr0_out))
sys.stderr.write('trainer 0 stderr: %s\n' % tr0_err)
sys.stderr.write('trainer 1 stdout: %s\n' % pickle.loads(tr1_out))
sys.stderr.write('trainer 1 stderr: %s\n' % tr1_err)
tr0_losses = tr0_loss_text.split("\n")
tr1_losses = tr1_loss_text.split("\n")
return tr0_losses, tr1_losses
# return tr0_losses, tr1_losses
return pickle.loads(tr0_out), pickle.loads(tr1_out)
def check_with_place(self,
model_file,
......@@ -411,9 +396,9 @@ class TestDistBase(unittest.TestCase):
check_error_log)
for step_id in range(RUN_STEP):
local_loss = eval(local_losses[step_id])[0]
tr0_loss = eval(tr0_losses[step_id])[0]
tr1_loss = eval(tr1_losses[step_id])[0]
dist_loss = (tr0_loss + tr1_loss) / 2
print(str(local_loss) + ":" + str(dist_loss))
self.assertAlmostEqual(local_loss, dist_loss, delta=delta)
local_loss = local_losses[step_id]
tr0_loss = tr0_losses[step_id]
tr1_loss = tr1_losses[step_id]
dist_loss = (np.array([tr0_loss]) + np.array([tr1_loss])) / 2
print("=======", local_loss, ":", dist_loss[0], "=======")
self.assertAlmostEqual(local_loss, dist_loss[0], delta=delta)
python/paddle/fluid/tests/unittests/test_dist_se_resnext.py
浏览文件 @ cf1944af
......@@ -23,16 +23,17 @@ class TestDistSeResneXt2x2(TestDistBase):
self._use_reader_alloc = False
def test_dist_train(self):
self.check_with_place("dist_se_resnext.py", delta=100)
self.check_with_place("dist_se_resnext.py", delta=1e-7)
class TestDistseResnXt2x2WithMemopt(TestDistBase):
def _setup_config(self):
self._sync_mode = True
self._mem_opt = True
self._use_reader_alloc = False
def test_dist_train(self):
self.check_with_place("dist_se_resnext.py", delta=100)
self.check_with_place("dist_se_resnext.py", delta=1e-7)
class TestDistSeResneXt2x2Async(TestDistBase):
......
python/paddle/fluid/tests/unittests/test_dist_transpiler.py
浏览文件 @ cf1944af
......@@ -283,6 +283,25 @@ class TestDecayedAdagrad(TranspilerTest):
trainer, _ = self.get_trainer()
class TestFtrl(TranspilerTest):
def net_conf(self):
x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
y_predict = fluid.layers.fc(input=x,
size=1000,
act=None,
param_attr=fluid.ParamAttr(name='fc_w'),
bias_attr=fluid.ParamAttr(name='fc_b'))
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(cost)
opt = fluid.optimizer.Ftrl(learning_rate=0.1)
opt.minimize(avg_cost)
def transpiler_test_impl(self):
pserver, startup = self.get_pserver(self.pserver1_ep)
trainer, _ = self.get_trainer()
class TestLRDecayConditional(TranspilerTest):
def net_conf(self):
x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
......@@ -405,18 +424,43 @@ class TestL2DecayWithPiecewise(TranspilerTest):
["sum", "scale", "scale", "elementwise_add", "momentum"])
class TestEmptyPserverOptimizeBlocks(TranspilerTest):
def net_conf(self):
x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
# only one parameter
y_predict = fluid.layers.fc(input=x,
size=1000,
act=None,
param_attr=fluid.ParamAttr(name='fc_w'),
bias_attr=False)
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(cost)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=1.0)
sgd_optimizer.minimize(avg_cost)
def transpiler_test_impl(self):
config = fluid.DistributeTranspilerConfig()
config.slice_var_up = False
pserver, startup = self.get_pserver(ep=self.pserver2_ep, config=config)
self.assertEqual(len(pserver.blocks), 2)
self.assertEqual(len(pserver.blocks[1].ops), 0)
class TestDistLookupTableBase(TranspilerTest):
def network_with_table(self, is_sparse, is_distributed):
self.table_size = 1000
self.emb_size = 64
self.lookup_table_name = 'shared_w'
def emb_pool(ids):
def emb_pool(ids, table_name, is_distributed):
emb = fluid.layers.embedding(
input=ids,
size=[self.table_size, self.emb_size],
dtype='float32',
param_attr=self.lookup_table_name, # share parameter
param_attr=table_name,
is_sparse=is_sparse,
is_distributed=is_distributed)
pool = fluid.layers.sequence_pool(input=emb, pool_type='average')
......@@ -426,9 +470,13 @@ class TestDistLookupTableBase(TranspilerTest):
name='title_ids', shape=[1], dtype='int64', lod_level=1)
brand_ids = fluid.layers.data(
name='brand_ids', shape=[1], dtype='int64', lod_level=1)
title_emb = emb_pool(title_ids)
brand_emb = emb_pool(brand_ids)
fc0 = fluid.layers.concat(input=[title_emb, brand_emb], axis=1)
profile_ids = fluid.layers.data(
name='brand_ids', shape=[1], dtype='int64', lod_level=1)
title_emb = emb_pool(title_ids, self.lookup_table_name, is_distributed)
brand_emb = emb_pool(brand_ids, self.lookup_table_name, is_distributed)
profile_emb = emb_pool(profile_ids, "profile_emb", False)
fc0 = fluid.layers.concat(
