async_executor.cc 11.7 KB
Newer Older
W
Wang Guibao 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
/* Copyright (c) 2016 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/async_executor.h"
#include "google/protobuf/io/zero_copy_stream_impl.h"
#include "google/protobuf/message.h"
#include "google/protobuf/text_format.h"

#include "gflags/gflags.h"
#include "paddle/fluid/framework/data_feed_factory.h"
#include "paddle/fluid/framework/executor_thread_worker.h"
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
#include "paddle/fluid/framework/lod_rank_table.h"
#include "paddle/fluid/framework/lod_tensor_array.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/reader.h"
#include "paddle/fluid/inference/io.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
H
pslib  
heqiaozhi 已提交
32
#include "pslib.h"
W
Wang Guibao 已提交
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50

namespace paddle {
namespace framework {
AsyncExecutor::AsyncExecutor(Scope* scope, const platform::Place& place)
    : root_scope_(scope), place_(place) {}

void AsyncExecutor::CreateThreads(
    ExecutorThreadWorker* worker, const ProgramDesc& main_program,
    const std::shared_ptr<DataFeed>& reader,
    const std::vector<std::string>& fetch_var_names, Scope* root_scope,
    const int thread_index, const bool debug) {
  worker->SetThreadId(thread_index);
  worker->SetDebug(debug);
  worker->SetRootScope(root_scope);
  worker->CreateThreadResource(main_program, place_);
  worker->SetDataFeed(reader);
  worker->SetFetchVarNames(fetch_var_names);
  worker->BindingDataFeedMemory();
51 52 53 54
  worker->SetPSlibPtr(_pslib_ptr);
  worker->SetPullDenseThread(_pull_dense_thread);
  worker->BindingSlotVariableMemory();
  worker->SetParamConfig(&_param_config);
W
Wang Guibao 已提交
55 56 57 58 59 60 61 62 63 64 65 66 67
}

void PrepareReaders(std::vector<std::shared_ptr<DataFeed>>& readers,  // NOLINT
                    const int thread_num, const DataFeedDesc& data_feed_desc,
                    const std::vector<std::string>& filelist) {
  readers.resize(thread_num);
  for (size_t i = 0; i < readers.size(); ++i) {
    readers[i] = DataFeedFactory::CreateDataFeed(data_feed_desc.name());
    readers[i]->Init(data_feed_desc);  // set batch_size and queue_size here
  }
  readers[0]->SetFileList(filelist);
}

H
heqiaozhi 已提交
68
void AsyncExecutor::InitServer(const std::string& dist_desc, int index) {
69
    _pslib_ptr = std::shared_ptr<paddle::distributed::PSlib>(new paddle::distributed::PSlib());
H
heqiaozhi 已提交
70 71 72
    _pslib_ptr->init_server(dist_desc, index);//TODO done

    InitParamConfig();
73 74
}

H
heqiaozhi 已提交
75 76 77 78
void AsyncExecutor::InitWorker(const std::string& dist_desc, std::vector<uint64_t>& host_sign_list, int node_num, int index) {
    _pslib_ptr = std::shared_ptr<paddle::distributed::PSlib>(new paddle::distributed::PSlib());
    _pslib_ptr->init_worker(dist_desc, host_sign_list.data(), node_num, index);//TODO done

H
heqiaozhi 已提交
79
    InitParamConfig();
H
heqiaozhi 已提交
80 81 82 83 84 85
}

uint64_t AsyncExecutor::StartServer() {
    return _pslib_ptr->run_server();
}

H
heqiaozhi 已提交
86 87 88 89
void AsyncExecutor::StopServer() {
    _pslib_ptr->stop_server();
}

H
heqiaozhi 已提交
90 91
void AsyncExecutor::GatherServers(std::vector<uint64_t>& host_sign_list, int node_num) {
    _pslib_ptr->gather_servers(host_sign_list.data(), node_num);
92 93
}

H
heqiaozhi 已提交
94
void AsyncExecutor::InitParamConfig() {
H
heqiaozhi 已提交
95 96 97 98 99 100
    for (int i = 0; i < _pslib_ptr->get_param()->server_param().downpour_server_param().downpour_table_param_size(); ++i) {
        if (_pslib_ptr->get_param()->server_param().downpour_server_param().downpour_table_param(i).table_class().find("SparseTable") != -1) {
            _param_config.fea_dim = _pslib_ptr->get_param()->server_param().downpour_server_param().downpour_table_param(i).accessor().fea_dim(); //TODO
            break;
        }
    }
H
heqiaozhi 已提交
101
    _param_config.slot_dim = _param_config.fea_dim - 2; //TODO
H
heqiaozhi 已提交
102 103 104 105 106 107
    _param_config.tmp_push_dense_wait_times = (int32_t)(_pslib_ptr->get_param()->trainer_param().push_dense_per_batch());
    _param_config.tmp_push_sparse_wait_times = (int32_t)(_pslib_ptr->get_param()->trainer_param().push_sparse_per_batch());

