data_feed.cc 30.2 KB
Newer Older
W
Wang Guibao 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15
#include "paddle/fluid/framework/data_feed.h"
D
dongdaxiang 已提交
16
#include <stdio_ext.h>
17
#include <utility>
18
#include "gflags/gflags.h"
W
Wang Guibao 已提交
19 20 21
#include "google/protobuf/io/zero_copy_stream_impl.h"
#include "google/protobuf/message.h"
#include "google/protobuf/text_format.h"
22 23
#include "io/fs.h"
#include "io/shell.h"
W
Wang Guibao 已提交
24 25
#include "paddle/fluid/framework/feed_fetch_method.h"
#include "paddle/fluid/framework/feed_fetch_type.h"
26
#include "paddle/fluid/platform/timer.h"
W
Wang Guibao 已提交
27 28 29 30 31 32 33 34

namespace paddle {
namespace framework {

void DataFeed::AddFeedVar(Variable* var, const std::string& name) {
  CheckInit();
  for (size_t i = 0; i < use_slots_.size(); ++i) {
    if (name == use_slots_[i]) {
35
      feed_vec_[i] = var->GetMutable<LoDTensor>();
W
Wang Guibao 已提交
36 37 38 39 40
    }
  }
}

bool DataFeed::SetFileList(const std::vector<std::string>& files) {
41
  std::unique_lock<std::mutex> lock(*mutex_for_pick_file_);
W
Wang Guibao 已提交
42
  CheckInit();
43 44 45
  // Do not set finish_set_filelist_ flag,
  // since a user may set file many times after init reader
  /*
W
Wang Guibao 已提交
46 47 48 49
  if (finish_set_filelist_) {
    VLOG(3) << "info: you have set the filelist.";
    return false;
  }
50
  */
51
  //PADDLE_ENFORCE(files.size(), "You have set an empty filelist.");
W
Wang Guibao 已提交
52 53 54 55 56 57 58 59 60 61 62 63
  filelist_.assign(files.begin(), files.end());

  finish_set_filelist_ = true;
  return true;
}

void DataFeed::SetBatchSize(int batch_size) {
  PADDLE_ENFORCE(batch_size > 0, "Illegal batch size: %d.", batch_size);
  default_batch_size_ = batch_size;
}

bool DataFeed::PickOneFile(std::string* filename) {
64 65 66 67 68 69
  PADDLE_ENFORCE(mutex_for_pick_file_ != nullptr,
                 "should call SetFileListMutex before PickOneFile");
  PADDLE_ENFORCE(file_idx_ != nullptr,
                 "should call SetFileListIndex before PickOneFile");
  std::unique_lock<std::mutex> lock(*mutex_for_pick_file_);
  if (*file_idx_ == filelist_.size()) {
70
    VLOG(3) << "DataFeed::PickOneFile no more file to pick";
W
Wang Guibao 已提交
71 72
    return false;
  }
73 74
  VLOG(3) << "file_idx_=" << *file_idx_;
  *filename = filelist_[(*file_idx_)++];
D
dongdaxiang 已提交
75
  // LOG(ERROR) << "pick file:" << *filename;
W
Wang Guibao 已提交
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
  return true;
}

void DataFeed::CheckInit() {
  PADDLE_ENFORCE(finish_init_, "Initialization did not succeed.");
}

void DataFeed::CheckSetFileList() {
  PADDLE_ENFORCE(finish_set_filelist_, "Set filelist did not succeed.");
}

void DataFeed::CheckStart() {
  PADDLE_ENFORCE(finish_start_, "Datafeed has not started running yet.");
}

template <typename T>
void PrivateQueueDataFeed<T>::SetQueueSize(int queue_size) {
  PADDLE_ENFORCE(queue_size > 0, "Illegal queue size: %d.", queue_size);
  queue_size_ = queue_size;
  queue_ = std::unique_ptr<paddle::operators::reader::BlockingQueue<T>>(
      new paddle::operators::reader::BlockingQueue<T>(queue_size_));
}

template <typename T>
bool PrivateQueueDataFeed<T>::Start() {
  CheckSetFileList();
102 103
  read_thread_ = std::thread(&PrivateQueueDataFeed::ReadThread, this);
  read_thread_.detach();
W
Wang Guibao 已提交
104 105 106 107 108 109 110

  finish_start_ = true;
  return true;
}

template <typename T>
void PrivateQueueDataFeed<T>::ReadThread() {
111 112 113 114 115 116 117 118 119 120
  std::string filename;
  while (PickOneFile(&filename)) {
    int err_no = 0;
    fp_ = fs_open_read(filename, &err_no, pipe_command_);
    __fsetlocking(&*fp_, FSETLOCKING_BYCALLER);
    thread_local string::LineFileReader reader;
    T instance;
    while (ParseOneInstanceFromPipe(&instance)) {
      queue_->Send(instance);
    }
W
Wang Guibao 已提交
121
  }
122
  queue_->Close();
W
Wang Guibao 已提交
123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
}

template <typename T>
int PrivateQueueDataFeed<T>::Next() {
  CheckStart();
  int index = 0;
  T instance;
  T ins_vec;
  while (index < default_batch_size_) {
    if (!queue_->Receive(&instance)) {
      break;
    }
    AddInstanceToInsVec(&ins_vec, instance, index++);
  }
  batch_size_ = index;
  if (batch_size_ != 0) {
    PutToFeedVec(ins_vec);
  }
  return batch_size_;
}

