Supported RunConfig Parameters

Parameters Supported by NPURunConfig

• model_dir: path for saving the model. Defaults to None.

  If model_dir set in NPURunConfig is different from that in NPUEstimator, an error is reported.

  If either NPURunConfig or NPUEstimator is configured with model_dir, the configured path applies.

  If neither NPURunConfig nor NPUEstimator is configured with model_dir, a model_dir_xxxxxxxxxx directory is created in the current script execution path to save the model file.

• tf_random_seed: seed of the initialization variable. Defaults to None.
• save_summary_steps: interval (in steps) for saving the summary. Defaults to 0.

  Applies only to the scenario where iterations_per_loop = 1. If iterations_per_loop > 1, the configured value may not be saved. For details about how to save information, see "Log and Summary Operators."

• save_checkpoints_steps: interval (in steps) for saving the checkpoints. Defaults to None.
  • This parameter is mutually exclusive with save_checkpoints_secs.
  • If save_checkpoints_steps and save_checkpoints_secs are set to None, the checkpoints are saved every 100 steps.
  • If the value of iterations_per_loop is greater than 1, set save_checkpoints_steps to a positive integer multiple of iterations_per_loop. Failure to do so may lead to checkpoint data not saved as defined by save_checkpoints_steps.

  To save the checkpoint data on only a specific device, modify the training script as follows:

  Original TensorFlow code:

  self._classifier=tf.estimator.Estimator(
    model_fn=cnn_model_fn,
    model_dir=self._model_dir,
    config=tf.estimator.RunConfig(
        save_checkpoints_steps=50 if hvd.rank() == 0 else None,
        keep_checkpoint_max=1))

  Code after porting:

  self._classifier=NPUEstimator(
    model_fn=cnn_model_fn,
    model_dir=self._model_dir,
    config=tf.estimator.NPURunConfig(
        save_checkpoints_steps=50 if get_rank_id() == 0 else 0,
        keep_checkpoint_max=1))

• save_checkpoints_secs: interval (in seconds) for saving the checkpoints. Defaults to None.

  This parameter is mutually exclusive with save_checkpoints_steps.

• session_config: ConfigProto object of session configuration. Defaults to None.
• keep_checkpoint_max: maximum number of checkpoint files that can be stored. Defaults to 5.
• keep_checkpoint_every_n_hours: checkpoint file saving duration in hours. Defaults to 10000. This function can be disabled.

  To use this function, set keep_checkpoint_max to a large value.

• log_step_count_steps: interval (in steps) for recording the global_step and loss values. Defaults to 100.

  Applies only to the scenario where iterations_per_loop = 1. If iterations_per_loop > 1, the configured value may not be saved. For details about how to save information, see "Log and Summary Operators."

Parameters Not Supported by NPURunConfig

The following parameters in RunConfig are not supported in NPURunConfig.

• train_distribute: distributed training enable. The distributed configuration is specified by experimental_distribute.

  This parameter is used only by TensorFlow Adapter. You are advised not to set it.

• device_fn: function of the Device field of each operation.
• protocol: (optional) protocol used to start the server. If the parameter is empty, the gRPC is used by default.
• eval_distribute: distributed evaluation enable. The distributed configuration is specified by experimental_distribute.

  This parameter is used only by TensorFlow Adapter. You are advised not to set it.

• experimental_distribute: distributed configuration.