Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 2 年 前同步成功

通知 2325
Star 20933
Fork 5424
P
Paddle
“5e153bf16e25f3ab5039531be61c3d7ee09137bb”上不存在“test/ir/inference/test_trt_convert_shape.py”
提交 d31e1940 编写于 作者: W wangguibao
浏览文件
操作

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

上级 48031c41 2eaa291e
显示空白变更内容
内联 并排
Showing with 798 addition and 389 deletion
CMakeLists.txt
浏览文件 @ d31e1940
...@@ -62,7 +62,6 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF) ...@@ -62,7 +62,6 @@ option(WITH_DISTRIBUTE "Compile with distributed support" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
option(EIGEN_USE_THREADS "Compile with multi-threaded 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_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(WITH_CONTRIB "Compile the third-party contributation" OFF)
option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better debug." 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_ANAKIN "Compile with Anakin library" OFF)
......
README.md
浏览文件 @ d31e1940
...@@ -2,8 +2,8 @@ ...@@ -2,8 +2,8 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle) [![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/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/1.0/beginners_guide/index.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) [![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) [![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. ...@@ -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. 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: ### Install Latest Stable Release:
``` ```
# Linux CPU # Linux CPU
...@@ -27,9 +27,9 @@ pip install paddlepaddle ...@@ -27,9 +27,9 @@ pip install paddlepaddle
# Linux GPU cuda9cudnn7 # Linux GPU cuda9cudnn7
pip install paddlepaddle-gpu pip install paddlepaddle-gpu
# Linux GPU cuda8cudnn7 # Linux GPU cuda8cudnn7
pip install paddlepaddle-gpu==1.0.1.post87 pip install paddlepaddle-gpu==1.1.0.post87
# Linux GPU cuda8cudnn5 # 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/ # For installation on other platform, refer to http://paddlepaddle.org/
``` ```
...@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.0.1.post85 ...@@ -76,26 +76,26 @@ pip install paddlepaddle-gpu==1.0.1.post85
## Installation ## 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 ## Documentation
We provide [English](http://paddlepaddle.org/documentation/docs/en/1.0.0/getstarted/index_en.html) and We provide [English](http://paddlepaddle.org/documentation/docs/en/1.1/getstarted/index_en.html) and
[Chinese](http://paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/index.html) documentation. [Chinese](http://paddlepaddle.org/documentation/docs/zh/1.1/beginners_guide/index.html) documentation.
- [Deep Learning 101](https://github.com/PaddlePaddle/book) - [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. 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. 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. 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! We appreciate your contributions!
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ d31e1940
...@@ -56,6 +56,7 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu ...@@ -56,6 +56,7 @@ cc_library(scope_buffered_ssa_graph_executor SRCS scope_buffered_ssa_graph_execu
# device_context reduce_op_handle ) # device_context reduce_op_handle )
cc_library(fast_threaded_ssa_graph_executor SRCS fast_threaded_ssa_graph_executor.cc 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) 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 cc_library(build_strategy SRCS build_strategy.cc DEPS
graph_viz_pass multi_devices_graph_pass graph_viz_pass multi_devices_graph_pass
......
paddle/fluid/framework/details/broadcast_op_handle_test.cc
浏览文件 @ d31e1940
...@@ -12,232 +12,12 @@ ...@@ -12,232 +12,12 @@
// See the License for the specific language governing permissions and // See the License for the specific language governing permissions and
// limitations under the License. // limitations under the License.
#include "paddle/fluid/framework/details/broadcast_op_handle.h" #include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
#include "gtest/gtest.h"
#include "paddle/fluid/platform/device_context.h"
namespace paddle { namespace paddle {
namespace framework { namespace framework {
namespace details { 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) { TEST(BroadcastTester, TestCPUBroadcastTestLodTensor) {
TestBroadcastOpHandle test_op; TestBroadcastOpHandle test_op;
size_t input_scope_idx = 0; size_t input_scope_idx = 0;
......
paddle/fluid/framework/details/broadcast_op_handle_test.h 0 → 100644
浏览文件 @ d31e1940
// 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/fused_broadcast_op_handle_test.cc 0 → 100644
浏览文件 @ d31e1940
// 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/ir/graph_pattern_detector.cc
浏览文件 @ d31e1940
...@@ -259,6 +259,15 @@ GraphPatternDetector::DetectPatterns() { ...@@ -259,6 +259,15 @@ GraphPatternDetector::DetectPatterns() {
return result; 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 // TODO(Superjomn) enhance the function as it marks unique unique as duplicates
// see https://github.com/PaddlePaddle/Paddle/issues/13550 // see https://github.com/PaddlePaddle/Paddle/issues/13550
void GraphPatternDetector::UniquePatterns( void GraphPatternDetector::UniquePatterns(
...@@ -267,12 +276,16 @@ void GraphPatternDetector::UniquePatterns( ...@@ -267,12 +276,16 @@ void GraphPatternDetector::UniquePatterns(
std::vector<GraphPatternDetector::subgraph_t> result; std::vector<GraphPatternDetector::subgraph_t> result;
std::unordered_set<size_t> set; std::unordered_set<size_t> set;
std::hash<std::string> hasher;
for (auto &g : *subgraphs) { for (auto &g : *subgraphs) {
size_t key = 0; // Sort the items in the sub-graph, and transform to a string key.