144
// explicit instantiation
W
Wang Guibao 已提交
145 146
template class PrivateQueueDataFeed<std::vector<MultiSlotType>>;

147 148 149
template <typename T>
InMemoryDataFeed<T>::InMemoryDataFeed() {
  cur_channel_ = 0;
150 151
  shuffled_ins_ = std::make_shared<paddle::framework::BlockingQueue<T>>();
  shuffled_ins_out_ = std::make_shared<paddle::framework::BlockingQueue<T>>();
152 153 154 155 156
  fleet_send_batch_size_ = 80000;
  memory_data_ = nullptr;
  mutex_for_update_memory_data_ = nullptr;
  this->file_idx_ = nullptr;
  this->mutex_for_pick_file_ = nullptr;
157 158 159 160 161
}

template <typename T>
bool InMemoryDataFeed<T>::Start() {
  DataFeed::CheckSetFileList();
162 163
  if (shuffled_ins_->Size() == 0 && shuffled_ins_out_->Size() == 0) {
    FillMemoryDataToChannel();
164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
  }
  DataFeed::finish_start_ = true;
  return true;
}

template <typename T>
int InMemoryDataFeed<T>::Next() {
  DataFeed::CheckStart();
  std::shared_ptr<paddle::framework::BlockingQueue<T>> in_channel = nullptr;
  std::shared_ptr<paddle::framework::BlockingQueue<T>> out_channel = nullptr;
  if (cur_channel_ == 0) {
    in_channel = shuffled_ins_;
    out_channel = shuffled_ins_out_;
  } else {
    in_channel = shuffled_ins_out_;
    out_channel = shuffled_ins_;
  }
  CHECK(in_channel != nullptr);
  CHECK(out_channel != nullptr);
X
xujiaqi01 已提交
183 184 185
  VLOG(3) << "in_channel size=" << in_channel->Size()
          << ", out_channel size=" << out_channel->Size()
          << ", thread_id=" << thread_id_;
186
  int index = 0;
D
dongdaxiang 已提交
187 188 189 190 191
  T instance;
  T ins_vec;
  while (index < DataFeed::default_batch_size_) {
    if (in_channel->Size() == 0) {
      break;
192
    }
D
dongdaxiang 已提交
193 194 195 196 197
    in_channel->Pop(instance);
    AddInstanceToInsVec(&ins_vec, instance, index++);
    out_channel->Push(std::move(instance));
  }
  DataFeed::batch_size_ = index;
198 199
  VLOG(3) << "batch_size_=" << DataFeed::batch_size_
          << ", thread_id=" << thread_id_;
D
dongdaxiang 已提交
200 201 202 203 204 205
  if (DataFeed::batch_size_ != 0) {
    PutToFeedVec(ins_vec);
  } else {
    cur_channel_ = 1 - cur_channel_;
  }
  return DataFeed::batch_size_;
206 207
}

208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
template <typename T>
void InMemoryDataFeed<T>::SetMemoryData(void* memory_data) {
  memory_data_ = static_cast<std::vector<T>*>(memory_data);
}

template <typename T>
void InMemoryDataFeed<T>::SetMemoryDataMutex(std::mutex* mutex) {
  mutex_for_update_memory_data_ = mutex;
}

template <typename T>
void InMemoryDataFeed<T>::SetThreadId(int thread_id) {
  thread_id_ = thread_id;
}

template <typename T>
void InMemoryDataFeed<T>::SetThreadNum(int thread_num) {
  thread_num_ = thread_num;
}

template <typename T>
void InMemoryDataFeed<T>::SetTrainerNum(int trainer_num) {
  trainer_num_ = trainer_num;
}

233 234
template <typename T>
void InMemoryDataFeed<T>::PutInsToChannel(const std::string& ins_str) {
235
  std::vector<T> ins;
X
xujiaqi01 已提交
236
  DeserializeIns(&ins, ins_str);
237 238 239 240
  shuffled_ins_->Extend(std::move(ins));
  VLOG(3) << "PutInsToChannel put ins num=" << ins.size()
          << " to channel, channel size=" << shuffled_ins_->Size()
          << " thread_id=" << thread_id_;
241 242
}

