RSVP for your your local TensorFlow Everywhere event today!


Triggers saves policy checkpoints an agent's policy.

Inherits From: IntervalTrigger

Used in the notebooks

Used in the tutorials

On construction this trigger will generate a saved_model for a: greedy_policy, a collect_policy, and a raw_policy. When triggered a checkpoint will be saved which can be used to updated any of the saved_model policies.

saved_model_dir Base dir where checkpoints will be saved.
agent Agent to extract policies from.
train_step tf.Variable which keeps track of the number of train steps.
interval How often, in train_steps, the trigger will save. Note that as long as the >= interval number of steps have passed since the last trigger, the event gets triggered. The current value is not necessarily interval steps away from the last triggered value.
async_saving If True saving will be done asynchronously in a separate thread. Note if this is on the variable values in the saved checkpoints/models are not deterministic.
metadata_metrics A dictionary of metrics, whose result() method returns a scalar to be saved along with the policy. Currently only supported when async_saving is False.
start Initial value for the trigger passed directly to the base class. It helps control from which train step the weigts of the model are saved.



View source

Resets the trigger interval.


View source

Maybe trigger the event based on the interval.

value the value for triggering.
force_trigger If True, the trigger will be forced triggered unless the last trigger value is equal to value.