for (auto &item : g) { std::vector<std::pair<PDNode *, Node *>> sorted_keys(g.begin(), g.end());
key ^= std::hash<void *>{}(item.first); std::sort(sorted_keys.begin(), sorted_keys.end(), GraphItemCMP);
key ^= std::hash<void *>{}(item.second); std::stringstream ss;
} for (auto &item : sorted_keys) {
ss << item.first << ":" << item.second;
}
auto key = hasher(ss.str());
if (!set.count(key)) { if (!set.count(key)) {
result.emplace_back(g); result.emplace_back(g);
set.insert(key); set.insert(key);
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ d31e1940
...@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor( ...@@ -418,7 +418,7 @@ void LoDTensor::MergeLoDTensor(
PADDLE_ENFORCE_EQ(new_lod.size(), lod.size()); PADDLE_ENFORCE_EQ(new_lod.size(), lod.size());
for (size_t j = 0; j < lod.size(); ++j) { for (size_t j = 0; j < lod.size(); ++j) {
auto &sub_lod = new_lod[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) { for (size_t k = 1; k < lod[j].size(); ++k) {
sub_lod.push_back(lod[j][k] + offset); sub_lod.push_back(lod[j][k] + offset);
} }
......
paddle/fluid/framework/operator.cc
浏览文件 @ d31e1940
...@@ -354,18 +354,18 @@ void OperatorBase::GenerateTemporaryNames() { ...@@ -354,18 +354,18 @@ void OperatorBase::GenerateTemporaryNames() {
} }
} }
static bool VarIsTensor(const Variable* var) { static bool VarIsTensor(const Variable& var) {
return var->IsType<LoDTensor>() || var->IsType<SelectedRows>(); return var.IsType<LoDTensor>() || var.IsType<SelectedRows>();
} }
const Tensor* GetTensorFromVar(Variable* var) { const Tensor* GetTensorFromVar(const Variable& var) {
if (var->IsType<LoDTensor>()) { if (var.IsType<LoDTensor>()) {
return var->GetMutable<LoDTensor>(); return static_cast<const Tensor*>(&(var.Get<LoDTensor>()));
} else if (var->IsType<SelectedRows>()) { } else if (var.IsType<SelectedRows>()) {
return var->GetMutable<SelectedRows>()->mutable_value(); return &(var.Get<SelectedRows>().value());
} else { } else {
PADDLE_THROW("Variable type_id %s, expect LoDTensor/SelectedRows.", 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 { ...@@ -415,8 +415,7 @@ bool ExecutionContext::HasOutput(const std::string& name) const {
template <> template <>
const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const { const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
auto* var = InputVar(name); auto* var = InputVar(name);
return var == nullptr ? nullptr return var == nullptr ? nullptr : GetTensorFromVar(*var);
: GetTensorFromVar(const_cast<Variable*>(var));
} }
template <> template <>
...@@ -428,7 +427,7 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>( ...@@ -428,7 +427,7 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
std::transform(names.begin(), names.end(), std::back_inserter(res), std::transform(names.begin(), names.end(), std::back_inserter(res),
[&](const std::string& sub_name) { [&](const std::string& sub_name) {
auto var = scope_.FindVar(sub_name); auto var = scope_.FindVar(sub_name);
return var == nullptr ? nullptr : GetTensorFromVar(var); return var == nullptr ? nullptr : GetTensorFromVar(*var);
}); });
return res; return res;
} }
...@@ -770,8 +769,10 @@ void OperatorWithKernel::TransferInplaceVarsBack( ...@@ -770,8 +769,10 @@ void OperatorWithKernel::TransferInplaceVarsBack(
for (auto& var_name : inplace_vars) { for (auto& var_name : inplace_vars) {
VLOG(3) << "share inplace var " + var_name + " back to it's original scope"; VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name)); auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
auto* transformed_tensor = auto* var = transfer_scope.FindVar(var_name);
GetTensorFromVar(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); original_tensor->ShareDataWith(*transformed_tensor);
} }
} }
...@@ -784,11 +785,11 @@ Scope* OperatorWithKernel::TryTransferData( ...@@ -784,11 +785,11 @@ Scope* OperatorWithKernel::TryTransferData(
for (auto& var_name : var_name_item.second) { for (auto& var_name : var_name_item.second) {
auto* var = scope.FindVar(var_name); auto* var = scope.FindVar(var_name);
// Only tensor can be tranfer to another device. // Only tensor can be tranfer to another device.
if (var == nullptr || !VarIsTensor(var)) { if (var == nullptr || !VarIsTensor(*var)) {
continue; continue;
} }
auto* tensor_in = GetTensorFromVar(var); auto* tensor_in = GetTensorFromVar(*var);
if (!tensor_in->IsInitialized()) { if (!tensor_in->IsInitialized()) {
continue; continue;
} }
......