243 244
template <typename T>
void InMemoryDataFeed<T>::FillMemoryDataToChannel() {
X
xujiaqi01 已提交
245
  VLOG(3) << "FillMemoryDataToChannel, thread_id=" << thread_id_;
246 247 248 249 250
  auto interval = GetMemoryDataInterval();
  VLOG(3) << "memory data size=" << memory_data_->size()
          << ", fill data from  [" << interval.first << ", "
          << interval.second << "), thread_id=" << thread_id_;
  for (int64_t i = interval.first; i < interval.second; ++i) {
251 252 253 254 255 256 257
    T& t = (*memory_data_)[i];
    shuffled_ins_->Push(std::move(t));
  }
}

template <typename T>
void InMemoryDataFeed<T>::FillChannelToMemoryData() {
X
xujiaqi01 已提交
258
  VLOG(3) << "FillChannelToMemoryData, thread_id=" << thread_id_;
259 260
  std::vector<T> local_vec;
  std::shared_ptr<paddle::framework::BlockingQueue<T>> channel = nullptr;
261
  std::shared_ptr<paddle::framework::BlockingQueue<T>> pre_channel = nullptr;
262 263
  if (cur_channel_ == 0) {
    channel = shuffled_ins_;
264
    pre_channel = shuffled_ins_out_;
265 266
  } else {
    channel = shuffled_ins_out_;
267
    pre_channel = shuffled_ins_;
268 269
  }
  CHECK(channel != nullptr);
270 271
  CHECK(pre_channel != nullptr);
  CHECK(pre_channel->Size() == 0);
X
xujiaqi01 已提交
272
  local_vec.resize(channel->Size());
273
  for (int64_t i = 0; i < local_vec.size(); ++i) {
274 275
    channel->Pop(local_vec[i]);
  }
X
xujiaqi01 已提交
276 277 278 279 280 281 282 283 284
  VLOG(3) << "local_vec size=" << local_vec.size() <<", thread_id=" << thread_id_;
  {
    std::lock_guard<std::mutex> g(*mutex_for_update_memory_data_);
    VLOG(3) << "before insert, memory_data_ size=" << memory_data_->size()
            << ", thread_id=" << thread_id_;
    memory_data_->insert(memory_data_->end(), local_vec.begin(), local_vec.end());
    VLOG(3) << "after insert memory_data_ size=" << memory_data_->size()
            << ", thread_id=" << thread_id_;
  }
285 286 287
  std::vector<T>().swap(local_vec);
}

288 289
template <typename T>
void InMemoryDataFeed<T>::LoadIntoMemory() {
X
xujiaqi01 已提交
290
  VLOG(3) << "LoadIntoMemory() begin, thread_id=" << thread_id_;
291 292 293
  std::vector<T> local_vec;
  std::string filename;
  while (DataFeed::PickOneFile(&filename)) {
X
xujiaqi01 已提交
294 295
    VLOG(3) << "PickOneFile, filename=" << filename
            << ", thread_id=" << thread_id_;
296
    int err_no = 0;
D
dongdaxiang 已提交
297 298
    PrivateQueueDataFeed<T>::fp_ =
        fs_open_read(filename, &err_no, PrivateQueueDataFeed<T>::pipe_command_);
299
    CHECK(PrivateQueueDataFeed<T>::fp_ != nullptr);
300 301
    __fsetlocking(&*PrivateQueueDataFeed<T>::fp_, FSETLOCKING_BYCALLER);
    T instance;
302 303
    platform::Timer timeline;
    timeline.Start();
D
dongdaxiang 已提交
304
    while (ParseOneInstanceFromPipe(&instance)) {
305 306
      local_vec.push_back(instance);
    }
307
    timeline.Pause();
X
xujiaqi01 已提交
308
    VLOG(3) << "LoadIntoMemory() read all lines, file="
309 310
            << filename << ", cost time=" << timeline.ElapsedSec()
            << " seconds, thread_id=" << thread_id_;
311 312
    {
      std::lock_guard<std::mutex> lock(*mutex_for_update_memory_data_);
313
      timeline.Start();
X
xujiaqi01 已提交
314
      memory_data_->insert(memory_data_->end(),
315 316 317 318 319 320
                           std::make_move_iterator(local_vec.begin()),
                           std::make_move_iterator(local_vec.end()));
      timeline.Pause();
      VLOG(3) << "LoadIntoMemory() memory_data insert, cost time="
              << timeline.ElapsedSec() << " seconds, thread_id="
              << thread_id_;
321
    }
322
    local_vec.clear();
323
  }
324
  std::vector<T>().swap(local_vec);
X
xujiaqi01 已提交
325
  VLOG(3) << "LoadIntoMemory() end, thread_id=" << thread_id_;
326 327 328 329
}

template <typename T>
void InMemoryDataFeed<T>::LocalShuffle() {
X
xujiaqi01 已提交
330
  VLOG(3) << "LocalShuffle() begin, thread_id=" << thread_id_;
331
  FillMemoryDataToChannel();
X
xujiaqi01 已提交
332
  VLOG(3) << "LocalShuffle() end, thread_id=" << thread_id_;
333 334
}

