analyzer_lac_tester.cc 5.7 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
// 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.
T
tensor-tang 已提交
14

L
luotao1 已提交
15
#include "paddle/fluid/inference/tests/api/tester_helper.h"
T
tensor-tang 已提交
16

T
tensor-tang 已提交
17 18 19
namespace paddle {
namespace inference {
namespace analysis {
T
tensor-tang 已提交
20

21 22
using contrib::AnalysisConfig;

T
tensor-tang 已提交
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
struct DataRecord {
  std::vector<int64_t> data;
  std::vector<size_t> lod;
  // for dataset and nextbatch
  size_t batch_iter{0};
  std::vector<std::vector<size_t>> batched_lods;
  std::vector<std::vector<int64_t>> batched_datas;
  std::vector<std::vector<int64_t>> datasets;
  DataRecord() = default;
  explicit DataRecord(const std::string &path, int batch_size = 1) {
    Load(path);
    Prepare(batch_size);
    batch_iter = 0;
  }
  void Load(const std::string &path) {
    std::ifstream file(path);
    std::string line;
    int num_lines = 0;
    datasets.resize(0);
    while (std::getline(file, line)) {
      num_lines++;
      std::vector<std::string> data;
      split(line, ';', &data);
      std::vector<int64_t> words_ids;
      split_to_int64(data[1], ' ', &words_ids);
      datasets.emplace_back(words_ids);
    }
  }
  void Prepare(int bs) {
    if (bs == 1) {
      batched_datas = datasets;
      for (auto one_sentence : datasets) {
        batched_lods.push_back({0, one_sentence.size()});
      }
    } else {
      std::vector<int64_t> one_batch;
      std::vector<size_t> lod{0};
      int bs_id = 0;
      for (auto one_sentence : datasets) {
        bs_id++;
        one_batch.insert(one_batch.end(), one_sentence.begin(),
                         one_sentence.end());
        lod.push_back(lod.back() + one_sentence.size());
        if (bs_id == bs) {
          bs_id = 0;
          batched_datas.push_back(one_batch);
          batched_lods.push_back(lod);
          one_batch.clear();
          one_batch.resize(0);
          lod.clear();
          lod.resize(0);
          lod.push_back(0);
        }
      }
      if (one_batch.size() != 0) {
        batched_datas.push_back(one_batch);
        batched_lods.push_back(lod);
      }
    }
  }
83

T
tensor-tang 已提交
84 85 86 87 88 89 90 91 92 93 94
  DataRecord NextBatch() {
    DataRecord data;
    data.data = batched_datas[batch_iter];
    data.lod = batched_lods[batch_iter];
    batch_iter++;
    if (batch_iter >= batched_datas.size()) {
      batch_iter = 0;
    }
    return data;
  }
};
T
tensor-tang 已提交
95

T
tensor-tang 已提交
96 97 98 99 100 101 102 103 104 105 106 107
void GetOneBatch(std::vector<PaddleTensor> *input_slots, DataRecord *data,
                 int batch_size) {
  auto one_batch = data->NextBatch();
  PaddleTensor input_tensor;
  input_tensor.name = "word";
  input_tensor.shape.assign({static_cast<int>(one_batch.data.size()), 1});
  input_tensor.lod.assign({one_batch.lod});
  input_tensor.dtype = PaddleDType::INT64;
  TensorAssignData<int64_t>(&input_tensor, {one_batch.data});
  PADDLE_ENFORCE_EQ(batch_size, static_cast<int>(one_batch.lod.size() - 1));
  input_slots->assign({input_tensor});
}
T
tensor-tang 已提交
108

T
Tao Luo 已提交
109 110 111 112 113 114 115
void SetConfig(AnalysisConfig *cfg) {
  cfg->model_dir = FLAGS_infer_model;
  cfg->use_gpu = false;
  cfg->device = 0;
  cfg->specify_input_name = true;
  cfg->enable_ir_optim = true;
}
L
luotao1 已提交
116