input=[title_emb, brand_emb, profile_emb], axis=1)
predict = fluid.layers.fc(input=fc0,
size=2,
act=None,
......@@ -449,7 +497,7 @@ class TestLocalLookupTable(TestDistLookupTableBase):
def transpiler_test_impl(self):
pserver1, startup1 = self.get_pserver(self.pserver1_ep)
self.assertEqual(len(pserver1.blocks), 3)
self.assertEqual(len(pserver1.blocks), 4)
# 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops],
......@@ -459,16 +507,23 @@ class TestLocalLookupTable(TestDistLookupTableBase):
self.assertEqual([op.type for op in pserver1.blocks[2].ops],
["sum", "scale", "adam", "scale", "scale"])
# 3 optimize for table 2 adam
# NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
self.assertEqual([op.type for op in pserver1.blocks[3].ops],
["sum", "scale", "adam", "scale", "scale"])
trainer, _ = self.get_trainer()
self.assertEqual(len(trainer.blocks), 1)
ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean',
'fill_constant', 'mean_grad', 'cross_entropy_grad',
'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
'lookup_table_grad', 'sum', 'split_selected_rows', 'send',
'send_barrier', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat'
'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'split_selected_rows', 'send', 'sequence_pool_grad',
'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
'sum', 'split_selected_rows', 'send', 'send_barrier', 'recv',
'recv', 'recv', 'recv', 'fetch_barrier', 'concat', 'concat'
]
self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
......@@ -480,39 +535,45 @@ class TestDistLookupTable(TestDistLookupTableBase):
def transpiler_test_impl(self):
pserver1, startup1 = self.get_pserver(self.pserver1_ep)
self.assertEqual(len(pserver1.blocks), 5)
self.assertEqual(len(pserver1.blocks), 6)
# 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops],
["sum", "scale", "adam", "scale", "scale"])
# 2 optimize for table sgd
# 4 prefetch -> lookup_sparse_table for data0
self.assertEqual([op.type for op in pserver1.blocks[2].ops],
["sum", "scale", "adam", "scale", "scale"])
# 2 optimize for table sgd
self.assertEqual([op.type for op in pserver1.blocks[3].ops],
["sum", "sgd"])
# 3 prefetch -> lookup_sparse_table for data0
self.assertEqual([op.type for op in pserver1.blocks[3].ops],
self.assertEqual([op.type for op in pserver1.blocks[4].ops],
["lookup_sparse_table"])
# 4 save table
self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["save"])
# 5 save table
self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
trainer, trainer_startup = self.get_trainer()
self.assertEqual(len(trainer.blocks), 1)
ops = [
'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sum', 'split_ids',
'send', 'send_barrier', 'recv', 'recv', 'fetch_barrier'
'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
'lookup_table_grad', 'split_selected_rows', 'send',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
'lookup_table_grad', 'sum', 'split_ids', 'send', 'send_barrier',
'recv', 'recv', 'recv', 'fetch_barrier', 'concat'
]
self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
startup_ops = [
'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
'fill_constant', 'fill_constant', 'uniform_random', 'recv', 'recv',
'fetch_barrier', 'fake_init'
'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
'fill_constant', 'fill_constant', 'uniform_random',
'uniform_random', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat',
'fake_init'
]
self.assertEqual([op.type for op in trainer_startup.blocks[0].ops],
startup_ops)
......@@ -526,7 +587,7 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
config = fluid.DistributeTranspilerConfig()
pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False)
self.assertEqual(len(pserver1.blocks), 3)
self.assertEqual(len(pserver1.blocks), 4)
# 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops],
......@@ -535,17 +596,23 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
# NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
self.assertEqual([op.type for op in pserver1.blocks[2].ops],
["adam", "scale", "scale"])
# 3 optimize for table adam
# NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num
self.assertEqual([op.type for op in pserver1.blocks[3].ops],
["adam", "scale", "scale"])
trainer, _ = self.get_trainer(config)
self.assertEqual(len(trainer.blocks), 1)
ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean',
'fill_constant', 'mean_grad', 'cross_entropy_grad',
'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
'lookup_table_grad', 'sum', 'split_selected_rows', 'send', 'recv',