335
template <typename T>
336
void InMemoryDataFeed<T>::GlobalShuffle() {
337
  VLOG(3) << "GlobalShuffle() begin, thread_id=" << thread_id_;
338
  auto fleet_ptr = FleetWrapper::GetInstance();
339 340 341 342 343 344 345 346 347 348
  std::vector<std::vector<T*>> send_vec(trainer_num_);
  for (auto& vec : send_vec) {
    vec.reserve(fleet_send_batch_size_);
  }
  std::vector<std::future<int32_t>> total_status;
  auto interval = GetMemoryDataInterval();
  VLOG(3) << "global shuffle data from  [" << interval.first << ", "
          << interval.second << "), thread_id=" << thread_id_;
  for (int64_t i = interval.first; i < interval.second; ++i) {
    // if get ins id, can also use hash
X
xujiaqi01 已提交
349
    // std::string ins_id = memory_data_[i].ins_id;
350
    int64_t random_num = fleet_ptr->LocalRandomEngine()();
X
xujiaqi01 已提交
351
    int64_t node_id = random_num % trainer_num_;
352
    send_vec[node_id].push_back(&((*memory_data_)[i]));
353
    if (i % fleet_send_batch_size_ == 0 && i != 0) {
354 355 356 357 358 359 360 361 362 363
      for (int j = 0; j < send_vec.size(); ++j) {
        std::string send_str;
        SerializeIns(send_vec[j], &send_str);
        VLOG(3) << "send str_length=" << send_str.length()
                << ", ins num=" << send_vec[j].size() << " to node_id="
                << j << ", thread_id=" << thread_id_;
        auto ret = fleet_ptr->SendClientToClientMsg(0, j, send_str);
        VLOG(3) << "end send, thread_id=" << thread_id_;
        send_vec[j].clear();
        total_status.push_back(std::move(ret));
364 365 366
      }
    }
  }
367 368 369 370 371 372 373 374 375
  for (int j = 0; j < send_vec.size(); ++j) {
    if (send_vec[j].size() != 0) {
      std::string send_str;
      SerializeIns(send_vec[j], &send_str);
      VLOG(3) << "send str_length=" << send_str.length()
              << " to node_id=" << j << ", thread_id=" << thread_id_;
      auto ret = fleet_ptr->SendClientToClientMsg(0, j, send_str);
      VLOG(3) << "end send, thread_id=" << thread_id_;
      total_status.push_back(std::move(ret));
376
    }
377 378 379 380
    std::vector<T*>().swap(send_vec[j]);
  }
  for (auto& t : total_status) {
    t.wait();
381
  }
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396
  VLOG(3) << "GlobalShuffle() end, thread_id=" << thread_id_;
}

template <typename T>
std::pair<int64_t, int64_t> InMemoryDataFeed<T>::GetMemoryDataInterval() {
  int64_t start = 0;
  int64_t end = 0;
  int64_t size = memory_data_->size();
  for (int64_t i = 0; i <= static_cast<int64_t>(thread_id_); ++i) {
    int64_t len = size / static_cast<int64_t>(thread_num_) +
                  (i < (size % static_cast<int64_t>(thread_num_)));
    start = end;
    end += len;
  }
  return std::make_pair(start, end);
397 398
}

399 400 401
// explicit instantiation
template class InMemoryDataFeed<std::vector<MultiSlotType>>;

W
Wang Guibao 已提交
402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430
void MultiSlotDataFeed::Init(
    const paddle::framework::DataFeedDesc& data_feed_desc) {
  finish_init_ = false;
  finish_set_filelist_ = false;
  finish_start_ = false;

  PADDLE_ENFORCE(data_feed_desc.has_multi_slot_desc(),
                 "Multi_slot_desc has not been set.");
  paddle::framework::MultiSlotDesc multi_slot_desc =
      data_feed_desc.multi_slot_desc();
  SetBatchSize(data_feed_desc.batch_size());
  SetQueueSize(data_feed_desc.batch_size());
  size_t all_slot_num = multi_slot_desc.slots_size();
  all_slots_.resize(all_slot_num);
  all_slots_type_.resize(all_slot_num);
  use_slots_index_.resize(all_slot_num);
  use_slots_.clear();
  use_slots_is_dense_.clear();
  for (size_t i = 0; i < all_slot_num; ++i) {
    const auto& slot = multi_slot_desc.slots(i);
    all_slots_[i] = slot.name();
    all_slots_type_[i] = slot.type();
    use_slots_index_[i] = slot.is_used() ? use_slots_.size() : -1;
    if (slot.is_used()) {
      use_slots_.push_back(all_slots_[i]);
      use_slots_is_dense_.push_back(slot.is_dense());
    }
  }
  feed_vec_.resize(use_slots_.size());
431
  pipe_command_ = data_feed_desc.pipe_command();
W
Wang Guibao 已提交
432 433 434
  finish_init_ = true;
}