paddle/fluid/framework/operator.h
浏览文件 @ d31e1940
...@@ -63,7 +63,7 @@ inline std::string GradVarName(const std::string& var_name) { ...@@ -63,7 +63,7 @@ inline std::string GradVarName(const std::string& var_name) {
} }
proto::VarType::Type GetDataTypeOfVar(const Variable* var); proto::VarType::Type GetDataTypeOfVar(const Variable* var);
const Tensor* GetTensorFromVar(Variable* var); const Tensor* GetTensorFromVar(const Variable& var);
class OperatorBase; class OperatorBase;
class ExecutionContext; class ExecutionContext;
......
paddle/fluid/framework/tensor_test.cc
浏览文件 @ d31e1940
...@@ -75,6 +75,19 @@ TEST(Tensor, MutableData) { ...@@ -75,6 +75,19 @@ TEST(Tensor, MutableData) {
platform::CPUPlace()); platform::CPUPlace());
EXPECT_EQ(p1, p2); 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 #ifdef PADDLE_WITH_CUDA
{ {
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ d31e1940
cc_library(inference_io SRCS io.cc) if(WITH_TESTING)
include(test.cmake) # some generic cmake funtion for inference
endif()
# analysis and tensorrt must be added before creating static library, # analysis and tensorrt must be added before creating static library,
# otherwise, there would be undefined reference to them in static library. # otherwise, there would be undefined reference to them in static library.
add_subdirectory(analysis) add_subdirectory(analysis)
...@@ -8,6 +10,8 @@ endif() ...@@ -8,6 +10,8 @@ endif()
set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor) set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor)
cc_library(inference_io SRCS io.cc)
# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal? # TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
cc_library(paddle_fluid_api cc_library(paddle_fluid_api
SRCS io.cc SRCS io.cc
......
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @ d31e1940
...@@ -20,22 +20,17 @@ cc_test(test_node SRCS node_tester.cc DEPS analysis) ...@@ -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_test(test_dot SRCS dot_tester.cc DEPS analysis)
cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid) cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis paddle_fluid)
function (inference_analysis_test TARGET) function(inference_analysis_test TARGET)
if(WITH_TESTING) if(WITH_TESTING)
set(options "") set(options "")
set(oneValueArgs "") set(oneValueArgs "")
set(multiValueArgs SRCS ARGS EXTRA_DEPS) set(multiValueArgs SRCS ARGS EXTRA_DEPS)
cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN}) cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(mem_opt "") inference_base_test(${TARGET}
if(WITH_GPU) SRCS ${analysis_test_SRCS}
set(mem_opt "--fraction_of_gpu_memory_to_use=0.5")
endif()
cc_test(${TARGET}
SRCS "${analysis_test_SRCS}"
DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS} DEPS analysis pass ${GLOB_PASS_LIB} ${analysis_test_EXTRA_DEPS}
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt} ${analysis_test_ARGS}) ARGS --inference_model_dir=${WORD2VEC_MODEL_DIR} ${analysis_test_ARGS})
set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec) endif()
endif(WITH_TESTING)
endfunction(inference_analysis_test) endfunction(inference_analysis_test)
inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api) inference_analysis_test(test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api)
......
paddle/fluid/inference/api/CMakeLists.txt
浏览文件 @ d31e1940
...@@ -17,39 +17,12 @@ if(APPLE) ...@@ -17,39 +17,12 @@ if(APPLE)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
endif(APPLE) 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) if(WITH_GPU AND TENSORRT_FOUND)
set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor) set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine analysis_predictor)
endif() endif()
function(inference_api_test TARGET_NAME)
if (WITH_TESTING)
set(options "")
set(oneValueArgs SRC)
set(multiValueArgs ARGS)
cmake_parse_arguments(inference_test "${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(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(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) cc_library(analysis_predictor SRCS analysis_predictor.cc DEPS paddle_inference_api analysis naive_executor zero_copy_tensor)
...@@ -59,8 +32,11 @@ cc_test(test_paddle_inference_api ...@@ -59,8 +32,11 @@ cc_test(test_paddle_inference_api
SRCS api_tester.cc SRCS api_tester.cc
DEPS paddle_inference_api) DEPS paddle_inference_api)
inference_api_test(test_api_impl SRC api_impl_tester.cc if(WITH_TESTING)
ARGS test_word2vec test_image_classification) 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 cc_test(test_analysis_predictor SRCS analysis_predictor_tester.cc DEPS analysis_predictor ${inference_deps} paddle_inference_api
ARGS --dirname=${PYTHON_TESTS_DIR}/book) ARGS --dirname=${PYTHON_TESTS_DIR}/book)
...@@ -68,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND) ...@@ -68,8 +44,10 @@ if(WITH_GPU AND TENSORRT_FOUND)
cc_library(paddle_inference_tensorrt_subgraph_engine cc_library(paddle_inference_tensorrt_subgraph_engine
SRCS api_tensorrt_subgraph_engine.cc 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) DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api tensorrt_converter zero_copy_tensor_dummy)
if(WITH_TESTING)
inference_api_test(test_api_tensorrt_subgraph_engine SRC api_tensorrt_subgraph_engine_tester.cc ARGS test_word2vec) 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() endif()
if (WITH_ANAKIN AND WITH_MKL) # only needed in CI if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
......
