analyzer_bert_tester.cc 7.4 KB
Newer Older
F
fuchang01 已提交
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.

15
#include "paddle/fluid/framework/transfer_scope_cache.h"
16
#include "paddle/fluid/inference/tests/api/tester_helper.h"
F
fuchang01 已提交
17 18 19 20 21 22 23 24 25 26 27 28 29 30 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 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132

namespace paddle {
namespace inference {

using paddle::PaddleTensor;

template <typename T>
void GetValueFromStream(std::stringstream *ss, T *t) {
  (*ss) >> (*t);
}

template <>
void GetValueFromStream<std::string>(std::stringstream *ss, std::string *t) {
  *t = ss->str();
}

// Split string to vector
template <typename T>
void Split(const std::string &line, char sep, std::vector<T> *v) {
  std::stringstream ss;
  T t;
  for (auto c : line) {
    if (c != sep) {
      ss << c;
    } else {
      GetValueFromStream<T>(&ss, &t);
      v->push_back(std::move(t));
      ss.str({});
      ss.clear();
    }
  }

  if (!ss.str().empty()) {
    GetValueFromStream<T>(&ss, &t);
    v->push_back(std::move(t));
    ss.str({});
    ss.clear();
  }
}

// Parse tensor from string
template <typename T>
bool ParseTensor(const std::string &field, paddle::PaddleTensor *tensor) {
  std::vector<std::string> data;
  Split(field, ':', &data);
  if (data.size() < 2) return false;

  std::string shape_str = data[0];

  std::vector<int> shape;
  Split(shape_str, ' ', &shape);

  std::string mat_str = data[1];

  std::vector<T> mat;
  Split(mat_str, ' ', &mat);

  tensor->shape = shape;
  auto size =
      std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>()) *
      sizeof(T);
  tensor->data.Resize(size);
  std::copy(mat.begin(), mat.end(), static_cast<T *>(tensor->data.data()));
  tensor->dtype = GetPaddleDType<T>();

  return true;
}

// Parse input tensors from string
bool ParseLine(const std::string &line,
               std::vector<paddle::PaddleTensor> *tensors) {
  std::vector<std::string> fields;
  Split(line, ';', &fields);

  if (fields.size() < 5) return false;

  tensors->clear();
  tensors->reserve(5);

  int i = 0;
  // src_id
  paddle::PaddleTensor src_id;
  ParseTensor<int64_t>(fields[i++], &src_id);
  tensors->push_back(src_id);

  // pos_id
  paddle::PaddleTensor pos_id;
  ParseTensor<int64_t>(fields[i++], &pos_id);
  tensors->push_back(pos_id);

  // segment_id
  paddle::PaddleTensor segment_id;
  ParseTensor<int64_t>(fields[i++], &segment_id);
  tensors->push_back(segment_id);

  // self_attention_bias
  paddle::PaddleTensor self_attention_bias;
  ParseTensor<float>(fields[i++], &self_attention_bias);
  tensors->push_back(self_attention_bias);

  // next_segment_index
  paddle::PaddleTensor next_segment_index;
  ParseTensor<int64_t>(fields[i++], &next_segment_index);
  tensors->push_back(next_segment_index);

  return true;
}

bool LoadInputData(std::vector<std::vector<paddle::PaddleTensor>> *inputs) {
  if (FLAGS_infer_data.empty()) {
    LOG(ERROR) << "please set input data path";
    return false;
  }

  std::ifstream fin(FLAGS_infer_data);
  std::string line;
133
  int sample = 0;
F
fuchang01 已提交
134

135
  // The unit-test dataset only have 10 samples, each sample have 5 feeds.
F
fuchang01 已提交
136 137
  while (std::getline(fin, line)) {
    std::vector<paddle::PaddleTensor> feed_data;
138 139 140 141
    ParseLine(line, &feed_data);
    inputs->push_back(std::move(feed_data));
    sample++;
    if (!FLAGS_test_all_data && sample == FLAGS_batch_size) break;
F
fuchang01 已提交
142
  }
143
  LOG(INFO) << "number of samples: " << sample;
F
fuchang01 已提交
144 145 146 147

  return true;
}

148
void SetConfig(AnalysisConfig *config) { config->SetModel(FLAGS_infer_model); }
F
fuchang01 已提交
149

150
void profile(bool use_mkldnn = false, bool use_ngraph = false) {
151
  AnalysisConfig config;
F
fuchang01 已提交
152 153 154 155
  SetConfig(&config);

  if (use_mkldnn) {
    config.EnableMKLDNN();
156
    config.pass_builder()->AppendPass("fc_mkldnn_pass");
F
fuchang01 已提交
157 158
  }