D
dongdaxiang 已提交
435
void MultiSlotDataFeed::ReadThread() {
436 437 438 439
  std::string filename;
  while (PickOneFile(&filename)) {
    int err_no = 0;
    fp_ = fs_open_read(filename, &err_no, pipe_command_);
D
dongdaxiang 已提交
440
    CHECK(fp_ != nullptr);
441 442 443 444 445 446 447 448
    __fsetlocking(&*fp_, FSETLOCKING_BYCALLER);
    thread_local string::LineFileReader reader;
    std::vector<MultiSlotType> instance;
    int ins_num = 0;
    while (ParseOneInstanceFromPipe(&instance)) {
      ins_num++;
      queue_->Send(instance);
    }
D
dongdaxiang 已提交
449
    VLOG(3) << "filename: " << filename << " inst num: " << ins_num;
D
dongdaxiang 已提交
450
  }
451
  queue_->Close();
D
dongdaxiang 已提交
452 453
}

W
Wang Guibao 已提交
454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482
bool MultiSlotDataFeed::CheckFile(const char* filename) {
  CheckInit();  // get info of slots
  std::ifstream fin(filename);
  if (!fin.good()) {
    VLOG(1) << "error: open file<" << filename << "> fail";
    return false;
  }
  std::string line;
  int instance_cout = 0;
  std::string all_slots_alias = "";
  for (const auto& alias : all_slots_) {
    all_slots_alias += alias + " ";
  }
  std::string use_slots_alias = "";
  for (const auto& alias : use_slots_) {
    use_slots_alias += alias + " ";
  }
  VLOG(3) << "total slots num: " << all_slots_.size();
  VLOG(3) << "total slots alias: " << all_slots_alias;
  VLOG(3) << "used slots num: " << use_slots_.size();
  VLOG(3) << "used slots alias: " << use_slots_alias;
  while (getline(fin, line)) {
    ++instance_cout;
    const char* str = line.c_str();
    char* endptr = const_cast<char*>(str);
    int len = line.length();
    for (size_t i = 0; i < all_slots_.size(); ++i) {
      int num = strtol(endptr, &endptr, 10);
      if (num < 0) {
483 484
        VLOG(0) << "error: the number of ids is a negative number: " << num;
        VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
485 486 487
                << filename << ">";
        return false;
      } else if (num == 0) {
488
        VLOG(0)
W
Wang Guibao 已提交
489 490 491 492
            << "error: the number of ids can not be zero, you need "
               "padding it in data generator; or if there is something wrong"
               " with the data, please check if the data contains unresolvable "
               "characters.";
493
        VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
494 495 496
                << filename << ">";
        return false;
      } else if (errno == ERANGE || num > INT_MAX) {
497 498
        VLOG(0) << "error: the number of ids greater than INT_MAX";
        VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
499 500 501 502 503 504 505
                << filename << ">";
        return false;
      }
      if (all_slots_type_[i] == "float") {
        for (int i = 0; i < num; ++i) {
          strtof(endptr, &endptr);
          if (errno == ERANGE) {
506
            VLOG(0) << "error: the value is out of the range of "
W
Wang Guibao 已提交
507
                       "representable values for float";
508
            VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
509 510 511 512
                    << filename << ">";
            return false;
          }
          if (i + 1 != num && endptr - str == len) {
513 514
            VLOG(0) << "error: there is a wrong with the number of ids.";
            VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
515 516 517 518 519 520 521 522
                    << filename << ">";
            return false;
          }
        }
      } else if (all_slots_type_[i] == "uint64") {
        for (int i = 0; i < num; ++i) {
          strtoull(endptr, &endptr, 10);
          if (errno == ERANGE) {
523
            VLOG(0) << "error: the value is out of the range of "
W
Wang Guibao 已提交
524
                       "representable values for uint64_t";
525
            VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
526 527 528 529
                    << filename << ">";
            return false;
          }
          if (i + 1 != num && endptr - str == len) {
530 531
            VLOG(0) << "error: there is a wrong with the number of ids.";
            VLOG(0) << "please check line<" << instance_cout << "> in file<"
W
Wang Guibao 已提交
532 533 534 535 536
                    << filename << ">";
            return false;
          }
        }
      } else {
537
        VLOG(0) << "error: this type<" << all_slots_type_[i]
W
Wang Guibao 已提交
538 539 540 541
                << "> is not supported";
        return false;
      }
    }
542 543 544
    // It may be added '\t' character to the end of the output of reduce
    // task when processes data by Hadoop(when the output of the reduce
    // task of Hadoop has only one field, it will add a '\t' at the end
545 546 547 548 549
    // of the line by default, and you can use this option to avoid it:
    // `-D mapred.textoutputformat.ignoreseparator=true`), which does
    // not affect the correctness of the data. Therefore, it should be
    // judged that the data is not normal when the end of each line of
    // data contains characters which are not spaces.
550 551 552 553 554 555 556 557
    while (endptr - str != len) {
      if (!isspace(*(endptr++))) {
        VLOG(0)
            << "error: there is some extra characters at the end of the line.";
        VLOG(0) << "please check line<" << instance_cout << "> in file<"
                << filename << ">";
        return false;
      }
W
Wang Guibao 已提交
558 559 560 561 562 563 564
    }
  }
  VLOG(3) << "instances cout: " << instance_cout;
  VLOG(3) << "The file format is correct";
  return true;
}