paddle/fluid/inference/api/api_impl_tester.cc
浏览文件 @ d31e1940
...@@ -22,12 +22,14 @@ limitations under the License. */ ...@@ -22,12 +22,14 @@ limitations under the License. */
#include "paddle/fluid/inference/tests/test_helper.h" #include "paddle/fluid/inference/tests/test_helper.h"
#ifdef __clang__ #ifdef __clang__
#define ACC_DIFF 4e-2 #define ACC_DIFF 4e-3
#else #else
#define ACC_DIFF 1e-2 #define ACC_DIFF 1e-3
#endif #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 { namespace paddle {
...@@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) { ...@@ -49,7 +51,7 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
NativeConfig GetConfig() { NativeConfig GetConfig() {
NativeConfig config; NativeConfig config;
config.model_dir = FLAGS_dirname + "/word2vec.inference.model"; config.model_dir = FLAGS_word2vec_dirname;
LOG(INFO) << "dirname " << config.model_dir; LOG(INFO) << "dirname " << config.model_dir;
config.fraction_of_gpu_memory = 0.15; config.fraction_of_gpu_memory = 0.15;
#ifdef PADDLE_WITH_CUDA #ifdef PADDLE_WITH_CUDA
...@@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) { ...@@ -116,7 +118,7 @@ void MainImageClassification(bool use_gpu) {
NativeConfig config = GetConfig(); NativeConfig config = GetConfig();
config.use_gpu = use_gpu; config.use_gpu = use_gpu;
config.model_dir = config.model_dir =
FLAGS_dirname + "/image_classification_resnet.inference.model"; FLAGS_book_dirname + "/image_classification_resnet.inference.model";
const bool is_combined = false; const bool is_combined = false;
std::vector<std::vector<int64_t>> feed_target_shapes = std::vector<std::vector<int64_t>> feed_target_shapes =
...@@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) { ...@@ -220,7 +222,7 @@ void MainThreadsImageClassification(bool use_gpu) {
NativeConfig config = GetConfig(); NativeConfig config = GetConfig();
config.use_gpu = use_gpu; config.use_gpu = use_gpu;
config.model_dir = config.model_dir =
FLAGS_dirname + "/image_classification_resnet.inference.model"; FLAGS_book_dirname + "/image_classification_resnet.inference.model";
auto main_predictor = CreatePaddlePredictor<NativeConfig>(config); auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
std::vector<framework::LoDTensor> jobs(num_jobs); std::vector<framework::LoDTensor> jobs(num_jobs);
......
paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc
浏览文件 @ d31e1940
...@@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) { ...@@ -29,13 +29,13 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
//# 1. Create PaddlePredictor with a config. //# 1. Create PaddlePredictor with a config.
NativeConfig config0; NativeConfig config0;
config0.model_dir = FLAGS_dirname + "word2vec.inference.model"; config0.model_dir = FLAGS_dirname;
config0.use_gpu = true; config0.use_gpu = true;
config0.fraction_of_gpu_memory = 0.3; config0.fraction_of_gpu_memory = 0.3;
config0.device = 0; config0.device = 0;
MixedRTConfig config1; MixedRTConfig config1;
config1.model_dir = FLAGS_dirname + "word2vec.inference.model"; config1.model_dir = FLAGS_dirname;
config1.use_gpu = true; config1.use_gpu = true;
config1.fraction_of_gpu_memory = 0.3; config1.fraction_of_gpu_memory = 0.3;
config1.device = 0; config1.device = 0;
......
paddle/fluid/inference/api/demo_ci/run.sh
浏览文件 @ d31e1940
...@@ -62,7 +62,7 @@ for WITH_STATIC_LIB in ON OFF; do ...@@ -62,7 +62,7 @@ for WITH_STATIC_LIB in ON OFF; do
-DWITH_GPU=$TEST_GPU_CPU \ -DWITH_GPU=$TEST_GPU_CPU \
-DWITH_STATIC_LIB=$WITH_STATIC_LIB -DWITH_STATIC_LIB=$WITH_STATIC_LIB
make -j 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 if [ -d $word2vec_model ]; then
for use_gpu in $use_gpu_list; do for use_gpu in $use_gpu_list; do
./simple_on_word2vec \ ./simple_on_word2vec \
......
paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc
浏览文件 @ d31e1940
...@@ -70,12 +70,8 @@ void Main(bool use_gpu) { ...@@ -70,12 +70,8 @@ void Main(bool use_gpu) {
// The outputs' buffers are in CPU memory. // The outputs' buffers are in CPU memory.
for (size_t i = 0; i < std::min(static_cast<size_t>(5), num_elements); for (size_t i = 0; i < std::min(static_cast<size_t>(5), num_elements);
i++) { i++) {
// Here will result random fail, for that the model is trained by CI, the CHECK_NEAR(static_cast<float*>(outputs.front().data.data())[i], result[i],
// train phase is not stable, so the result will be random. 0.001);
// TODO(Superjomn) will restore after the model is upload.
