Function TestInference

test/cpp/inference/test_helper.h:176–303

template <typename Place, bool CreateVars = true, bool PrepareContext = false>
void TestInference(
    const std::string& dirname,
    const std::vector<phi::DenseTensor*>& cpu_feeds,
    const std::vector<paddle::framework::FetchType*>& cpu_fetches,
    const int repeat = 1,
    const bool is_combined = false) {
  // 1. Define place, executor, scope
  auto place = Place();
  auto executor = paddle::framework::Executor(place);
  auto* scope = new paddle::framework::Scope();

  // Profile the performance
  paddle::platform::ProfilerState state;
  if (phi::is_cpu_place(place)) {
    state = paddle::platform::ProfilerState::kCPU;
  } else {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
    state = paddle::platform::ProfilerState::kAll;
    // The default device_id of phi::GPUPlace is 0.
    // Users can get the device_id using:
    //   int device_id = place.GetDeviceId();
    paddle::platform::SetDeviceId(0);
#else
    PADDLE_THROW(common::errors::Unavailable(
        "'CUDAPlace' is not supported in CPU only device."));
#endif
  }

  // 2. Initialize the inference_program and load parameters
  std::unique_ptr<paddle::framework::ProgramDesc> inference_program;

  // Enable the profiler
  paddle::platform::EnableProfiler(state);
  {
    phi::RecordEvent record_event("init_program");
    inference_program = InitProgram(&executor, scope, dirname, is_combined);
  }

  // Disable the profiler and print the timing information
  paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
                                    "load_program_profiler");
  paddle::platform::ResetProfiler();

  // 3. Get the feed_target_names and fetch_target_names
  const std::vector<std::string>& feed_target_names =
      inference_program->GetFeedTargetNames();
  const std::vector<std::string>& fetch_target_names =
      inference_program->GetFetchTargetNames();

  // 4. Prepare inputs: set up maps for feed targets
  std::map<std::string, const phi::DenseTensor*> feed_targets;
  for (size_t i = 0; i < feed_target_names.size(); ++i) {
    // Please make sure that cpu_feeds[i] is right for feed_target_names[i]
    feed_targets[feed_target_names[i]] = cpu_feeds[i];
  }

  // 5. Define Tensor to get the outputs: set up maps for fetch targets
  std::map<std::string, paddle::framework::FetchType*> fetch_targets;