D
dongdaxiang 已提交
565 566
bool MultiSlotDataFeed::ParseOneInstanceFromPipe(
    std::vector<MultiSlotType>* instance) {
567 568 569
  thread_local string::LineFileReader reader;

  if (!reader.getline(&*(fp_.get()))) {
D
dongdaxiang 已提交
570 571
    return false;
  } else {
572 573 574
    int use_slots_num = use_slots_.size();
    instance->resize(use_slots_num);

D
dongdaxiang 已提交
575 576
    const char* str = reader.get();
    std::string line = std::string(str);
577
    //VLOG(3) << line;
D
dongdaxiang 已提交
578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605
    char* endptr = const_cast<char*>(str);
    int pos = 0;
    for (size_t i = 0; i < use_slots_index_.size(); ++i) {
      int idx = use_slots_index_[i];
      int num = strtol(&str[pos], &endptr, 10);
      PADDLE_ENFORCE(
          num,
          "The number of ids can not be zero, you need padding "
          "it in data generator; or if there is something wrong with "
          "the data, please check if the data contains unresolvable "
          "characters.\nplease check this error line: %s",
          str);
      if (idx != -1) {
        (*instance)[idx].Init(all_slots_type_[i]);
        if ((*instance)[idx].GetType()[0] == 'f') {  // float
          for (int j = 0; j < num; ++j) {
            float feasign = strtof(endptr, &endptr);
            (*instance)[idx].AddValue(feasign);
          }
        } else if ((*instance)[idx].GetType()[0] == 'u') {  // uint64
          for (int j = 0; j < num; ++j) {
            uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10);
            (*instance)[idx].AddValue(feasign);
          }
        }
        pos = endptr - str;
      } else {
        for (int j = 0; j <= num; ++j) {
D
dongdaxiang 已提交
606 607 608 609
          // pos = line.find_first_of(' ', pos + 1);
          while (line[pos + 1] != ' ') {
            pos++;
          }
D
dongdaxiang 已提交
610 611 612 613 614 615 616
        }
      }
    }
    return true;
  }
}

W
Wang Guibao 已提交
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635
bool MultiSlotDataFeed::ParseOneInstance(std::vector<MultiSlotType>* instance) {
  std::string line;
  if (getline(file_, line)) {
    int use_slots_num = use_slots_.size();
    instance->resize(use_slots_num);
    // parse line
    const char* str = line.c_str();
    char* endptr = const_cast<char*>(str);
    int pos = 0;
    for (size_t i = 0; i < use_slots_index_.size(); ++i) {
      int idx = use_slots_index_[i];
      int num = strtol(&str[pos], &endptr, 10);
      PADDLE_ENFORCE(
          num,
          "The number of ids can not be zero, you need padding "
          "it in data generator; or if there is something wrong with "
          "the data, please check if the data contains unresolvable "
          "characters.\nplease check this error line: %s",
          str);
636

W
Wang Guibao 已提交
637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672
      if (idx != -1) {
        (*instance)[idx].Init(all_slots_type_[i]);
        if ((*instance)[idx].GetType()[0] == 'f') {  // float
          for (int j = 0; j < num; ++j) {
            float feasign = strtof(endptr, &endptr);
            (*instance)[idx].AddValue(feasign);
          }
        } else if ((*instance)[idx].GetType()[0] == 'u') {  // uint64
          for (int j = 0; j < num; ++j) {
            uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10);
            (*instance)[idx].AddValue(feasign);
          }
        }
        pos = endptr - str;
      } else {
        for (int j = 0; j <= num; ++j) {
          pos = line.find_first_of(' ', pos + 1);
        }
      }
    }
  } else {
    return false;
  }
  return true;
}

void MultiSlotDataFeed::AddInstanceToInsVec(
    std::vector<MultiSlotType>* ins_vec,
    const std::vector<MultiSlotType>& instance, int index) {
  if (index == 0) {
    ins_vec->resize(instance.size());
    for (size_t i = 0; i < instance.size(); ++i) {
      (*ins_vec)[i].Init(instance[i].GetType());
      (*ins_vec)[i].InitOffset();
    }
  }
673