// CHECK_NEAR(static_cast<float*>(outputs.front().data.data())[i],
// result[i],
// 0.001);
} }
} }
} }
......
paddle/fluid/inference/test.cmake 0 → 100644
浏览文件 @ d31e1940
set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com" CACHE STRING "inference download url")
set(INFERENCE_DEMO_INSTALL_DIR "${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} ${base_test_ARGS})
endfunction()
paddle/fluid/inference/tests/api/CMakeLists.txt
浏览文件 @ d31e1940
set(INFERENCE_URL "http://paddle-inference-dist.cdn.bcebos.com")
set(INFERENCE_DEMO_INSTALL_DIR "${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) 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) function(download_model_and_data install_dir model_name data_name)
if (NOT EXISTS ${install_dir}) if (NOT EXISTS ${install_dir})
......
paddle/fluid/operators/sum_op.cc
浏览文件 @ d31e1940
...@@ -67,6 +67,7 @@ class SumOp : public framework::OperatorWithKernel { ...@@ -67,6 +67,7 @@ class SumOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType( framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override { const framework::ExecutionContext& ctx) const override {
auto x_vars = ctx.MultiInputVar("X"); auto x_vars = ctx.MultiInputVar("X");
auto x_vars_name = ctx.Inputs("X");
framework::LibraryType library{framework::LibraryType::kPlain}; framework::LibraryType library{framework::LibraryType::kPlain};
framework::DataLayout layout{framework::DataLayout::kAnyLayout}; framework::DataLayout layout{framework::DataLayout::kAnyLayout};
...@@ -81,10 +82,11 @@ class SumOp : public framework::OperatorWithKernel { ...@@ -81,10 +82,11 @@ class SumOp : public framework::OperatorWithKernel {
if (x_vars[0]->IsType<framework::LoDTensor>()) { if (x_vars[0]->IsType<framework::LoDTensor>()) {
int dtype = -1; int dtype = -1;
for (auto& x_var : x_vars) { for (size_t idx = 0; idx < x_vars.size(); ++idx) {
PADDLE_ENFORCE(x_vars[idx] != nullptr,
"Input var[%s] should not be nullptr", x_vars_name[idx]);
// FIXME(zcd): The input x_var may be SelectedRows or LoDTensor. // FIXME(zcd): The input x_var may be SelectedRows or LoDTensor.
auto tensor = framework::GetTensorFromVar( auto tensor = framework::GetTensorFromVar(*x_vars[idx]);
const_cast<framework::Variable*>(x_var));
if (tensor->numel() == 0) { if (tensor->numel() == 0) {
continue; continue;
} }
......
paddle/scripts/paddle_build.sh
浏览文件 @ d31e1940
...@@ -147,7 +147,6 @@ function cmake_gen() { ...@@ -147,7 +147,6 @@ function cmake_gen() {
-DWITH_SWIG_PY=${WITH_SWIG_PY:-ON} -DWITH_SWIG_PY=${WITH_SWIG_PY:-ON}
-DCUDNN_ROOT=/usr/ -DCUDNN_ROOT=/usr/
-DWITH_TESTING=${WITH_TESTING:-ON} -DWITH_TESTING=${WITH_TESTING:-ON}
-DWITH_FAST_BUNDLE_TEST=ON
-DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
-DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF}
...@@ -180,7 +179,6 @@ EOF ...@@ -180,7 +179,6 @@ EOF
-DWITH_PYTHON=${WITH_PYTHON:-ON} \ -DWITH_PYTHON=${WITH_PYTHON:-ON} \
-DCUDNN_ROOT=/usr/ \ -DCUDNN_ROOT=/usr/ \
-DWITH_TESTING=${WITH_TESTING:-ON} \ -DWITH_TESTING=${WITH_TESTING:-ON} \
-DWITH_FAST_BUNDLE_TEST=ON \
-DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \ -DCMAKE_MODULE_PATH=/opt/rocm/hip/cmake \
-DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \ -DWITH_FLUID_ONLY=${WITH_FLUID_ONLY:-OFF} \
-DCMAKE_EXPORT_COMPILE_COMMANDS=ON \ -DCMAKE_EXPORT_COMPILE_COMMANDS=ON \
......