159 160 161 162
  if (use_ngraph) {
    config.EnableNgraph();
  }

163
  std::vector<std::vector<PaddleTensor>> outputs;
F
fuchang01 已提交
164 165 166 167 168 169
  std::vector<std::vector<PaddleTensor>> inputs;
  LoadInputData(&inputs);
  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&config),
                 inputs, &outputs, FLAGS_num_threads);
}

170 171
TEST(Analyzer_bert, profile) { profile(); }
#ifdef PADDLE_WITH_MKLDNN
172 173 174 175 176
TEST(Analyzer_bert, profile_mkldnn) { profile(true, false); }
#endif

#ifdef PADDLE_WITH_NGRAPH
TEST(Analyzer_bert, profile_ngraph) { profile(false, true); }
177 178 179 180 181 182 183 184 185 186 187 188 189 190
#endif

// Check the fuse status
TEST(Analyzer_bert, fuse_statis) {
  AnalysisConfig cfg;
  SetConfig(&cfg);
  int num_ops;
  auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
  auto fuse_statis = GetFuseStatis(
      static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
  LOG(INFO) << "num_ops: " << num_ops;
}

// Compare result of NativeConfig and AnalysisConfig
191
void compare(bool use_mkldnn = false, bool use_ngraph = false) {
192 193 194 195
  AnalysisConfig cfg;
  SetConfig(&cfg);
  if (use_mkldnn) {
    cfg.EnableMKLDNN();
196
    cfg.pass_builder()->AppendPass("fc_mkldnn_pass");
197
  }
F
fuchang01 已提交
198

199 200 201 202
  if (use_ngraph) {
    cfg.EnableNgraph();
  }

F
fuchang01 已提交
203 204 205
  std::vector<std::vector<PaddleTensor>> inputs;
  LoadInputData(&inputs);
  CompareNativeAndAnalysis(
206
      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), inputs);
F
fuchang01 已提交
207 208
}

209
TEST(Analyzer_bert, compare) { compare(); }
F
fuchang01 已提交
210
#ifdef PADDLE_WITH_MKLDNN
211 212 213 214 215 216 217 218 219
TEST(Analyzer_bert, compare_mkldnn) {
  compare(true, false /* use_mkldnn, no use_ngraph */);
}
#endif

#ifdef PADDLE_WITH_NGRAPH
TEST(Analyzer_bert, compare_ngraph) {
  compare(false, true /* no use_mkldnn, use_ngraph */);
}
F
fuchang01 已提交
220
#endif
221 222

// Compare Deterministic result
223 224 225 226 227 228 229 230 231
TEST(Analyzer_bert, compare_determine) {
  AnalysisConfig cfg;
  SetConfig(&cfg);

  std::vector<std::vector<PaddleTensor>> inputs;
  LoadInputData(&inputs);
  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
                       inputs);
}
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270

TEST(Analyzer_bert, transfer_scope_cache) {
  AnalysisConfig config;
  SetConfig(&config);

  std::vector<PaddleTensor> input, output;
  auto predictor = CreatePaddlePredictor<AnalysisConfig>(config);

  int threads_num = 10;
  std::vector<std::thread> threads;
  std::unordered_set<std::unordered_set<paddle::framework::Scope *> *>
      global_transfer_scope_cache;
  std::unordered_set<std::unordered_map<size_t, paddle::framework::Scope *> *>
      global_transfer_data_cache;

  std::ifstream fin(FLAGS_infer_data);
  std::string line;

  for (int i = 0; i < threads_num; i++) {
    threads.emplace_back([&, i]() {
      std::getline(fin, line);
      ParseLine(line, &input);
      predictor->Run(input, &output, FLAGS_batch_size);
      global_transfer_scope_cache.insert(
          &paddle::framework::global_transfer_scope_cache());
      global_transfer_data_cache.insert(
          &paddle::framework::global_transfer_data_cache());
    });
    threads[0].join();
    threads.clear();
    std::vector<PaddleTensor>().swap(input);
  }
  // Since paddle::framework::global_transfer_scope_cache() and
  // paddle::framework::global_transfer_data_cache() are thread_local,
  // their pointer should be different among different thread id.
  PADDLE_ENFORCE(global_transfer_scope_cache.size(), threads_num);
  PADDLE_ENFORCE(global_transfer_data_cache.size(), threads_num);
}

F
fuchang01 已提交
271 272
}  // namespace inference
}  // namespace paddle