Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into accelerate_ddpg
test=developrevert-15207-remove_op_handle_lock_and_fix_var
minqiyang
6 years ago
commit
8ed0233924
@ -0,0 +1,54 @@
|
|||||||
|
// 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/scope_pool.h"
|
||||||
|
#include "paddle/fluid/framework/threadpool.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace framework {
|
||||||
|
|
||||||
|
ScopePool &ScopePool::Instance() { // NOLINT
|
||||||
|
static ScopePool pool;
|
||||||
|
return pool;
|
||||||
|
}
|
||||||
|
|
||||||
|
void ScopePool::DeleteScope(Scope *scope) { delete scope; }
|
||||||
|
|
||||||
|
void ScopePool::Insert(std::unique_ptr<Scope> &&s) {
|
||||||
|
std::lock_guard<std::mutex> guard(mtx_);
|
||||||
|
scopes_.insert(s.release());
|
||||||
|
}
|
||||||
|
|
||||||
|
void ScopePool::Remove(Scope *s) {
|
||||||
|
size_t has_scope;
|
||||||
|
{
|
||||||
|
std::lock_guard<std::mutex> guard(mtx_);
|
||||||
|
has_scope = scopes_.erase(s);
|
||||||
|
}
|
||||||
|
PADDLE_ENFORCE(has_scope > 0, "Delete non-existing global scope");
|
||||||
|
DeleteScope(s);
|
||||||
|
}
|
||||||
|
|
||||||
|
ScopePool::~ScopePool() { Clear(); }
|
||||||
|
|
||||||
|
void ScopePool::Clear() {
|
||||||
|
std::lock_guard<std::mutex> guard(mtx_);
|
||||||
|
for (auto *s : scopes_) {
|
||||||
|
DeleteScope(s);
|
||||||
|
}
|
||||||
|
scopes_.clear();
|
||||||
|
}
|
||||||
|
|
||||||
|
} // namespace framework
|
||||||
|
} // namespace paddle
|
||||||
@ -0,0 +1,46 @@
|
|||||||
|
// 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 <mutex> // NOLINT
|
||||||
|
#include <unordered_set>
|
||||||
|
#include "paddle/fluid/framework/scope.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace framework {
|
||||||
|
|
||||||
|
class ScopePool {
|
||||||
|
public:
|
||||||
|
static ScopePool &Instance(); // NOLINT
|
||||||
|
|
||||||
|
void Insert(std::unique_ptr<Scope> &&s);
|
||||||
|
|
||||||
|
void Remove(Scope *s);
|
||||||
|
|
||||||
|
void Clear();
|
||||||
|
|
||||||
|
~ScopePool();
|
||||||
|
|
||||||
|
private:
|
||||||
|
ScopePool() = default;
|
||||||
|
|
||||||
|
static void DeleteScope(Scope *scope);
|
||||||
|
|
||||||
|
std::unordered_set<Scope *> scopes_;
|
||||||
|
std::mutex mtx_;
|
||||||
|
};
|
||||||
|
|
||||||
|
} // namespace framework
|
||||||
|
} // namespace paddle
|
||||||
@ -0,0 +1,178 @@
|
|||||||
|
// 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/inference/tests/api/tester_helper.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace inference {
|
||||||
|
using contrib::AnalysisConfig;
|
||||||
|
|
||||||
|
struct DataRecord {
|
||||||
|
std::vector<std::vector<int64_t>> query_data_all, title_data_all;
|
||||||
|
std::vector<size_t> lod1, lod2;
|
||||||
|
size_t batch_iter{0};
|
||||||
|
size_t batch_size{1};
|
||||||
|
size_t num_samples; // total number of samples
|
||||||
|
DataRecord() = default;
|
||||||
|
explicit DataRecord(const std::string &path, int batch_size = 1)
|
||||||
|
: batch_size(batch_size) {
|
||||||
|
Load(path);
|
||||||
|
}
|
||||||
|
DataRecord NextBatch() {
|
||||||
|
DataRecord data;
|
||||||
|
size_t batch_end = batch_iter + batch_size;
|
||||||
|
// NOTE skip the final batch, if no enough data is provided.