python/paddle/fluid/tests/unittests/test_dist_transpiler.py
浏览文件 @ d31e1940
...@@ -283,6 +283,25 @@ class TestDecayedAdagrad(TranspilerTest): ...@@ -283,6 +283,25 @@ class TestDecayedAdagrad(TranspilerTest):
trainer, _ = self.get_trainer() 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): class TestLRDecayConditional(TranspilerTest):
def net_conf(self): def net_conf(self):
x = fluid.layers.data(name='x', shape=[1000], dtype='float32') x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
...@@ -405,18 +424,43 @@ class TestL2DecayWithPiecewise(TranspilerTest): ...@@ -405,18 +424,43 @@ class TestL2DecayWithPiecewise(TranspilerTest):
["sum", "scale", "scale", "elementwise_add", "momentum"]) ["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): class TestDistLookupTableBase(TranspilerTest):
def network_with_table(self, is_sparse, is_distributed): def network_with_table(self, is_sparse, is_distributed):
self.table_size = 1000 self.table_size = 1000
self.emb_size = 64 self.emb_size = 64
self.lookup_table_name = 'shared_w' self.lookup_table_name = 'shared_w'
def emb_pool(ids): def emb_pool(ids, table_name, is_distributed):
emb = fluid.layers.embedding( emb = fluid.layers.embedding(
input=ids, input=ids,
size=[self.table_size, self.emb_size], size=[self.table_size, self.emb_size],
dtype='float32', dtype='float32',
param_attr=self.lookup_table_name, # share parameter param_attr=table_name,
is_sparse=is_sparse, is_sparse=is_sparse,
is_distributed=is_distributed) is_distributed=is_distributed)
pool = fluid.layers.sequence_pool(input=emb, pool_type='average') pool = fluid.layers.sequence_pool(input=emb, pool_type='average')
...@@ -426,9 +470,13 @@ class TestDistLookupTableBase(TranspilerTest): ...@@ -426,9 +470,13 @@ class TestDistLookupTableBase(TranspilerTest):
name='title_ids', shape=[1], dtype='int64', lod_level=1) name='title_ids', shape=[1], dtype='int64', lod_level=1)
brand_ids = fluid.layers.data( brand_ids = fluid.layers.data(
name='brand_ids', shape=[1], dtype='int64', lod_level=1) name='brand_ids', shape=[1], dtype='int64', lod_level=1)
title_emb = emb_pool(title_ids) profile_ids = fluid.layers.data(
brand_emb = emb_pool(brand_ids) name='brand_ids', shape=[1], dtype='int64', lod_level=1)
fc0 = fluid.layers.concat(input=[title_emb, brand_emb], axis=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, predict = fluid.layers.fc(input=fc0,
size=2, size=2,
act=None, act=None,
...@@ -449,7 +497,7 @@ class TestLocalLookupTable(TestDistLookupTableBase): ...@@ -449,7 +497,7 @@ class TestLocalLookupTable(TestDistLookupTableBase):
def transpiler_test_impl(self): def transpiler_test_impl(self):
pserver1, startup1 = self.get_pserver(self.pserver1_ep) pserver1, startup1 = self.get_pserver(self.pserver1_ep)
self.assertEqual(len(pserver1.blocks), 3) self.assertEqual(len(pserver1.blocks), 4)
# 0 listen_and_serv # 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam # 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops], self.assertEqual([op.type for op in pserver1.blocks[1].ops],
...@@ -459,16 +507,23 @@ class TestLocalLookupTable(TestDistLookupTableBase): ...@@ -459,16 +507,23 @@ class TestLocalLookupTable(TestDistLookupTableBase):
self.assertEqual([op.type for op in pserver1.blocks[2].ops], self.assertEqual([op.type for op in pserver1.blocks[2].ops],
["sum", "scale", "adam", "scale", "scale"]) ["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() trainer, _ = self.get_trainer()
self.assertEqual(len(trainer.blocks), 1) self.assertEqual(len(trainer.blocks), 1)
ops = [ ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', 'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'fill_constant', 'mean_grad', 'cross_entropy_grad', 'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', 'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'lookup_table_grad', 'sum', 'split_selected_rows', 'send', 'split_selected_rows', 'send', 'sequence_pool_grad',
'send_barrier', 'recv', 'recv', 'recv', 'fetch_barrier', 'concat' '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) self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
...@@ -480,39 +535,45 @@ class TestDistLookupTable(TestDistLookupTableBase): ...@@ -480,39 +535,45 @@ class TestDistLookupTable(TestDistLookupTableBase):
def transpiler_test_impl(self): def transpiler_test_impl(self):
pserver1, startup1 = self.get_pserver(self.pserver1_ep) pserver1, startup1 = self.get_pserver(self.pserver1_ep)
self.assertEqual(len(pserver1.blocks), 5) self.assertEqual(len(pserver1.blocks), 6)
# 0 listen_and_serv # 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam # 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops], self.assertEqual([op.type for op in pserver1.blocks[1].ops],
["sum", "scale", "adam", "scale", "scale"]) ["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], 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"]) ["sum", "sgd"])
# 3 prefetch -> lookup_sparse_table for data0 # 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"]) ["lookup_sparse_table"])
# 4 save table # 5 save table
self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["save"]) self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