    for (auto t = 0u; t < _pslib_ptr->get_param()->trainer_param().skip_op_size(); ++t) {
        _param_config.skip_op.push_back(_pslib_ptr->get_param()->trainer_param().skip_op(t));
    }
H
heqiaozhi 已提交
108 109 110 111 112 113
    //sparse
    for (auto t = 0u; t < _pslib_ptr->get_param()->trainer_param().sparse_table_size(); ++t) {
        auto& table = _pslib_ptr->get_param()->trainer_param().sparse_table(t);
        std::vector<std::string> tmp_sparse_variable_name;
        for (int i = 0u; i < table.slot_value_size(); ++i) {
            tmp_sparse_variable_name.push_back(table.slot_value(i));
H
heqiaozhi 已提交
114
            _param_config.slot_alias_to_table[table.slot_key(i)] = table.table_id();
H
heqiaozhi 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
        }
        std::vector<std::string> tmp_sparse_gradient_variable_name;
        for (auto i = 0u; i < table.slot_gradient_size(); ++i) {
            tmp_sparse_gradient_variable_name.push_back(
                    table.slot_gradient(i));
        }
        _param_config.slot_input_vec[table.table_id()] = std::move(tmp_sparse_variable_name);
        _param_config.gradient_var[table.table_id()] = std::move(tmp_sparse_gradient_variable_name);
        _param_config.sparse_table_id.push_back(table.table_id());
    }
    //dense
    for (auto t = 0u; t < _pslib_ptr->get_param()->trainer_param().dense_table_size(); ++t) {
        auto& table = _pslib_ptr->get_param()->trainer_param().dense_table(t);
        std::vector<std::string> tmp_dense_variable_name;
        for (int i = 0u; i < table.dense_variable_name_size(); ++i) {
            tmp_dense_variable_name.push_back(table.dense_variable_name(i));
        }
        std::vector<std::string> tmp_dense_gradient_variable_name;
        for (auto i = 0u; i < table.dense_gradient_variable_name_size(); ++i) {
            tmp_dense_gradient_variable_name.push_back(
                    table.dense_gradient_variable_name(i));
        }
        _param_config.dense_variable_name[table.table_id()] = std::move(tmp_dense_variable_name);
        _param_config.dense_gradient_variable_name[table.table_id()] = std::move(tmp_dense_gradient_variable_name);
        _param_config.dense_table_id.push_back(table.table_id());
        _param_config.dense_table_size.push_back(table.fea_dim()); //TODO
    }
}

144 145
void AsyncExecutor::InitModel() {
    //TODO only rank = 0 do this
H
heqiaozhi 已提交
146 147 148
    //std::vector<int> all_dense_table_id; //TODO 
    //all_dense_table_id.push_back(0); //done
    for (auto table_id: _param_config.dense_table_id) {
149
        std::vector<paddle::ps::Region> regions;
H
heqiaozhi 已提交
150 151
        //std::vector<std::string> variables;  //TODO
        for (auto& t : _param_config.dense_variable_name[table_id]) {
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193
            Variable* var = root_scope_->FindVar(t);
            CHECK(var != nullptr) << "var[" << t << "] not found";
            LoDTensor* tensor = var->GetMutable<LoDTensor>();

            float* g = tensor->data<float>();
            CHECK(g != nullptr) << "var[" << t << "] value not initialized";

            float init_range = 0.2;
            int rown = tensor->dims()[0];
            init_range /= sqrt(rown);

            std::normal_distribution<float> ndistr(0.0, 1.0);
            for (auto i = 0u; i < tensor->numel(); ++i) {
                g[i] = ndistr(local_random_engine()) * init_range;
            }

            paddle::ps::Region reg(g, tensor->numel());
            regions.emplace_back(std::move(reg));
        }

        auto push_status = _pslib_ptr->_worker_ptr->push_dense_param(regions.data(), regions.size(), table_id);
        push_status.wait();
        auto status = push_status.get();
        if (status != 0) {
            LOG(FATAL) << "push dense param failed, status[" << status << "]";
            exit(-1);
        } 
    }
}

void AsyncExecutor::SaveModel(const std::string& path) {
    auto ret = _pslib_ptr->_worker_ptr->flush();
    ret.wait();
    ret = _pslib_ptr->_worker_ptr->save(path, 0);
    ret.wait();
    int32_t feasign_cnt = ret.get();
    if (feasign_cnt == -1) { // TODO should be feasign_cnt < 0, because server bug
        LOG(FATAL) << "save model failed";
        exit(-1);
    }
}

