Function define_performance

official/utils/flags/_performance.py:53–294

Register flags for specifying performance tuning arguments. Args: num_parallel_calls: Create a flag to specify parallelism of data loading. inter_op: Create a flag to allow specification of inter op threads. intra_op: Create a flag to allow specification of intra op threads. synth

(num_parallel_calls=False,
                       inter_op=False,
                       intra_op=False,
                       synthetic_data=False,
                       max_train_steps=False,
                       dtype=False,
                       all_reduce_alg=False,
                       num_packs=False,
                       tf_gpu_thread_mode=False,
                       datasets_num_private_threads=False,
                       datasets_num_parallel_batches=False,
                       fp16_implementation=False,
                       loss_scale=False,
                       tf_data_experimental_slack=False,
                       enable_xla=False,
                       training_dataset_cache=False)

def define_performance(num_parallel_calls=False,
                       inter_op=False,
                       intra_op=False,
                       synthetic_data=False,
                       max_train_steps=False,
                       dtype=False,
                       all_reduce_alg=False,
                       num_packs=False,
                       tf_gpu_thread_mode=False,
                       datasets_num_private_threads=False,
                       datasets_num_parallel_batches=False,
                       fp16_implementation=False,
                       loss_scale=False,
                       tf_data_experimental_slack=False,
                       enable_xla=False,
                       training_dataset_cache=False):
  """Register flags for specifying performance tuning arguments.

  Args:
    num_parallel_calls: Create a flag to specify parallelism of data loading.
    inter_op: Create a flag to allow specification of inter op threads.
    intra_op: Create a flag to allow specification of intra op threads.
    synthetic_data: Create a flag to allow the use of synthetic data.
    max_train_steps: Create a flags to allow specification of maximum number of
      training steps
    dtype: Create flags for specifying dtype.
    all_reduce_alg: If set forces a specific algorithm for multi-gpu.
    num_packs: If set provides number of packs for MirroredStrategy's cross
      device ops.
    tf_gpu_thread_mode: gpu_private triggers us of private thread pool.
    datasets_num_private_threads: Number of private threads for datasets.
    datasets_num_parallel_batches: Determines how many batches to process in
      parallel when using map and batch from tf.data.
    fp16_implementation: Create fp16_implementation flag.
    loss_scale: Controls the loss scaling, normally for mixed-precision
      training. Can only be turned on if dtype is also True.
    tf_data_experimental_slack: Determines whether to enable tf.data's
      `experimental_slack` option.
    enable_xla: Determines if XLA (auto clustering) is turned on.
    training_dataset_cache: Whether to cache the training dataset on workers.
      Typically used to improve training performance when training data is in
      remote storage and can fit into worker memory.

  Returns:
    A list of flags for core.py to marks as key flags.
  """

  key_flags = []
  if num_parallel_calls:
    flags.DEFINE_integer(
        name="num_parallel_calls",
        short_name="npc",
        default=multiprocessing.cpu_count(),
        help=help_wrap("The number of records that are  processed in parallel "
                       "during input processing. This can be optimized per "
                       "data set but for generally homogeneous data sets, "
                       "should be approximately the number of available CPU "
                       "cores. (default behavior)"))