W
Wang Guibao 已提交
674 675 676 677 678 679 680 681 682 683 684
  for (size_t i = 0; i < instance.size(); ++i) {
    (*ins_vec)[i].AddIns(instance[i]);
  }
}

void MultiSlotDataFeed::PutToFeedVec(
    const std::vector<MultiSlotType>& ins_vec) {
  for (size_t i = 0; i < use_slots_.size(); ++i) {
    const auto& type = ins_vec[i].GetType();
    const auto& offset = ins_vec[i].GetOffset();
    int total_instance = static_cast<int>(offset.back());
685

W
Wang Guibao 已提交
686 687
    if (type[0] == 'f') {  // float
      const auto& feasign = ins_vec[i].GetFloatData();
688 689 690
      float* tensor_ptr = feed_vec_[i]->mutable_data<float>(
          {total_instance, 1}, platform::CPUPlace());
      memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
W
Wang Guibao 已提交
691 692 693
    } else if (type[0] == 'u') {  // uint64
      // no uint64_t type in paddlepaddle
      const auto& feasign = ins_vec[i].GetUint64Data();
694 695 696 697
      int64_t* tensor_ptr = feed_vec_[i]->mutable_data<int64_t>(
          {total_instance, 1}, platform::CPUPlace());
      memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
    }
698

699 700 701 702 703
    LoD data_lod{offset};
    feed_vec_[i]->set_lod(data_lod);
    if (use_slots_is_dense_[i]) {
      int dim = total_instance / batch_size_;
      feed_vec_[i]->Resize({batch_size_, dim});
W
Wang Guibao 已提交
704 705 706 707
    }
  }
}

708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752
void MultiSlotInMemoryDataFeed::Init(
    const paddle::framework::DataFeedDesc& data_feed_desc) {
  finish_init_ = false;
  finish_set_filelist_ = false;
  finish_start_ = false;

  PADDLE_ENFORCE(data_feed_desc.has_multi_slot_desc(),
                 "Multi_slot_desc has not been set.");
  paddle::framework::MultiSlotDesc multi_slot_desc =
      data_feed_desc.multi_slot_desc();
  SetBatchSize(data_feed_desc.batch_size());
  SetQueueSize(data_feed_desc.batch_size());
  size_t all_slot_num = multi_slot_desc.slots_size();
  all_slots_.resize(all_slot_num);
  all_slots_type_.resize(all_slot_num);
  use_slots_index_.resize(all_slot_num);
  use_slots_.clear();
  use_slots_is_dense_.clear();
  for (size_t i = 0; i < all_slot_num; ++i) {
    const auto& slot = multi_slot_desc.slots(i);
    all_slots_[i] = slot.name();
    all_slots_type_[i] = slot.type();
    use_slots_index_[i] = slot.is_used() ? use_slots_.size() : -1;
    if (slot.is_used()) {
      use_slots_.push_back(all_slots_[i]);
      use_slots_is_dense_.push_back(slot.is_dense());
    }
  }
  feed_vec_.resize(use_slots_.size());
  pipe_command_ = data_feed_desc.pipe_command();
  finish_init_ = true;
}

bool MultiSlotInMemoryDataFeed::ParseOneInstanceFromPipe(
    std::vector<MultiSlotType>* instance) {
  thread_local string::LineFileReader reader;

  if (!reader.getline(&*(fp_.get()))) {
    return false;
  } else {
    int use_slots_num = use_slots_.size();
    instance->resize(use_slots_num);

    const char* str = reader.get();
    std::string line = std::string(str);
753
    //VLOG(3) << line;
754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792
    char* endptr = const_cast<char*>(str);
    int pos = 0;
    for (size_t i = 0; i < use_slots_index_.size(); ++i) {
      int idx = use_slots_index_[i];
      int num = strtol(&str[pos], &endptr, 10);
      PADDLE_ENFORCE(
          num,
          "The number of ids can not be zero, you need padding "
          "it in data generator; or if there is something wrong with "
          "the data, please check if the data contains unresolvable "
          "characters.\nplease check this error line: %s",
          str);
      if (idx != -1) {
        (*instance)[idx].Init(all_slots_type_[i]);
        if ((*instance)[idx].GetType()[0] == 'f') {  // float
          for (int j = 0; j < num; ++j) {
            float feasign = strtof(endptr, &endptr);
            (*instance)[idx].AddValue(feasign);
          }
        } else if ((*instance)[idx].GetType()[0] == 'u') {  // uint64
          for (int j = 0; j < num; ++j) {
            uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10);
            (*instance)[idx].AddValue(feasign);
          }
        }
        pos = endptr - str;
      } else {
        for (int j = 0; j <= num; ++j) {
          // pos = line.find_first_of(' ', pos + 1);
          while (line[pos + 1] != ' ') {
            pos++;
          }
        }
      }
    }
    return true;
  }
}