|
||||||
|
if (batch_end <= query_data_all.size()) {
|
||||||
|
data.query_data_all.assign(query_data_all.begin() + batch_iter,
|
||||||
|
query_data_all.begin() + batch_end);
|
||||||
|
data.title_data_all.assign(title_data_all.begin() + batch_iter,
|
||||||
|
title_data_all.begin() + batch_end);
|
||||||
|
// Prepare LoDs
|
||||||
|
data.lod1.push_back(0);
|
||||||
|
data.lod2.push_back(0);
|
||||||
|
CHECK(!data.query_data_all.empty());
|
||||||
|
CHECK(!data.title_data_all.empty());
|
||||||
|
CHECK_EQ(data.query_data_all.size(), data.title_data_all.size());
|
||||||
|
for (size_t j = 0; j < data.query_data_all.size(); j++) {
|
||||||
|
// calculate lod
|
||||||
|
data.lod1.push_back(data.lod1.back() + data.query_data_all[j].size());
|
||||||
|
data.lod2.push_back(data.lod2.back() + data.title_data_all[j].size());
|
||||||
|
}
|
||||||
|
}
|
||||||
|
batch_iter += batch_size;
|
||||||
|
return data;
|
||||||
|
}
|
||||||
|
void Load(const std::string &path) {
|
||||||
|
std::ifstream file(path);
|
||||||
|
std::string line;
|
||||||
|
int num_lines = 0;
|
||||||
|
while (std::getline(file, line)) {
|
||||||
|
num_lines++;
|
||||||
|
std::vector<std::string> data;
|
||||||
|
split(line, '\t', &data);
|
||||||
|
// load query data
|
||||||
|
std::vector<int64_t> query_data;
|
||||||
|
split_to_int64(data[0], ' ', &query_data);
|
||||||
|
// load title data
|
||||||
|
std::vector<int64_t> title_data;
|
||||||
|
split_to_int64(data[1], ' ', &title_data);
|
||||||
|
query_data_all.push_back(std::move(query_data));
|
||||||
|
title_data_all.push_back(std::move(title_data));
|
||||||
|
}
|
||||||
|
num_samples = num_lines;
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
|
||||||
|
int batch_size) {
|
||||||
|
PaddleTensor lod_query_tensor, lod_title_tensor;
|
||||||
|
lod_query_tensor.name = "left";
|
||||||
|
lod_title_tensor.name = "right";
|
||||||
|
auto one_batch = data->NextBatch();
|
||||||
|
int size1 = one_batch.lod1[one_batch.lod1.size() - 1]; // token batch size
|
||||||
|
int size2 = one_batch.lod2[one_batch.lod2.size() - 1]; // token batch size
|
||||||
|
lod_query_tensor.shape.assign({size1, 1});
|
||||||
|
lod_query_tensor.lod.assign({one_batch.lod1});
|
||||||
|
lod_title_tensor.shape.assign({size2, 1});
|
||||||
|
lod_title_tensor.lod.assign({one_batch.lod2});
|
||||||
|
// assign data
|
||||||
|
TensorAssignData<int64_t>(&lod_query_tensor, one_batch.query_data_all);
|
||||||
|
TensorAssignData<int64_t>(&lod_title_tensor, one_batch.title_data_all);
|
||||||
|
// Set inputs.
|
||||||
|
input_slots->assign({lod_query_tensor, lod_title_tensor});
|
||||||
|
for (auto &tensor : *input_slots) {
|
||||||
|
tensor.dtype = PaddleDType::INT64;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
void SetConfig(contrib::AnalysisConfig *cfg) {
|
||||||
|
cfg->model_dir = FLAGS_infer_model;
|
||||||
|
cfg->use_gpu = false;
|
||||||
|
cfg->device = 0;
|
||||||
|
cfg->specify_input_name = true;
|
||||||
|
cfg->enable_ir_optim = true;
|
||||||
|
}
|
||||||
|
|
||||||
|
void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
|
||||||
|
DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
|
||||||
|
std::vector<PaddleTensor> input_slots;
|
||||||
|
int epoch = FLAGS_test_all_data ? data.num_samples / FLAGS_batch_size : 1;
|
||||||
|
LOG(INFO) << "number of samples: " << epoch * FLAGS_batch_size;
|
||||||
|
for (int bid = 0; bid < epoch; ++bid) {
|
||||||
|
PrepareInputs(&input_slots, &data, FLAGS_batch_size);
|
||||||
|
(*inputs).emplace_back(input_slots);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
// Easy for profiling independently.
|
||||||
|
TEST(Analyzer_MM_DNN, profile) {
|
||||||
|
contrib::AnalysisConfig cfg;
|
||||||
|
SetConfig(&cfg);
|
||||||
|
std::vector<PaddleTensor> outputs;
|
||||||
|
|
||||||
|
std::vector<std::vector<PaddleTensor>> input_slots_all;
|
||||||
|
SetInput(&input_slots_all);
|
||||||
|
TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
|
||||||
|
input_slots_all, &outputs, FLAGS_num_threads);
|
||||||
|
|
||||||
|
if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
|
||||||
|
PADDLE_ENFORCE_EQ(outputs.size(), 2UL);
|
||||||
|
for (auto &output : outputs) {
|
||||||
|
size_t size = GetSize(output);
|
||||||
|
PADDLE_ENFORCE_GT(size, 0);
|
||||||
|
float *result = static_cast<float *>(output.data.data());
|
||||||
|
// output is probability, which is in (-1, 1).