H
heqiaozhi 已提交
194 195 196 197 198 199 200 201 202 203 204 205 206
void AsyncExecutor::PrepareDenseThread(const std::string& mode) {
    if (mode == "mpi") {
        DensePullThreadParam param;
        param.ps_client = _pslib_ptr->_worker_ptr;;
        param.threshold = 1;//GlobalConfig::instance().pull_dense_per_batch; //TODO
        param.training_thread_num = actual_thread_num;
        param.root_scope = root_scope_;
        //param.dense_params = &GlobalConfig::instance().dense_variable_name; //TODO
        param.dense_params = &_param_config.dense_variable_name;

        _pull_dense_thread = std::shared_ptr<DensePullThread>(new DensePullThread(param));
        _pull_dense_thread->start();
    }
207 208
}

W
Wang Guibao 已提交
209 210 211 212 213
void AsyncExecutor::RunFromFile(const ProgramDesc& main_program,
                                const std::string& data_feed_desc_str,
                                const std::vector<std::string>& filelist,
                                const int thread_num,
                                const std::vector<std::string>& fetch_var_names,
H
heqiaozhi 已提交
214
                                const std::string& mode,
W
Wang Guibao 已提交
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
                                const bool debug) {
  std::vector<std::thread> threads;

  auto& block = main_program.Block(0);
  for (auto var_name : fetch_var_names) {
    auto var_desc = block.FindVar(var_name);
    auto shapes = var_desc->GetShape();
    PADDLE_ENFORCE(shapes[shapes.size() - 1] == 1,
                   "var %s: Fetched var has wrong shape, "
                   "only variables with the last dimension size 1 supported",
                   var_name);
  }

  DataFeedDesc data_feed_desc;
  google::protobuf::TextFormat::ParseFromString(data_feed_desc_str,
                                                &data_feed_desc);

232
  actual_thread_num = thread_num;
W
Wang Guibao 已提交
233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255
  int file_cnt = filelist.size();
  PADDLE_ENFORCE(file_cnt > 0, "File list cannot be empty");

  if (actual_thread_num > file_cnt) {
    VLOG(1) << "Thread num = " << thread_num << ", file num = " << file_cnt
            << ". Changing thread_num = " << file_cnt;
    actual_thread_num = file_cnt;
  }

  /*
    readerDesc: protobuf description for reader initlization
    argument: class_name, batch_size, use_slot, queue_size, buffer_size,
    padding_index

    reader:
    1) each thread has a reader, reader will read input data and
    put it into input queue
    2) each reader has a Next() iterface, that can fetch an instance
    from the input queue
   */
  // todo: should be factory method for creating datafeed
  std::vector<std::shared_ptr<DataFeed>> readers;
  PrepareReaders(readers, actual_thread_num, data_feed_desc, filelist);
H
heqiaozhi 已提交
256
  PrepareDenseThread(mode);
W
Wang Guibao 已提交
257 258 259
  std::vector<std::shared_ptr<ExecutorThreadWorker>> workers;
  workers.resize(actual_thread_num);
  for (auto& worker : workers) {
H
heqiaozhi 已提交
260 261 262 263 264
    if (mode == "mpi") {
        worker.reset(new AsyncExecutorThreadWorker);
    } else {
        worker.reset(new ExecutorThreadWorker);
    }
W
Wang Guibao 已提交
265 266 267 268 269 270 271 272
  }

  // prepare thread resource here
  for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
    CreateThreads(workers[thidx].get(), main_program, readers[thidx],
                  fetch_var_names, root_scope_, thidx, debug);
  }

H
heqiaozhi 已提交
273
  
W
Wang Guibao 已提交
274 275 276 277 278 279 280 281 282
  // start executing ops in multiple threads
  for (int thidx = 0; thidx < actual_thread_num; ++thidx) {
    threads.push_back(
        std::thread(&ExecutorThreadWorker::TrainFiles, workers[thidx].get()));
  }

  for (auto& th : threads) {
    th.join();
  }
H
heqiaozhi 已提交
283 284 285
  if (mode == "mpi") {
    _pull_dense_thread->stop();
  }
W
Wang Guibao 已提交
286 287 288 289 290 291 292
  root_scope_->DropKids();

  return;
}

}  // einit_modelnd namespace framework
}  // end namespace paddle