hogwild_worker.cc 4.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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/device_worker.h"
16
#include "paddle/fluid/framework/device_worker_factory.h"
17
#include "paddle/fluid/platform/cpu_helper.h"
D
dongdaxiang 已提交
18
#include "paddle/fluid/platform/lodtensor_printer.h"
19 20 21 22

namespace paddle {
namespace framework {

D
dongdaxiang 已提交
23 24 25 26 27 28 29 30
void HogwildWorker::Initialize(const TrainerDesc& desc) {
  fetch_var_names_.resize(desc.fetch_var_names_size());
  for (size_t i = 0; i < desc.fetch_var_names_size(); ++i) {
    fetch_var_names_[i] = desc.fetch_var_names(i);
  }
  batch_cnt_per_print_ = static_cast<int>(desc.batch_per_print());
}

31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
void HogwildWorker::CreateThreadOperators(const ProgramDesc& program) {
  auto& block = program.Block(0);
  op_names_.clear();
  for (auto& op_desc : block.AllOps()) {
    std::unique_ptr<OperatorBase> local_op = OpRegistry::CreateOp(*op_desc);
    op_names_.push_back(op_desc->Type());
    OperatorBase* local_op_ptr = local_op.release();
    ops_.push_back(local_op_ptr);
    continue;
  }
}

void HogwildWorker::CreateThreadScope(const ProgramDesc& program) {
  auto& block = program.Block(0);

  PADDLE_ENFORCE_NOT_NULL(
      root_scope_, "root_scope should be set before creating thread scope");

  thread_scope_ = &root_scope_->NewScope();
  for (auto& var : block.AllVars()) {
    if (var->Persistable()) {
      auto* ptr = root_scope_->Var(var->Name());
      InitializeVariable(ptr, var->GetType());
    } else {
      auto* ptr = thread_scope_->Var(var->Name());
      InitializeVariable(ptr, var->GetType());
    }
  }
}

void HogwildWorker::BindingDataFeedMemory() {
  const std::vector<std::string>& input_feed =
63
      device_reader_->GetUseSlotAlias();
64
  for (auto name : input_feed) {
65
    device_reader_->AddFeedVar(thread_scope_->Var(name), name);
66 67 68 69 70 71 72 73 74 75
  }
}

void HogwildWorker::CreateDeviceResource(const ProgramDesc& main_prog) {
  CreateThreadScope(main_prog);
  CreateThreadOperators(main_prog);
}

void HogwildWorker::TrainFilesWithProfiler() {
  platform::SetNumThreads(1);
76
  device_reader_->Start();
77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
  std::vector<double> op_total_time;
  std::vector<std::string> op_name;
  for (auto& op : ops_) {
    op_name.push_back(op->Type());
  }
  op_total_time.resize(ops_.size());
  for (size_t i = 0; i < op_total_time.size(); ++i) {
    op_total_time[i] = 0.0;
  }
  platform::Timer timeline;
  double total_time = 0.0;
  double read_time = 0.0;
  int cur_batch;
  int batch_cnt = 0;
  timeline.Start();
D
dongdaxiang 已提交
92
  uint64_t total_inst = 0;
93
  while ((cur_batch = device_reader_->Next()) > 0) {
94
    VLOG(3) << "read a batch in thread " << thread_id_;
95 96 97 98 99 100 101 102 103 104
    timeline.Pause();
    read_time += timeline.ElapsedSec();
    total_time += timeline.ElapsedSec();
    for (size_t i = 0; i < ops_.size(); ++i) {
      timeline.Start();
      ops_[i]->Run(*thread_scope_, place_);
      timeline.Pause();
      op_total_time[i] += timeline.ElapsedSec();
      total_time += timeline.ElapsedSec();
    }
D
dongdaxiang 已提交
105
    total_inst += cur_batch;
106 107 108 109 110 111 112 113 114
    ++batch_cnt;
    thread_scope_->DropKids();
    if (thread_id_ == 0) {
      if (batch_cnt > 0 && batch_cnt % 100 == 0) {
        for (size_t i = 0; i < ops_.size(); ++i) {
          fprintf(stderr, "op_name:[%zu][%s], op_mean_time:[%fs]\n", i,
                  op_name[i].c_str(), op_total_time[i] / batch_cnt);
        }
        fprintf(stderr, "mean read time: %fs\n", read_time / batch_cnt);
D
dongdaxiang 已提交
115
        fprintf(stderr, "IO percent: %f\n", read_time / total_time * 100);
D
dongdaxiang 已提交
116
        fprintf(stderr, "%6.2f instances/s\n", total_inst / total_time);
117 118 119 120 121 122 123 124 125 126
      }
    }
    timeline.Start();
  }
}

void HogwildWorker::TrainFiles() {
  platform::SetNumThreads(1);

  // how to accumulate fetched values here
127
  device_reader_->Start();
128 129
  int cur_batch;
  int batch_cnt = 0;
130
  while ((cur_batch = device_reader_->Next()) > 0) {
131 132 133 134 135 136 137 138 139
    for (auto& op : ops_) {
      op->Run(*thread_scope_, place_);
    }

    ++batch_cnt;
    thread_scope_->DropKids();
  }
}

D
dongdaxiang 已提交
140 141 142 143 144 145 146 147 148 149 150
void HogwildWorker::PrintFetchVars(int batch_cnt) {
  if (thread_id_ == 0) {
    if (batch_cnt > 0 && batch_cnt % batch_cnt_per_print_ == 0) {
      int fetch_var_num = fetch_var_names_.size();
      for (int i = 0; i < fetch_var_num; ++i) {
        platform::PrintVar(thread_scope_, fetch_var_names_[i], "None");
      }
    }
  }
}

151 152
}  // end namespace framework
}  // end namespace paddle