|
||||||
|
for (size_t i = 0; i < size; i++) {
|
||||||
|
EXPECT_GT(result[i], -1);
|
||||||
|
EXPECT_LT(result[i], 1);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
|
// Check the fuse status
|
||||||
|
TEST(Analyzer_MM_DNN, fuse_statis) {
|
||||||
|
contrib::AnalysisConfig cfg;
|
||||||
|
SetConfig(&cfg);
|
||||||
|
|
||||||
|
int num_ops;
|
||||||
|
auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
|
||||||
|
auto fuse_statis = GetFuseStatis(
|
||||||
|
static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
|
||||||
|
}
|
||||||
|
|
||||||
|
// Compare result of NativeConfig and AnalysisConfig
|
||||||
|
TEST(Analyzer_MM_DNN, compare) {
|
||||||
|
contrib::AnalysisConfig cfg;
|
||||||
|
SetConfig(&cfg);
|
||||||
|
|
||||||
|
std::vector<std::vector<PaddleTensor>> input_slots_all;
|
||||||
|
SetInput(&input_slots_all);
|
||||||
|
CompareNativeAndAnalysis(
|
||||||
|
reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
|
||||||
|
}
|
||||||
|
|
||||||
|
// Compare Deterministic result
|
||||||
|
TEST(Analyzer_MM_DNN, compare_determine) {
|
||||||
|
AnalysisConfig cfg;
|
||||||
|
SetConfig(&cfg);
|
||||||
|
|
||||||
|
std::vector<std::vector<PaddleTensor>> input_slots_all;
|
||||||
|
SetInput(&input_slots_all);
|
||||||
|
CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
|
||||||
|
input_slots_all);
|
||||||
|
}
|
||||||
|
|
||||||
|
} // namespace inference
|
||||||
|
} // namespace paddle
|
||||||
@ -0,0 +1,117 @@
|
|||||||
|
/* 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 <fstream>
|
||||||
|
#include <iostream>
|
||||||
|
#include "paddle/fluid/inference/tests/api/tester_helper.h"
|
||||||
|
|
||||||
|
namespace paddle {
|
||||||
|
namespace inference {
|
||||||
|
namespace analysis {
|
||||||
|
|
||||||
|
void SetConfig(AnalysisConfig *cfg) {
|
||||||
|
cfg->param_file = FLAGS_infer_model + "/params";
|
||||||
|
cfg->prog_file = FLAGS_infer_model + "/model";
|
||||||
|
cfg->use_gpu = false;
|
||||||
|
cfg->device = 0;
|
||||||
|
cfg->enable_ir_optim = true;
|
||||||
|
cfg->specify_input_name = true;
|
||||||
|
cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
|
||||||
|
}
|
||||||
|
|
||||||
|
void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
|
||||||
|
std::vector<std::string> feed_names = {
|
||||||
|
"slot10000_embed", "slot10001_embed", "slot10004_embed",
|
||||||
|
"slot10005_embed", "slot10008_embed", "slot10009_embed",
|
||||||
|
"slot10012_embed", "slot10013_embed", "slot10108_embed",
|
||||||
|
"slot13324_embed", "slot13325_embed", "slot13326_embed",
|
||||||
|
"slot13327_embed", "slot13328_embed", "slot13329_embed",
|
||||||
|
"slot13330_embed", "slot13331_embed", "slot15501_embed",
|
||||||
|
"slot15502_embed", "slot15503_embed", "slot15504_embed",
|
||||||
|
"slot15505_embed", "slot15506_embed", "slot15507_embed",
|
||||||
|
"slot15508_embed", "slot15516_embed", "slot15519_embed",
|
||||||
|
"slot15523_embed", "slot15531_embed", "slot15533_embed",
|
||||||
|
"slot15548_embed", "slot15564_embed", "slot15565_embed",
|
||||||
|
"slot15566_embed", "slot15570_embed", "slot15571_embed",
|
||||||
|
"slot15572_embed", "slot15573_embed", "slot15574_embed",
|
||||||
|
"slot15575_embed", "slot15576_embed", "slot15577_embed",
|
||||||
|
"slot15579_embed", "slot15581_embed", "slot15582_embed",
|
||||||
|