D
dongdaxiang 已提交
793 794
bool MultiSlotInMemoryDataFeed::ParseOneInstance(
    std::vector<MultiSlotType>* instance) {
795 796 797 798
  std::string line;
  if (getline(file_, line)) {
    int use_slots_num = use_slots_.size();
    instance->resize(use_slots_num);
799
    VLOG(3) << line;
800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886
    // parse line
    const char* str = line.c_str();
    char* endptr = const_cast<char*>(str);
    int pos = 0;
    for (size_t i = 0; i < use_slots_index_.size(); ++i) {
      int idx = use_slots_index_[i];
      int num = strtol(&str[pos], &endptr, 10);
      PADDLE_ENFORCE(
          num,
          "The number of ids can not be zero, you need padding "
          "it in data generator; or if there is something wrong with "
          "the data, please check if the data contains unresolvable "
          "characters.\nplease check this error line: %s",
          str);

      if (idx != -1) {
        (*instance)[idx].Init(all_slots_type_[i]);
        if ((*instance)[idx].GetType()[0] == 'f') {  // float
          for (int j = 0; j < num; ++j) {
            float feasign = strtof(endptr, &endptr);
            (*instance)[idx].AddValue(feasign);
          }
        } else if ((*instance)[idx].GetType()[0] == 'u') {  // uint64
          for (int j = 0; j < num; ++j) {
            uint64_t feasign = (uint64_t)strtoull(endptr, &endptr, 10);
            (*instance)[idx].AddValue(feasign);
          }
        }
        pos = endptr - str;
      } else {
        for (int j = 0; j <= num; ++j) {
          pos = line.find_first_of(' ', pos + 1);
        }
      }
    }
  } else {
    return false;
  }
  return true;
}

void MultiSlotInMemoryDataFeed::AddInstanceToInsVec(
    std::vector<MultiSlotType>* ins_vec,
    const std::vector<MultiSlotType>& instance, int index) {
  if (index == 0) {
    ins_vec->resize(instance.size());
    for (size_t i = 0; i < instance.size(); ++i) {
      (*ins_vec)[i].Init(instance[i].GetType());
      (*ins_vec)[i].InitOffset();
    }
  }

  for (size_t i = 0; i < instance.size(); ++i) {
    (*ins_vec)[i].AddIns(instance[i]);
  }
}

void MultiSlotInMemoryDataFeed::PutToFeedVec(
    const std::vector<MultiSlotType>& ins_vec) {
  for (size_t i = 0; i < use_slots_.size(); ++i) {
    const auto& type = ins_vec[i].GetType();
    const auto& offset = ins_vec[i].GetOffset();
    int total_instance = static_cast<int>(offset.back());

    if (type[0] == 'f') {  // float
      const auto& feasign = ins_vec[i].GetFloatData();
      float* tensor_ptr = feed_vec_[i]->mutable_data<float>(
          {total_instance, 1}, platform::CPUPlace());
      memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
    } else if (type[0] == 'u') {  // uint64
      // no uint64_t type in paddlepaddle
      const auto& feasign = ins_vec[i].GetUint64Data();
      int64_t* tensor_ptr = feed_vec_[i]->mutable_data<int64_t>(
          {total_instance, 1}, platform::CPUPlace());
      memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
    }

    LoD data_lod{offset};
    feed_vec_[i]->set_lod(data_lod);
    if (use_slots_is_dense_[i]) {
      int dim = total_instance / batch_size_;
      feed_vec_[i]->Resize({batch_size_, dim});
    }
  }
}

// todo serialize ins in global shuffle
D
dongdaxiang 已提交
887
void MultiSlotInMemoryDataFeed::SerializeIns(
888 889
    const std::vector<std::vector<MultiSlotType>*>& ins,
    std::string* str) {
890 891
  auto fleet_ptr = FleetWrapper::GetInstance();
  fleet_ptr->Serialize(ins, str);
892 893
}
// todo deserialize ins in global shuffle
894 895 896
void MultiSlotInMemoryDataFeed::DeserializeIns(
    std::vector<std::vector<MultiSlotType>>* ins,
    const std::string& str) {
897 898
  auto fleet_ptr = FleetWrapper::GetInstance();
  fleet_ptr->Deserialize(ins, str);
899 900
}

W
Wang Guibao 已提交
901 902
}  // namespace framework
}  // namespace paddle