analysis_predictor.cc 5.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// 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.

Y
Yan Chunwei 已提交
15
#include "paddle/fluid/inference/api/analysis_predictor.h"
16
#include <memory>
17 18
#include <string>
#include <vector>
Y
Yan Chunwei 已提交
19
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
20 21 22
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
L
luotao1 已提交
23
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
24
#include "paddle/fluid/inference/utils/singleton.h"
T
tensor-tang 已提交
25 26 27
#include "paddle/fluid/platform/profiler.h"

DECLARE_bool(profile);
28 29 30

namespace paddle {

Y
Yan Chunwei 已提交
31 32 33
bool AnalysisPredictor::Init(
    const std::shared_ptr<framework::Scope>& parent_scope) {
  VLOG(3) << "Predictor::init()";
T
tensor-tang 已提交
34 35 36 37 38 39 40 41 42 43
#if !defined(_WIN32)
  if (FLAGS_profile) {
    LOG(WARNING) << "Profiler is actived, might affect the performance";
    LOG(INFO) << "You can turn off by set gflags '-profile false'";
    auto tracking_device = config_.use_gpu ? platform::ProfilerState::kAll
                                           : platform::ProfilerState::kCPU;
    platform::EnableProfiler(tracking_device);
  }
#endif

Y
Yan Chunwei 已提交
44 45
  if (config_.use_gpu) {
    place_ = paddle::platform::CUDAPlace(config_.device);
46 47
    LOG(WARNING) << "ir optimize only supports CPU currently";
    config_.enable_ir_optim = false;
Y
Yan Chunwei 已提交
48 49 50 51 52 53 54 55 56 57
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  if (parent_scope) {
    scope_ = parent_scope;
    sub_scope_ = &(parent_scope->NewScope());
  } else {
    paddle::framework::InitDevices(false);
    scope_.reset(new paddle::framework::Scope());
  }
58

Y
Yan Chunwei 已提交
59
  executor_.reset(new paddle::framework::Executor(place_));
60

Y
Yan Chunwei 已提交
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
  // Initialize the inference program
  if (!config_.model_dir.empty()) {
    // Parameters are saved in separate files sited in
    // the specified `dirname`.
    inference_program_ = paddle::inference::Load(executor_.get(), scope_.get(),
                                                 config_.model_dir);
  } else if (!config_.prog_file.empty() && !config_.param_file.empty()) {
    // All parameters are saved in a single file.
    // The file names should be consistent with that used
    // in Python API `fluid.io.save_inference_model`.
    inference_program_ = paddle::inference::Load(
        executor_.get(), scope_.get(), config_.prog_file, config_.param_file);
  } else {
    LOG(ERROR) << "fail to load inference model.";
    return false;
76 77
  }

Y
Yan Chunwei 已提交
78 79
  OptimizeInferenceProgram();
  ctx_ = executor_->Prepare(*inference_program_, 0);
T
tensor-tang 已提交
80
  if (config_._use_mkldnn) {
T
tensor-tang 已提交
81 82
    executor_->EnableMKLDNN(*inference_program_);
  }
83

Y
Yan Chunwei 已提交
84 85 86 87 88 89 90 91
  VLOG(5) << "to create variables";
  PADDLE_ENFORCE(scope_.get());
  executor_->CreateVariables(*inference_program_,
                             sub_scope_ ? sub_scope_ : scope_.get(), 0);
  // Get the feed_target_names and fetch_target_names
  PrepareFeedFetch();
  return true;
}
92

Y
Yan Chunwei 已提交
93 94
void AnalysisPredictor::OptimizeInferenceProgram() {
  LOG(INFO) << "optimize begin";
95
  FLAGS_IA_enable_ir = config_.enable_ir_optim;
Y
Yan Chunwei 已提交
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
  FLAGS_IA_enable_tensorrt_subgraph_engine = false;
  FLAGS_IA_output_storage_path = "";  // Don't output the model.
  // Analyze inference_program
  if (!config_.model_dir.empty()) {
    argument_.fluid_model_dir.reset(new std::string(config_.model_dir));
  } else {
    PADDLE_ENFORCE(
        !config_.param_file.empty(),
        "Either model_dir or (param_file, prog_file) should be set.");
    PADDLE_ENFORCE(!config_.prog_file.empty());
    argument_.fluid_model_program_path.reset(
        new std::string(config_.prog_file));
    argument_.fluid_model_param_path.reset(new std::string(config_.param_file));
  }
  argument_.origin_program_desc.reset(
      new ProgramDesc(*inference_program_->Proto()));
112 113 114 115
  PADDLE_ENFORCE(config_.ir_mode == AnalysisConfig::IrPassMode::kExclude,
                 "Only kExclude is supported yet.");
  Analyzer().DisableIrPasses(config_.ir_passes).Run(&argument_);

Y
Yan Chunwei 已提交
116 117 118 119
  CHECK(argument_.transformed_program_desc);
  VLOG(5) << "to prepare executor";
  inference_program_.reset(
      new framework::ProgramDesc(*argument_.transformed_program_desc));
120 121 122 123 124 125
  if (argument_.Has(framework::ir::kParamScopeAttr)) {
    // Update scope.
    scope_.reset(
        argument_.Release<framework::Scope>(framework::ir::kParamScopeAttr));
  }
  LOG(INFO) << "== optimize end ==";
Y
Yan Chunwei 已提交
126
}
127 128 129

template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
130 131
    AnalysisConfig, PaddleEngineKind::kAnalysis>(const AnalysisConfig& config) {
  VLOG(3) << "create AnalysisConfig";
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157
  if (config.use_gpu) {
    // 1. GPU memeroy
    PADDLE_ENFORCE_GT(
        config.fraction_of_gpu_memory, 0.f,
        "fraction_of_gpu_memory in the config should be set to range (0., 1.]");
    PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device);
    std::vector<std::string> flags;
    if (config.fraction_of_gpu_memory >= 0.0f ||
        config.fraction_of_gpu_memory <= 0.95f) {
      flags.push_back("dummpy");
      std::string flag = "--fraction_of_gpu_memory_to_use=" +
                         std::to_string(config.fraction_of_gpu_memory);
      flags.push_back(flag);
      VLOG(3) << "set flag: " << flag;
      framework::InitGflags(flags);
    }
  }

  std::unique_ptr<PaddlePredictor> predictor(new AnalysisPredictor(config));
  if (!dynamic_cast<AnalysisPredictor*>(predictor.get())->Init(nullptr)) {
    return nullptr;
  }
  return predictor;
}

}  // namespace paddle