device_worker.cc 6.4 KB
Newer Older
D
dongdaxiang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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"

17 18 19 20
namespace pten {
class DenseTensor;
}  // namespace pten

D
dongdaxiang 已提交
21 22 23
namespace paddle {
namespace framework {

W
wanghuancoder 已提交
24 25
class Scope;

D
dongdaxiang 已提交
26 27
void DeviceWorker::SetRootScope(Scope* root_scope) { root_scope_ = root_scope; }

J
jiaqi 已提交
28
void DeviceWorker::SetDataFeed(DataFeed* data_feed) {
D
dongdaxiang 已提交
29 30 31
  device_reader_ = data_feed;
}

32
template <typename T>
33
std::string PrintLodTensorType(Tensor* tensor, int64_t start, int64_t end) {
34 35 36 37 38 39 40 41 42 43 44 45
  auto count = tensor->numel();
  if (start < 0 || end > count) {
    VLOG(3) << "access violation";
    return "access violation";
  }
  std::ostringstream os;
  for (int64_t i = start; i < end; i++) {
    os << ":" << tensor->data<T>()[i];
  }
  return os.str();
}

46
std::string PrintLodTensorIntType(Tensor* tensor, int64_t start, int64_t end) {
47 48 49 50 51 52 53 54 55 56 57 58
  auto count = tensor->numel();
  if (start < 0 || end > count) {
    VLOG(3) << "access violation";
    return "access violation";
  }
  std::ostringstream os;
  for (int64_t i = start; i < end; i++) {
    os << ":" << static_cast<uint64_t>(tensor->data<int64_t>()[i]);
  }
  return os.str();
}

59
std::string PrintLodTensor(Tensor* tensor, int64_t start, int64_t end) {
60 61 62 63 64 65 66 67 68 69 70 71 72 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
  std::string out_val;
  if (tensor->type() == proto::VarType::FP32) {
    out_val = PrintLodTensorType<float>(tensor, start, end);
  } else if (tensor->type() == proto::VarType::INT64) {
    out_val = PrintLodTensorIntType(tensor, start, end);
  } else if (tensor->type() == proto::VarType::FP64) {
    out_val = PrintLodTensorType<double>(tensor, start, end);
  } else {
    out_val = "unsupported type";
  }
  return out_val;
}

std::pair<int64_t, int64_t> GetTensorBound(LoDTensor* tensor, int index) {
  auto& dims = tensor->dims();
  if (tensor->lod().size() != 0) {
    auto& lod = tensor->lod()[0];
    return {lod[index] * dims[1], lod[index + 1] * dims[1]};
  } else {
    return {index * dims[1], (index + 1) * dims[1]};
  }
}

bool CheckValidOutput(LoDTensor* tensor, size_t batch_size) {
  auto& dims = tensor->dims();
  if (dims.size() != 2) return false;
  if (tensor->lod().size() != 0) {
    auto& lod = tensor->lod()[0];
    if (lod.size() != batch_size + 1) {
      return false;
    }
  } else {
    if (dims[0] != static_cast<int>(batch_size)) {
      return false;
    }
  }
  return true;
}

H
hutuxian 已提交
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118
void DeviceWorker::DumpParam(const Scope& scope, const int batch_id) {
  std::ostringstream os;
  for (auto& param : *dump_param_) {
    os.str("");
    Variable* var = scope.FindVar(param);
    if (var == nullptr) {
      continue;
    }
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
    framework::LoDTensor cpu_tensor;
    if (platform::is_gpu_place(tensor->place())) {
      TensorCopySync(*tensor, platform::CPUPlace(), &cpu_tensor);
      tensor = &cpu_tensor;
    }
    int64_t len = tensor->numel();
    os << "(" << batch_id << "," << param << ")"
       << PrintLodTensor(tensor, 0, len);
    writer_ << os.str();
  }
}
X
xujiaqi01 已提交
119

H
hutuxian 已提交
120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171
void DeviceWorker::InitRandomDumpConfig(const TrainerDesc& desc) {
  bool enable_random_dump = desc.enable_random_dump();
  if (!enable_random_dump) {
    dump_mode_ = 0;
  } else {
    if (desc.random_with_lineid()) {
      dump_mode_ = 1;
    } else {
      dump_mode_ = 2;
    }
  }
  dump_interval_ = desc.dump_interval();
}

void DeviceWorker::DumpField(const Scope& scope, int dump_mode,
                             int dump_interval) {  // dump_mode: 0: no random,
                                                   // 1: random with insid hash,
                                                   // 2: random with random
                                                   // number
  size_t batch_size = device_reader_->GetCurBatchSize();
  auto& ins_id_vec = device_reader_->GetInsIdVec();
  auto& ins_content_vec = device_reader_->GetInsContentVec();
  if (ins_id_vec.size() > 0) {
    batch_size = ins_id_vec.size();
  }
  std::vector<std::string> ars(batch_size);
  std::vector<bool> hit(batch_size, false);

  std::default_random_engine engine(0);
  std::uniform_int_distribution<size_t> dist(0U, INT_MAX);
  for (size_t i = 0; i < batch_size; i++) {
    size_t r = 0;
    if (dump_mode == 1) {
      r = XXH64(ins_id_vec[i].data(), ins_id_vec[i].length(), 0);
    } else if (dump_mode == 2) {
      r = dist(engine);
    }
    if (r % dump_interval != 0) {
      continue;
    }
    hit[i] = true;
  }
  for (size_t i = 0; i < ins_id_vec.size(); i++) {
    if (!hit[i]) {
      continue;
    }
    ars[i] += ins_id_vec[i];
    ars[i] = ars[i] + "\t" + ins_content_vec[i];
  }
  for (auto& field : *dump_fields_) {
    Variable* var = scope.FindVar(field);
    if (var == nullptr) {
172 173
      VLOG(0) << "Note: field[" << field
              << "] cannot be find in scope, so it was skipped.";
H
hutuxian 已提交
174 175 176
      continue;
    }
    LoDTensor* tensor = var->GetMutable<LoDTensor>();
177 178 179 180 181
    if (!tensor->IsInitialized()) {
      VLOG(0) << "Note: field[" << field
              << "] is not initialized, so it was skipped.";
      continue;
    }
H
hutuxian 已提交
182 183 184
    framework::LoDTensor cpu_tensor;
    if (platform::is_gpu_place(tensor->place())) {
      TensorCopySync(*tensor, platform::CPUPlace(), &cpu_tensor);
185
      cpu_tensor.set_lod(tensor->lod());
H
hutuxian 已提交
186 187 188
      tensor = &cpu_tensor;
    }
    if (!CheckValidOutput(tensor, batch_size)) {
189 190 191
      VLOG(0) << "Note: field[" << field << "] cannot pass check, so it was "
                                            "skipped. Maybe the dimension is "
                                            "wrong ";
H
hutuxian 已提交
192 193 194 195 196 197 198
      continue;
    }
    for (size_t i = 0; i < batch_size; ++i) {
      if (!hit[i]) {
        continue;
      }
      auto bound = GetTensorBound(tensor, i);
199 200
      ars[i] = ars[i] + "\t" + field + ":" +
               std::to_string(bound.second - bound.first);
H
hutuxian 已提交
201 202 203 204 205 206 207 208 209 210 211 212
      ars[i] += PrintLodTensor(tensor, bound.first, bound.second);
    }
  }
  // #pragma omp parallel for
  for (size_t i = 0; i < ars.size(); i++) {
    if (ars[i].length() == 0) {
      continue;
    }
    writer_ << ars[i];
  }
}

D
dongdaxiang 已提交
213 214
}  // namespace framework
}  // namespace paddle