'recv', 'recv', 'concat'
'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'split_selected_rows', 'send', 'sequence_pool_grad',
'lookup_table_grad', 'sequence_pool_grad', 'lookup_table_grad',
'sum', 'split_selected_rows', 'send', 'recv', 'recv', 'recv',
'recv', 'concat', 'concat'
]
self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
......@@ -559,29 +626,34 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase):
pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False)
self.assertEqual(len(pserver1.blocks), 5)
self.assertEqual(len(pserver1.blocks), 6)
# 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops],
["adam", "scale", "scale"])
# 2 optimize for table sgd
self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["sgd"])
# 3 prefetch -> lookup_sparse_table for data0
self.assertEqual([op.type for op in pserver1.blocks[3].ops],
# 2 optimize for table adam
self.assertEqual([op.type for op in pserver1.blocks[2].ops],
["adam", "scale", "scale"])
# 3 optimize for table sgd
self.assertEqual([op.type for op in pserver1.blocks[3].ops], ["sgd"])
# 4 prefetch -> lookup_sparse_table for data0
self.assertEqual([op.type for op in pserver1.blocks[4].ops],
["lookup_sparse_table"])
# 4 save table
self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["save"])
# 5 save table
self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
trainer, _ = self.get_trainer(config)
self.assertEqual(len(trainer.blocks), 1)
ops = [
'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
'sequence_pool', 'concat', 'mul', 'elementwise_add',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sum', 'split_ids',
'send', 'recv', 'recv'
'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
'lookup_table_grad', 'split_selected_rows', 'send',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad',
'lookup_table_grad', 'sum', 'split_ids', 'send', 'recv', 'recv',
'recv', 'concat'
]
self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
......
python/paddle/fluid/tests/unittests/test_listen_and_serv_op.py
浏览文件 @ cf1944af
......@@ -55,6 +55,46 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id):
exe.run(pserver_prog)
def run_pserver_with_empty_block(use_cuda, sync_mode, ip, port, trainers,
trainer_id):
x = fluid.layers.data(name='x', shape=[1], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None, bias_attr=False)
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
# loss function
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(cost)
# optimizer
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
exe = fluid.Executor(place)
ps1 = ip + ":" + str(int(port) + 1)
ps2 = ip + ":" + port
pserver_endpoints = ps1 + "," + ps2
config = fluid.DistributeTranspilerConfig()
config.slice_var_up = False
t = fluid.DistributeTranspiler(config=config)
t.transpile(
trainer_id,
pservers=pserver_endpoints,
trainers=trainers,
sync_mode=sync_mode)
pserver_prog = t.get_pserver_program(ps2)
# pserver2 have no parameter
assert (len(pserver_prog.blocks) == 2)
assert (len(pserver_prog.blocks[1].ops) == 0)
pserver_startup = t.get_startup_program(ps2, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
class TestListenAndServOp(OpTest):
def setUp(self):
self.ps_timeout = 5
......@@ -63,9 +103,9 @@ class TestListenAndServOp(OpTest):
self.trainers = 1
self.trainer_id = 0
def _start_pserver(self, use_cuda, sync_mode):
def _start_pserver(self, use_cuda, sync_mode, pserver_func):
p = Process(
target=run_pserver,
target=pserver_func,
args=(use_cuda, sync_mode, self.ip, self.port, self.trainers,
self.trainer_id))
p.daemon = True
......@@ -92,7 +132,24 @@ class TestListenAndServOp(OpTest):
def test_handle_signal_in_serv_op(self):
# run pserver on CPU in sync mode
p1 = self._start_pserver(False, True)
p1 = self._start_pserver(False, True, run_pserver)
self._wait_ps_ready(p1.pid)
# raise SIGTERM to pserver
os.kill(p1.pid, signal.SIGINT)
p1.join()
# run pserver on CPU in async mode
p2 = self._start_pserver(False, False, run_pserver)
self._wait_ps_ready(p2.pid)
# raise SIGTERM to pserver
os.kill(p2.pid, signal.SIGTERM)
p2.join()
def test_list_and_serv_run_empty_optimize_block(self):
# run pserver on CPU in sync mode
p1 = self._start_pserver(False, True, run_pserver_with_empty_block)
self._wait_ps_ready(p1.pid)
# raise SIGTERM to pserver
......@@ -100,7 +157,7 @@ class TestListenAndServOp(OpTest):
p1.join()
# run pserver on CPU in async mode
p2 = self._start_pserver(False, False)
p2 = self._start_pserver(False, False, run_pserver_with_empty_block)
self._wait_ps_ready(p2.pid)
# raise SIGTERM to pserver
......
