2d05fc8af6533babd8b5ebc44149758c6b2535e7,official/benchmark/retinanet_benchmark.py,RetinanetBenchmarkReal,benchmark_1_gpu_coco,#RetinanetBenchmarkReal#,251

Before Change

  @flagsaver.flagsaver
  def benchmark_1_gpu_coco(self):
    Run RetinaNet model accuracy test with 1 GPU.
    self.num_gpus = 1
    self._setup()
    params = copy.deepcopy(self.params_override)
    params["train"]["batch_size"] = 8
    params["train"]["total_steps"] = 200
    params["train"]["iterations_per_loop"] = 1
    params["eval"]["eval_samples"] = 8
    FLAGS.num_gpus = self.num_gpus
    FLAGS.params_override = json.dumps(params)
    FLAGS.model_dir = self._get_model_dir("real_benchmark_1_gpu_coco")
    FLAGS.strategy_type = "one_device"
    // Use negative value to avoid saving checkpoints.
    FLAGS.save_checkpoint_freq = -1
    if self.timer_callback is None:
      logging.error("Cannot measure performance without timer callback")
    else:
      self._run_and_report_benchmark()

  @flagsaver.flagsaver
  def benchmark_xla_1_gpu_coco(self):
    Run RetinaNet model accuracy test with 1 GPU and XLA enabled.
    self.num_gpus = 1

After Change

  def benchmark_1_gpu_coco(self):
    Run RetinaNet model accuracy test with 1 GPU.
    self._setup()
    params = self._params()
    params["train"]["batch_size"] = 8
    params["train"]["total_steps"] = 200
    params["train"]["iterations_per_loop"] = 1
    params["eval"]["eval_samples"] = 8
    FLAGS.num_gpus = 1
    FLAGS.model_dir = self._get_model_dir("real_benchmark_1_gpu_coco")
    FLAGS.strategy_type = "one_device"
    self._run_and_report_benchmark(params)

  @flagsaver.flagsaver
  def benchmark_xla_1_gpu_coco(self):
    Run RetinaNet model accuracy test with 1 GPU and XLA enabled.

Instances