T
Tao Luo 已提交
117 118 119 120 121 122 123 124
void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
  DataRecord data(FLAGS_infer_data, FLAGS_batch_size);
  std::vector<PaddleTensor> input_slots;
  int epoch = FLAGS_test_all_data ? data.batched_datas.size() : 1;
  LOG(INFO) << "number of samples: " << epoch;
  for (int bid = 0; bid < epoch; ++bid) {
    GetOneBatch(&input_slots, &data, FLAGS_batch_size);
    (*inputs).emplace_back(input_slots);
T
tensor-tang 已提交
125
  }
T
Tao Luo 已提交
126
}
T
tensor-tang 已提交
127

T
Tao Luo 已提交
128 129 130 131 132
// Easy for profiling independently.
TEST(Analyzer_LAC, profile) {
  AnalysisConfig cfg;
  SetConfig(&cfg);
  std::vector<PaddleTensor> outputs;
T
tensor-tang 已提交
133

T
Tao Luo 已提交
134 135 136
  std::vector<std::vector<PaddleTensor>> input_slots_all;
  SetInput(&input_slots_all);
  TestPrediction(cfg, input_slots_all, &outputs, FLAGS_num_threads);
T
tensor-tang 已提交
137

T
Tao Luo 已提交
138 139 140 141 142 143 144 145 146 147 148 149 150
  if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
    // the first inference result
    const int64_t lac_ref_data[] = {
        24, 25, 25, 25, 38, 30, 31, 14, 15, 44, 24, 25, 25, 25, 25, 25,
        44, 24, 25, 25, 25, 36, 42, 43, 44, 14, 15, 44, 14, 15, 44, 14,
        15, 44, 38, 39, 14, 15, 44, 22, 23, 23, 23, 23, 23, 23, 23};
    PADDLE_ENFORCE_EQ(outputs.size(), 1UL);
    size_t size = GetSize(outputs[0]);
    size_t batch1_size = sizeof(lac_ref_data) / sizeof(int64_t);
    PADDLE_ENFORCE_GE(size, batch1_size);
    int64_t *pdata = static_cast<int64_t *>(outputs[0].data.data());
    for (size_t i = 0; i < batch1_size; ++i) {
      EXPECT_EQ(pdata[i], lac_ref_data[i]);
T
tensor-tang 已提交
151
    }
T
tensor-tang 已提交
152
  }
T
tensor-tang 已提交
153
}
T
tensor-tang 已提交
154

T
Tao Luo 已提交
155 156 157 158 159 160
// Check the fuse status
TEST(Analyzer_LAC, fuse_statis) {
  AnalysisConfig cfg;
  SetConfig(&cfg);

  int num_ops;
161 162 163
  auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
  auto fuse_statis = GetFuseStatis(
      static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
T
Tao Luo 已提交
164 165 166 167 168
  ASSERT_TRUE(fuse_statis.count("fc_fuse"));
  ASSERT_TRUE(fuse_statis.count("fc_gru_fuse"));
  EXPECT_EQ(fuse_statis.at("fc_fuse"), 1);
  EXPECT_EQ(fuse_statis.at("fc_gru_fuse"), 4);
  EXPECT_EQ(num_ops, 11);
T
tensor-tang 已提交
169
}
T
tensor-tang 已提交
170

T
Tao Luo 已提交
171 172 173 174 175 176 177 178
// Compare result of NativeConfig and AnalysisConfig
TEST(Analyzer_LAC, compare) {
  AnalysisConfig cfg;
  SetConfig(&cfg);

  std::vector<std::vector<PaddleTensor>> input_slots_all;
  SetInput(&input_slots_all);
  CompareNativeAndAnalysis(cfg, input_slots_all);
T
tensor-tang 已提交
179 180
}

T
tensor-tang 已提交
181 182 183
}  // namespace analysis
}  // namespace inference
}  // namespace paddle