"slot15583_embed", "slot15584_embed", "slot5016_embed",
|
||||||
|
"slot5021_embed", "slot6002_embed", "slot6003_embed",
|
||||||
|
"slot6004_embed", "slot6005_embed", "slot6006_embed",
|
||||||
|
"slot6007_embed", "slot6008_embed", "slot6009_embed",
|
||||||
|
"slot6011_embed", "slot6014_embed", "slot6015_embed",
|
||||||
|
"slot6023_embed", "slot6024_embed", "slot6025_embed",
|
||||||
|
"slot6027_embed", "slot6029_embed", "slot6031_embed",
|
||||||
|
"slot6034_embed", "slot6035_embed", "slot6036_embed",
|
||||||
|
"slot6037_embed", "slot6039_embed", "slot6048_embed",
|
||||||
|
"slot6050_embed", "slot6058_embed", "slot6059_embed",
|
||||||
|
"slot6060_embed", "slot6066_embed", "slot6067_embed",
|
||||||
|
"slot6068_embed", "slot6069_embed", "slot6070_embed",
|
||||||
|
"slot6071_embed", "slot6072_embed", "slot6073_embed",
|
||||||
|
"slot6182_embed", "slot6183_embed", "slot6184_embed",
|
||||||
|
"slot6185_embed", "slot6186_embed", "slot6188_embed",
|
||||||
|
"slot6189_embed", "slot6190_embed", "slot6201_embed",
|
||||||
|
"slot6202_embed", "slot6203_embed", "slot6247_embed",
|
||||||
|
"slot6248_embed", "slot6250_embed", "slot6251_embed",
|
||||||
|
"slot6807_embed", "slot6808_embed", "slot6809_embed",
|
||||||
|
"slot6810_embed", "slot6811_embed", "slot6812_embed",
|
||||||
|
"slot6813_embed", "slot6814_embed", "slot6815_embed",
|
||||||
|
"slot6816_embed", "slot6817_embed", "slot6818_embed",
|
||||||
|
"slot6819_embed", "slot6820_embed", "slot6822_embed",
|
||||||
|
"slot6823_embed", "slot6826_embed", "slot7002_embed",
|
||||||
|
"slot7003_embed", "slot7004_embed", "slot7005_embed",
|
||||||
|
"slot7006_embed", "slot7008_embed", "slot7009_embed",
|
||||||
|
"slot7010_embed", "slot7011_embed", "slot7013_embed",
|
||||||
|
"slot7014_embed", "slot7015_embed", "slot7016_embed",
|
||||||
|
"slot7017_embed", "slot7019_embed", "slot7100_embed",
|
||||||
|
"slot7506_embed", "slot7507_embed", "slot7514_embed",
|
||||||
|
"slot7515_embed", "slot7516_embed"};
|
||||||
|
SetFakeImageInput(inputs, FLAGS_infer_model, true, "model", "params",
|
||||||
|
&feed_names);
|
||||||
|
}
|
||||||
|
|
||||||
|
// Easy for profiling independently.
|
||||||
|
void profile(bool use_mkldnn = false) {
|
||||||
|
AnalysisConfig cfg;
|
||||||
|
SetConfig(&cfg);
|
||||||
|
|
||||||
|
if (use_mkldnn) {
|
||||||
|
cfg.EnableMKLDNN();
|
||||||
|
}
|
||||||
|
std::vector<PaddleTensor> outputs;
|
||||||
|
|
||||||
|
std::vector<std::vector<PaddleTensor>> input_slots_all;
|
||||||
|
SetInput(&input_slots_all);
|
||||||
|
TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
|
||||||
|
input_slots_all, &outputs, FLAGS_num_threads);
|
||||||
|
}
|
||||||
|
|
||||||
|
TEST(Analyzer_seq_pool1, profile) { profile(); }
|
||||||
|
|
||||||
|
// Check the fuse status
|
||||||
|
TEST(Analyzer_seq_pool1, fuse_statis) {
|
||||||
|
AnalysisConfig cfg;
|
||||||
|
SetConfig(&cfg);
|
||||||
|
int num_ops;
|
||||||
|
auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
|
||||||
|
auto fuse_statis = GetFuseStatis(
|
||||||
|
static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
|
||||||
|
LOG(INFO) << "num_ops: " << num_ops;
|
||||||
|
EXPECT_EQ(num_ops, 314);
|
||||||
|
}
|
||||||
|
|
||||||
|
} // namespace analysis
|
||||||
|
} // namespace inference
|
||||||
|
} // namespace paddle
|
||||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue