analyzer_mm_dnn_tester.cc 5.1 KB
Newer Older
T
Tao Luo 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
// 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/inference/tests/api/tester_helper.h"

namespace paddle {
namespace inference {
using contrib::AnalysisConfig;

struct DataRecord {
22
  std::vector<std::vector<int64_t>> query, title;
T
Tao Luo 已提交
23
  std::vector<size_t> lod1, lod2;
24
  size_t batch_iter{0}, batch_size{1}, num_samples;  // total number of samples
T
Tao Luo 已提交
25 26 27 28 29 30 31 32 33
  DataRecord() = default;
  explicit DataRecord(const std::string &path, int batch_size = 1)
      : batch_size(batch_size) {
    Load(path);
  }
  DataRecord NextBatch() {
    DataRecord data;
    size_t batch_end = batch_iter + batch_size;
    // NOTE skip the final batch, if no enough data is provided.
34 35 36
    if (batch_end <= query.size()) {
      GetInputPerBatch(query, &data.query, &data.lod1, batch_iter, batch_end);
      GetInputPerBatch(title, &data.title, &data.lod2, batch_iter, batch_end);
T
Tao Luo 已提交
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
    }
    batch_iter += batch_size;
    return data;
  }
  void Load(const std::string &path) {
    std::ifstream file(path);
    std::string line;
    int num_lines = 0;
    while (std::getline(file, line)) {
      num_lines++;
      std::vector<std::string> data;
      split(line, '\t', &data);
      // load query data
      std::vector<int64_t> query_data;
      split_to_int64(data[0], ' ', &query_data);
      // load title data
      std::vector<int64_t> title_data;
      split_to_int64(data[1], ' ', &title_data);
55 56
      query.push_back(std::move(query_data));
      title.push_back(std::move(title_data));
T
Tao Luo 已提交
57 58 59 60 61 62 63 64 65 66 67 68
    }
    num_samples = num_lines;
  }
};

void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
                   int batch_size) {
  PaddleTensor lod_query_tensor, lod_title_tensor;
  lod_query_tensor.name = "left";
  lod_title_tensor.name = "right";
  auto one_batch = data->NextBatch();
  // assign data
69 70
  TensorAssignData<int64_t>(&lod_query_tensor, one_batch.query, one_batch.lod1);
  TensorAssignData<int64_t>(&lod_title_tensor, one_batch.title, one_batch.lod2);
T
Tao Luo 已提交
71 72 73 74 75 76 77 78
  // Set inputs.
  input_slots->assign({lod_query_tensor, lod_title_tensor});
  for (auto &tensor : *input_slots) {
    tensor.dtype = PaddleDType::INT64;
  }
}

void SetConfig(contrib::AnalysisConfig *cfg) {
79 80 81 82
  cfg->SetModel(FLAGS_infer_model);
  cfg->DisableGpu();
  cfg->SwitchSpecifyInputNames();
  cfg->SwitchIrOptim();
T
Tao Luo 已提交
83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156
}

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.num_samples / FLAGS_batch_size : 1;
  LOG(INFO) << "number of samples: " << epoch * FLAGS_batch_size;
  for (int bid = 0; bid < epoch; ++bid) {
    PrepareInputs(&input_slots, &data, FLAGS_batch_size);
    (*inputs).emplace_back(input_slots);
  }
}

// Easy for profiling independently.
TEST(Analyzer_MM_DNN, profile) {
  contrib::AnalysisConfig cfg;
  SetConfig(&cfg);
  std::vector<PaddleTensor> outputs;

  std::vector<std::vector<PaddleTensor>> input_slots_all;
  SetInput(&input_slots_all);
  TestPrediction(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
                 input_slots_all, &outputs, FLAGS_num_threads);

  if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) {
    PADDLE_ENFORCE_EQ(outputs.size(), 2UL);
    for (auto &output : outputs) {
      size_t size = GetSize(output);
      PADDLE_ENFORCE_GT(size, 0);
      float *result = static_cast<float *>(output.data.data());
      // output is probability, which is in (-1, 1).
      for (size_t i = 0; i < size; i++) {
        EXPECT_GT(result[i], -1);
        EXPECT_LT(result[i], 1);
      }
    }
  }
}

// Check the fuse status
TEST(Analyzer_MM_DNN, fuse_statis) {
  contrib::AnalysisConfig cfg;
  SetConfig(&cfg);

  int num_ops;
  auto predictor = CreatePaddlePredictor<AnalysisConfig>(cfg);
  auto fuse_statis = GetFuseStatis(
      static_cast<AnalysisPredictor *>(predictor.get()), &num_ops);
}

// Compare result of NativeConfig and AnalysisConfig
TEST(Analyzer_MM_DNN, compare) {
  contrib::AnalysisConfig cfg;
  SetConfig(&cfg);

  std::vector<std::vector<PaddleTensor>> input_slots_all;
  SetInput(&input_slots_all);
  CompareNativeAndAnalysis(
      reinterpret_cast<const PaddlePredictor::Config *>(&cfg), input_slots_all);
}

// Compare Deterministic result
TEST(Analyzer_MM_DNN, compare_determine) {
  AnalysisConfig cfg;
  SetConfig(&cfg);

  std::vector<std::vector<PaddleTensor>> input_slots_all;
  SetInput(&input_slots_all);
  CompareDeterministic(reinterpret_cast<const PaddlePredictor::Config *>(&cfg),
                       input_slots_all);
}

}  // namespace inference
}  // namespace paddle