trainer, trainer_startup = self.get_trainer() trainer, trainer_startup = self.get_trainer()
self.assertEqual(len(trainer.blocks), 1) self.assertEqual(len(trainer.blocks), 1)
ops = [ ops = [
'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
'sequence_pool', 'concat', 'mul', 'elementwise_add', 'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad', 'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad', 'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad', 'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sum', 'split_ids', 'lookup_table_grad', 'split_selected_rows', 'send',
'send', 'send_barrier', 'recv', 'recv', 'fetch_barrier' '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) self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
startup_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', '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', 'fill_constant', 'fill_constant', 'fill_constant', 'fill_constant',
'fetch_barrier', 'fake_init' '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], self.assertEqual([op.type for op in trainer_startup.blocks[0].ops],
startup_ops) startup_ops)
...@@ -526,7 +587,7 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase): ...@@ -526,7 +587,7 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
config = fluid.DistributeTranspilerConfig() config = fluid.DistributeTranspilerConfig()
pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False) 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 # 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam # 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops], self.assertEqual([op.type for op in pserver1.blocks[1].ops],
...@@ -535,17 +596,23 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase): ...@@ -535,17 +596,23 @@ class TestAsyncLocalLookupTable(TestDistLookupTableBase):
# NOTE: if param is not selected rows, the grad will scaled to grad / trainer_num # 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], self.assertEqual([op.type for op in pserver1.blocks[2].ops],
["adam", "scale", "scale"]) ["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) trainer, _ = self.get_trainer(config)
self.assertEqual(len(trainer.blocks), 1) self.assertEqual(len(trainer.blocks), 1)
ops = [ ops = [
'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool', 'lookup_table', 'sequence_pool',
'concat', 'mul', 'elementwise_add', 'cross_entropy', 'mean', 'lookup_table', 'sequence_pool', 'concat', 'mul', 'elementwise_add',
'fill_constant', 'mean_grad', 'cross_entropy_grad', 'cross_entropy', 'mean', 'fill_constant', 'mean_grad',
'elementwise_add_grad', 'send', 'mul_grad', 'send', 'concat_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sequence_pool_grad', 'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad',
'lookup_table_grad', 'sum', 'split_selected_rows', 'send', 'recv', 'split_selected_rows', 'send', 'sequence_pool_grad',
'recv', 'recv', 'concat' '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) self.assertEqual([op.type for op in trainer.blocks[0].ops], ops)
...@@ -559,29 +626,34 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase): ...@@ -559,29 +626,34 @@ class TestAsyncDistLookupTable(TestDistLookupTableBase):
pserver1, startup1 = self.get_pserver(self.pserver1_ep, config, False) 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 # 0 listen_and_serv
# 1 optimize for fc_w or fc_b adam # 1 optimize for fc_w or fc_b adam
self.assertEqual([op.type for op in pserver1.blocks[1].ops], self.assertEqual([op.type for op in pserver1.blocks[1].ops],
["adam", "scale", "scale"]) ["adam", "scale", "scale"])
# 2 optimize for table sgd # 2 optimize for table adam
self.assertEqual([op.type for op in pserver1.blocks[2].ops], ["sgd"]) self.assertEqual([op.type for op in pserver1.blocks[2].ops],
# 3 prefetch -> lookup_sparse_table for data0 ["adam", "scale", "scale"])
self.assertEqual([op.type for op in pserver1.blocks[3].ops], # 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"]) ["lookup_sparse_table"])
# 4 save table # 5 save table
self.assertEqual([op.type for op in pserver1.blocks[4].ops], ["save"]) self.assertEqual([op.type for op in pserver1.blocks[5].ops], ["save"])
trainer, _ = self.get_trainer(config) trainer, _ = self.get_trainer(config)
self.assertEqual(len(trainer.blocks), 1) self.assertEqual(len(trainer.blocks), 1)
ops = [ ops = [
'split_ids', 'prefetch', 'merge_ids', 'sequence_pool', 'split_ids', 'prefetch', 'merge_ids', 'sequence_pool',
'sequence_pool', 'concat', 'mul', 'elementwise_add', 'sequence_pool', 'lookup_table', 'sequence_pool', 'concat', 'mul',
'cross_entropy', 'mean', 'fill_constant', 'mean_grad', 'elementwise_add', 'cross_entropy', 'mean', 'fill_constant',
'cross_entropy_grad', 'elementwise_add_grad', 'send', 'mul_grad', 'mean_grad', 'cross_entropy_grad', 'elementwise_add_grad', 'send',
'send', 'concat_grad', 'sequence_pool_grad', 'lookup_table_grad', 'mul_grad', 'send', 'concat_grad', 'sequence_pool_grad',
'sequence_pool_grad', 'lookup_table_grad', 'sum', 'split_ids', 'lookup_table_grad', 'split_selected_rows', 'send',
