tensorflow lite example label_image 分析【二】

接上文

3.代码分析

main函数首先将入参写入参数结构体

Settings s;

struct Settings {
  bool verbose = false;
  bool accel = false;
  bool input_floating = false;
  bool profiling = false;
  int loop_count = 1;
  float input_mean = 127.5f;
  float input_std = 127.5f;
  string model_name = "./mobilenet_quant_v1_224.tflite";
  string input_bmp_name = "./grace_hopper.bmp";
  string labels_file_name = "./labels.txt";
  string input_layer_type = "uint8_t";
  int number_of_threads = 4;

};

运行执行函数 RunInference(&s);

【背景知识：https://blog.csdn.net/jILRvRTrc/article/details/80553561】

RunInference();

1）建立模型model = tflite::FlatBufferModel::BuildFromFile(s->model_name.c_str());

2）建立OpResolver 用于指向每个node的操作函数 tflite::ops::builtin::BuiltinOpResolver resolver;

3）建立解释器 tflite::InterpreterBuilder(*model, resolver)(&interpreter);

4）对解释器进行参数设置包括

interpreter->UseNNAPI(s->accel); 是否使用NNAPI加速

interpreter->SetNumThreads(s->number_of_threads);设置运行线程数

bmp文件读入并进行必要的resize

int image_width = 224;
  int image_height = 224;
  int image_channels = 3;
  std::vector<uint8_t> in = read_bmp(s->input_bmp_name, &image_width,

                                     &image_height, &image_channels, s);

resize<float>(interpreter->typed_tensor<float>(input), in.data(),
                    image_height, image_width, image_channels, wanted_height,

                    wanted_width, wanted_channels, s);

打印运行参数相关信息

if (s->verbose) PrintInterpreterState(interpreter.get());

  // output something like
  // time (ms) , Node xxx, OpCode xxx, symblic name

  //      5.352, Node   5, OpCode   4, DEPTHWISE_CONV_2D

运行模型及获得运行时间

struct timeval start_time, stop_time;
  gettimeofday(&start_time, nullptr);
  for (int i = 0; i < s->loop_count; i++) {
    if (interpreter->Invoke() != kTfLiteOk) {
      LOG(FATAL) << "Failed to invoke tflite!\n";
    }
  }

  gettimeofday(&stop_time, nullptr);

打印profiling

  if (s->profiling) {
    profiler->StopProfiling();
    auto profile_events = profiler->GetProfileEvents();
    for (int i = 0; i < profile_events.size(); i++) {
      auto op_index = profile_events[i]->event_metadata;
      const auto node_and_registration =
          interpreter->node_and_registration(op_index);
      const TfLiteRegistration registration = node_and_registration->second;
      PrintProfilingInfo(profile_events[i], op_index, registration);
    }

  }

获取输出

int output = interpreter->outputs()[0];
  switch (interpreter->tensor(output)->type) {
    case kTfLiteFloat32:
      get_top_n<float>(interpreter->typed_output_tensor<float>(0), output_size,

                       num_results, threshold, &top_results, true);

加载label并显示对应输出结果

 std::vector<string> labels;
  size_t label_count;


  if (ReadLabelsFile(s->labels_file_name, &labels, &label_count) != kTfLiteOk)
    exit(-1);


  for (const auto& result : top_results) {
    const float confidence = result.first;
    const int index = result.second;
    LOG(INFO) << confidence << ": " << index << " " << labels[index] << "\n";

  }