python/paddle/fluid/tests/unittests/test_seq_pool.py
浏览文件 @ cf1944af
......@@ -184,6 +184,20 @@ class TestSeqMaxPool2D(TestSeqAvgPool2D):
out[i] = np.reshape(np.amax(sub_x, axis=0), (3, 11))
class TestSeqMaxPool2DInference(TestSeqMaxPool2D):
def compute(self, x, offset, out):
self.attrs = {'pooltype': "MAX", 'is_test': True}
for i in range(len(offset[0]) - 1):
sub_x = np.reshape(x[offset[0][i]:offset[0][i + 1], :],
(-1, 3 * 11))
out[i] = np.reshape(np.amax(sub_x, axis=0), (3, 11))
def test_check_grad(self):
"""Grad computation does not apply to Sequence MAX
Pool executed when is_test is true """
return
class TestSeqLastPool2D(TestSeqAvgPool2D):
def compute(self, x, offset, out):
self.attrs = {'pooltype': "LAST"}
......
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ cf1944af
......@@ -35,6 +35,7 @@ import sys
import numpy as np
import collections
import six
import logging
from .ps_dispatcher import RoundRobin, HashName, PSDispatcher
from .. import core, framework
......@@ -767,6 +768,15 @@ in a single call.")
prefetch_var_name_to_block_id.extend(
lookup_table_var_name_to_block_id)
if len(optimize_blocks) == 0:
logging.warn("pserver [" + str(endpoint) +
"] has no optimize block!!")
pre_block_idx = pserver_program.num_blocks - 1
empty_block = pserver_program._create_block(pre_block_idx)
optimize_blocks.append(empty_block)
# In some case, some parameter server will have no parameter to optimize
# So we give an empty optimize block to parameter server.
attrs = {
"optimize_blocks": optimize_blocks,
"endpoint": endpoint,
......@@ -1065,7 +1075,12 @@ to transpile() call.")
continue_search_lookup_table_op = False
all_ops = program.global_block().ops
for op in all_ops:
if op.type == LOOKUP_TABLE_TYPE:
if op.type == LOOKUP_TABLE_TYPE and self.table_name == op.input(
"W")[0]:
if not op.attr('is_distributed'):
raise RuntimeError(
"lookup_table_op that lookup an distributed embedding table"
"should set is_distributed to true")
continue_search_lookup_table_op = True
lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list(
......@@ -1275,7 +1290,6 @@ to transpile() call.")
}
outputs = {"ParamOut": [param_var]}
# only support sgd now
import logging
logging.warn(
"distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of "
+ table_opt_op.type)
......@@ -1442,6 +1456,9 @@ to transpile() call.")
elif op_type == "decayed_adagrad":
if varkey == "Moment":
return param_shape
elif op_type == "ftrl":
if varkey in ["SquaredAccumulator", "LinearAccumulator"]:
return param_shape
elif op_type == "sgd":
pass
else:
......
python/paddle/fluid/transpiler/inference_transpiler.py
浏览文件 @ cf1944af
......@@ -61,6 +61,9 @@ class InferenceTranspiler(object):
raise TypeError("scope should be as Scope type or None")
use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False))
if use_mkldnn:
self._depthwise_conv_mkldnn(program)
self._fuse_batch_norm(program, place, scope)
if use_mkldnn:
self._fuse_conv_bias_mkldnn(program)
......@@ -70,6 +73,31 @@ class InferenceTranspiler(object):
program) # ResNet residual block merging
self._fuse_bn_relu_mkldnn(program)
def _depthwise_conv_mkldnn(self, program):
'''
Transpile the program by replacing depthwise_conv2d to conv2d for MKLDNN program.
The result is:
- before:
- any_other_op->depthwise_conv->any_other_op
- after:
- any_other_op->conv->any_other_op
:param program: program to transpile
:type program: Program
'''
self.block = program.block(0)
i = 0
while i < len(self.block.ops):
current_op = self.block.ops[i]
if current_op.type == 'depthwise_conv2d':
current_op.desc.set_type("conv2d")
i = i + 1
# TODO(luotao): use clone() method to flush the program.desc in force,
# since some large program.desc will not be flushed immediately.
# And a better solution will be considered later.
program = program.clone()
def _fuse_conv_eltwise_mkldnn(self, program):
'''
Transpile the program fusing elementwise_add into conv for MKLDNN
......
Markdown is supported
0% 或 .
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 或