'send', 'recv', 'recv' '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) 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 @ d31e1940
...@@ -55,6 +55,46 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id): ...@@ -55,6 +55,46 @@ def run_pserver(use_cuda, sync_mode, ip, port, trainers, trainer_id):
exe.run(pserver_prog) 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): class TestListenAndServOp(OpTest):
def setUp(self): def setUp(self):
self.ps_timeout = 5 self.ps_timeout = 5
...@@ -63,9 +103,9 @@ class TestListenAndServOp(OpTest): ...@@ -63,9 +103,9 @@ class TestListenAndServOp(OpTest):
self.trainers = 1 self.trainers = 1
self.trainer_id = 0 self.trainer_id = 0
def _start_pserver(self, use_cuda, sync_mode): def _start_pserver(self, use_cuda, sync_mode, pserver_func):
p = Process( p = Process(
target=run_pserver, target=pserver_func,
args=(use_cuda, sync_mode, self.ip, self.port, self.trainers, args=(use_cuda, sync_mode, self.ip, self.port, self.trainers,
self.trainer_id)) self.trainer_id))
p.daemon = True p.daemon = True
...@@ -92,7 +132,24 @@ class TestListenAndServOp(OpTest): ...@@ -92,7 +132,24 @@ class TestListenAndServOp(OpTest):
def test_handle_signal_in_serv_op(self): def test_handle_signal_in_serv_op(self):
# run pserver on CPU in sync mode # 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) self._wait_ps_ready(p1.pid)
# raise SIGTERM to pserver # raise SIGTERM to pserver
...@@ -100,7 +157,7 @@ class TestListenAndServOp(OpTest): ...@@ -100,7 +157,7 @@ class TestListenAndServOp(OpTest):
p1.join() p1.join()
# run pserver on CPU in async mode # 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) self._wait_ps_ready(p2.pid)
# raise SIGTERM to pserver # raise SIGTERM to pserver
......
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ d31e1940
...@@ -35,6 +35,7 @@ import sys ...@@ -35,6 +35,7 @@ import sys
import numpy as np import numpy as np
import collections import collections
import six import six
import logging
from .ps_dispatcher import RoundRobin, HashName, PSDispatcher from .ps_dispatcher import RoundRobin, HashName, PSDispatcher
from .. import core, framework from .. import core, framework
...@@ -767,6 +768,15 @@ in a single call.") ...@@ -767,6 +768,15 @@ in a single call.")
prefetch_var_name_to_block_id.extend( prefetch_var_name_to_block_id.extend(
lookup_table_var_name_to_block_id) 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 = { attrs = {
"optimize_blocks": optimize_blocks, "optimize_blocks": optimize_blocks,
"endpoint": endpoint, "endpoint": endpoint,
...@@ -1065,7 +1075,12 @@ to transpile() call.") ...@@ -1065,7 +1075,12 @@ to transpile() call.")
continue_search_lookup_table_op = False continue_search_lookup_table_op = False
all_ops = program.global_block().ops all_ops = program.global_block().ops
for op in all_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 continue_search_lookup_table_op = True
lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list( lookup_table_op_index = lookup_table_op_index if lookup_table_op_index != -1 else list(
...@@ -1275,7 +1290,6 @@ to transpile() call.") ...@@ -1275,7 +1290,6 @@ to transpile() call.")
} }
outputs = {"ParamOut": [param_var]} outputs = {"ParamOut": [param_var]}
# only support sgd now # only support sgd now
import logging
logging.warn( logging.warn(
"distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of " "distribute lookup table only support sgd optimizer, change it's optimizer to sgd instead of "
+ table_opt_op.type) + table_opt_op.type)
...@@ -1442,6 +1456,9 @@ to transpile() call.") ...@@ -1442,6 +1456,9 @@ to transpile() call.")
elif op_type == "decayed_adagrad": elif op_type == "decayed_adagrad":
if varkey == "Moment": if varkey == "Moment":
return param_shape return param_shape
elif op_type == "ftrl":
if varkey in ["SquaredAccumulator", "LinearAccumulator"]:
return param_shape
elif op_type == "sgd": elif op_type == "sgd":
pass pass
else: else:
......
python/paddle/fluid/transpiler/inference_transpiler.py
浏览文件 @ d31e1940
...@@ -61,6 +61,9 @@ class InferenceTranspiler(object): ...@@ -61,6 +61,9 @@ class InferenceTranspiler(object):
raise TypeError("scope should be as Scope type or None") raise TypeError("scope should be as Scope type or None")
use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False)) use_mkldnn = bool(os.getenv("FLAGS_use_mkldnn", False))
if use_mkldnn:
self._depthwise_conv_mkldnn(program)
self._fuse_batch_norm(program, place, scope) self._fuse_batch_norm(program, place, scope)
if use_mkldnn: if use_mkldnn:
self._fuse_conv_bias_mkldnn(program) self._fuse_conv_bias_mkldnn(program)
...@@ -70,6 +73,31 @@ class InferenceTranspiler(object): ...@@ -70,6 +73,31 @@ class InferenceTranspiler(object):
program) # ResNet residual block merging program) # ResNet residual block merging
self._fuse_bn_relu_mkldnn(program) 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): def _fuse_conv_eltwise_mkldnn(self, program):
''' '''
Transpile the program fusing elementwise_add into conv for MKLDNN 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