Method OptimizeHillClimb

tensorflow/core/framework/model.cc:886–952

}

void Model::OptimizeHillClimb(int64 cpu_budget, int64 ram_budget) {
  std::shared_ptr<Node> snapshot;
  {
    tf_shared_lock lock(mu_);
    snapshot = output_->Snapshot(nullptr);
  }
  VLOG(2) << "Starting optimization of tunable parameters with HillClimb";
  const double processing_time = TotalProcessingTime(snapshot);
  auto parameters = CollectTunableParameters(snapshot);
  // We add the number of model's buffered bytes because it is excluded from the
  // memory budget, but it is included in the maximum number of buffered bytes.
  ram_budget += TotalBufferedBytes(snapshot);
  // Buffer size parameter will only be incremented if the output latency
  // improvement is greater than this constant.
  constexpr double kBufferSizeMinDelta = 1.0L;

  for (auto& pair : parameters) {
    pair.second->value = pair.second->min;
  }
  while (true) {
    const double output_time = OutputTime(snapshot, /*gradient=*/nullptr);
    bool all_max = true;
    for (auto& pair : parameters) {
      if (pair.second->value < pair.second->max) {
        all_max = false;
        break;
      }
    }
    if (output_time < processing_time / cpu_budget || all_max ||
        TotalMaximumBufferedBytes(snapshot) > ram_budget) {
      break;
    }
    double best_delta = -1.0L;
    Parameter* best_parameter = nullptr;
    for (auto& pair : parameters) {
      if (pair.second->value == pair.second->max) {
        continue;
      }
      pair.second->value++;
      double new_output_time = OutputTime(snapshot, /*gradient=*/nullptr);
      double delta = output_time - new_output_time;
      if (delta > best_delta &&
          (delta > kBufferSizeMinDelta || pair.second->name != kBufferSize)) {
        best_delta = delta;
        best_parameter = pair.second.get();
      }
      pair.second->value--;
    }
    if (!best_parameter) {
      VLOG(2) << "Failed to find a tunable parameter that would decrease the "
                 "output time. This means that the autotuning optimization got "
                 "stuck in a local maximum. The optimization attempt will be "
                 "aborted.";
      return;
    }
    best_parameter->value++;
  }
  VLOG(2) << "Number of tunable parameters: " << parameters.size();