analyzer_int8_image_classification_tester.cc 5.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
/* Copyright (c) 2019 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 <fstream>
#include <iostream>
#include "paddle/fluid/inference/api/paddle_analysis_config.h"
#include "paddle/fluid/inference/tests/api/tester_helper.h"

namespace paddle {
namespace inference {
namespace analysis {

void SetConfig(AnalysisConfig *cfg) {
  cfg->SetModel(FLAGS_infer_model);
  cfg->DisableGpu();
  cfg->SwitchIrOptim();
28
  cfg->SwitchSpecifyInputNames();
29 30 31 32 33 34 35 36 37 38
  cfg->SetCpuMathLibraryNumThreads(FLAGS_paddle_num_threads);
  cfg->EnableMKLDNN();
}

template <typename T>
class TensorReader {
 public:
  TensorReader(std::ifstream &file, size_t beginning_offset,
               std::vector<int> shape, std::string name)
      : file_(file), position(beginning_offset), shape_(shape), name_(name) {
39 40
    numel = std::accumulate(shape_.begin(), shape_.end(), size_t{1},
                            std::multiplies<size_t>());
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
  }

  PaddleTensor NextBatch() {
    PaddleTensor tensor;
    tensor.name = name_;
    tensor.shape = shape_;
    tensor.dtype = GetPaddleDType<T>();
    tensor.data.Resize(numel * sizeof(T));

    file_.seekg(position);
    file_.read(static_cast<char *>(tensor.data.data()), numel * sizeof(T));
    position = file_.tellg();

    if (file_.eof()) LOG(ERROR) << name_ << ": reached end of stream";
    if (file_.fail())
      throw std::runtime_error(name_ + ": failed reading file.");

    return tensor;
  }

 protected:
  std::ifstream &file_;
  size_t position;
  std::vector<int> shape_;
  std::string name_;
  size_t numel;
};

std::shared_ptr<std::vector<PaddleTensor>> GetWarmupData(
70 71
    const std::vector<std::vector<PaddleTensor>> &test_data,
    int num_images = FLAGS_warmup_batch_size) {
72
  int test_data_batch_size = test_data[0][0].shape[0];
73
  auto iterations = test_data.size();
74
  PADDLE_ENFORCE(
75
      static_cast<size_t>(num_images) <= iterations * test_data_batch_size,
76 77
      "The requested quantization warmup data size " +
          std::to_string(num_images) + " is bigger than all test data size.");
78 79

  PaddleTensor images;
80
  images.name = "image";
81 82 83 84 85
  images.shape = {num_images, 3, 224, 224};
  images.dtype = PaddleDType::FLOAT32;
  images.data.Resize(sizeof(float) * num_images * 3 * 224 * 224);

  PaddleTensor labels;
86
  labels.name = "label";
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
  labels.shape = {num_images, 1};
  labels.dtype = PaddleDType::INT64;
  labels.data.Resize(sizeof(int64_t) * num_images);

  for (int i = 0; i < num_images; i++) {
    auto batch = i / test_data_batch_size;
    auto element_in_batch = i % test_data_batch_size;
    std::copy_n(static_cast<float *>(test_data[batch][0].data.data()) +
                    element_in_batch * 3 * 224 * 224,
                3 * 224 * 224,
                static_cast<float *>(images.data.data()) + i * 3 * 224 * 224);

    std::copy_n(static_cast<int64_t *>(test_data[batch][1].data.data()) +
                    element_in_batch,
                1, static_cast<int64_t *>(labels.data.data()) + i);
  }

  auto warmup_data = std::make_shared<std::vector<PaddleTensor>>(2);
  (*warmup_data)[0] = std::move(images);
  (*warmup_data)[1] = std::move(labels);
  return warmup_data;
}

void SetInput(std::vector<std::vector<PaddleTensor>> *inputs,
              int32_t batch_size = FLAGS_batch_size) {
  std::ifstream file(FLAGS_infer_data, std::ios::binary);
  if (!file) {
    FAIL() << "Couldn't open file: " << FLAGS_infer_data;
  }

  int64_t total_images{0};
  file.read(reinterpret_cast<char *>(&total_images), sizeof(total_images));
  LOG(INFO) << "Total images in file: " << total_images;

  std::vector<int> image_batch_shape{batch_size, 3, 224, 224};
  std::vector<int> label_batch_shape{batch_size, 1};
123
  auto images_offset_in_file = static_cast<size_t>(file.tellg());
124
  auto labels_offset_in_file =
125
      images_offset_in_file + sizeof(float) * total_images * 3 * 224 * 224;
126

127
  TensorReader<float> image_reader(file, images_offset_in_file,
128
                                   image_batch_shape, "image");
129 130 131
  TensorReader<int64_t> label_reader(file, labels_offset_in_file,
                                     label_batch_shape, "label");

132
  auto iterations_max = total_images / batch_size;
133 134 135 136 137
  auto iterations = iterations_max;
  if (FLAGS_iterations > 0 && FLAGS_iterations < iterations_max) {
    iterations = FLAGS_iterations;
  }
  for (auto i = 0; i < iterations; i++) {
138 139 140 141 142 143 144
    auto images = image_reader.NextBatch();
    auto labels = label_reader.NextBatch();
    inputs->emplace_back(
        std::vector<PaddleTensor>{std::move(images), std::move(labels)});
  }
}

145
TEST(Analyzer_int8_image_classification, quantization) {
146 147 148 149 150 151
  AnalysisConfig cfg;
  SetConfig(&cfg);

  AnalysisConfig q_cfg;
  SetConfig(&q_cfg);

152
  // read data from file and prepare batches with test data
153
  std::vector<std::vector<PaddleTensor>> input_slots_all;
154
  SetInput(&input_slots_all);
155

156 157
  // prepare warmup batch from input data read earlier
  // warmup batch size can be different than batch size
158
  std::shared_ptr<std::vector<PaddleTensor>> warmup_data =
159
      GetWarmupData(input_slots_all);
160

161
  // configure quantizer
162 163
  q_cfg.EnableMkldnnQuantizer();
  q_cfg.mkldnn_quantizer_config()->SetWarmupData(warmup_data);
164
  q_cfg.mkldnn_quantizer_config()->SetWarmupBatchSize(FLAGS_warmup_batch_size);
165

166
  CompareQuantizedAndAnalysis(&cfg, &q_cfg, input_slots_all);
167 168 169 170 171
}

}  // namespace analysis
}  // namespace inference
}  // namespace paddle