analysis_predictor.cc 5.2 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>
Y
Yan Chunwei 已提交
17
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
18 19 20 21 22 23 24
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/inference/utils/singleton.h"

namespace paddle {

Y
Yan Chunwei 已提交
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
bool AnalysisPredictor::Init(
    const std::shared_ptr<framework::Scope>& parent_scope) {
  VLOG(3) << "Predictor::init()";
  if (config_.use_gpu) {
    place_ = paddle::platform::CUDAPlace(config_.device);
  } else {
    place_ = paddle::platform::CPUPlace();
  }
  PADDLE_ENFORCE(!parent_scope);
  if (parent_scope) {
    scope_ = parent_scope;
    sub_scope_ = &(parent_scope->NewScope());
  } else {
    paddle::framework::InitDevices(false);
    scope_.reset(new paddle::framework::Scope());
  }
41

Y
Yan Chunwei 已提交
42
  executor_.reset(new paddle::framework::Executor(place_));
43

Y
Yan Chunwei 已提交
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58
  // 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;
59 60
  }

Y
Yan Chunwei 已提交
61 62
  OptimizeInferenceProgram();
  ctx_ = executor_->Prepare(*inference_program_, 0);
63

Y
Yan Chunwei 已提交
64 65 66 67 68 69 70 71
  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;
}
72

Y
Yan Chunwei 已提交
73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104
void AnalysisPredictor::OptimizeInferenceProgram() {
  LOG(INFO) << "optimize begin";
  FLAGS_IA_enable_ir = true;
  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()));
  Analyzer().Run(&argument_);
  CHECK(argument_.transformed_program_desc);
  VLOG(5) << "to prepare executor";
  // LOG(INFO) << "transformed_parogram_desc " <<
  // argument.transformed_program_desc->DebugString();
  inference_program_.reset(
      new framework::ProgramDesc(*argument_.transformed_program_desc));
  PADDLE_ENFORCE(argument_.Has(framework::ir::kParamScopeAttr));
  // Update scope.
  scope_.reset(
      argument_.Release<framework::Scope>(framework::ir::kParamScopeAttr));
  LOG(INFO) << "optimize end ==";
}
105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137

template <>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
    NativeConfig, PaddleEngineKind::kAnalysis>(const NativeConfig& config) {
  VLOG(3) << "create NativePredictor";
  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

USE_PASS(fc_fuse_pass);
L
luotao1 已提交
138 139
USE_PASS(seq_concat_fc_fuse_pass);
USE_PASS(fc_lstm_fuse_pass);
140 141
USE_PASS(graph_viz_pass);
USE_PASS(infer_clean_graph_pass);
L
luotao1 已提交
142
USE_PASS(attention_lstm_fuse_pass);