tf.data.experimental.CheckpointInputPipelineHook

Checkpoints input pipeline state every N steps or seconds.

Inherits From: SessionRunHook

This hook saves the state of the iterators in the Graph so that when training is resumed the input pipeline continues from where it left off. This could potentially avoid overfitting in certain pipelines where the number of training steps per eval are small compared to the dataset size or if the training pipeline is pre-empted.

Differences from CheckpointSaverHook:

  1. Saves only the input pipelines in the "iterators" collection and not the global variables or other saveable objects.
  2. Does not write the GraphDef and MetaGraphDef to the summary.

Example of checkpointing the training pipeline:

est = tf.estimator.Estimator(model_fn)
while True:
  est.train(
      train_input_fn,
      hooks=[tf.data.experimental.CheckpointInputPipelineHook(est)],
      steps=train_steps_per_eval)
  # Note: We do not pass the hook here.
  metrics = est.evaluate(eval_input_fn)
  if should_stop_the_training(metrics):
    break

This hook should be used if the input pipeline state needs to be saved separate from the model checkpoint. Doing so may be useful for a few reasons:

  1. The input pipeline checkpoint may be large, if there are large shuffle or prefetch buffers for instance, and may bloat the checkpoint size.
  2. If the input pipeline is shared between training and validation, restoring the checkpoint during validation may override the validation input pipeline.

For saving the input pipeline checkpoint alongside the model weights use tf.data.experimental.make_saveable_from_iterator directly to create a SaveableObject and add to the SAVEABLE_OBJECTS collection. Note, however, that you will need to be careful not to restore the training iterator during eval. You can do that by not adding the iterator to the SAVEABLE_OBJECTS collector when building the eval graph.

estimator Estimator.
external_state_policy A string that identifies how to handle input pipelines that depend on external state. Possible values are 'ignore': The external state is silently ignored. 'warn': The external state is ignored, logging a warning. 'fail': The operation fails upon encountering external state. By default we set it to 'fail'.