tf_agents.eval.metric_utils.eager_compute

Compute metrics using policy on the environment.

tf_agents.eval.metric_utils.eager_compute(
    metrics,
    environment,
    policy,
    num_episodes=1,
    train_step=None,
    summary_writer=None,
    summary_prefix='',
    use_function=True
)

Note: Because placeholders are not compatible with Eager mode we can not use python policies. Because we use tf_policies we need the environment time_steps to be tensors making it easier to use a tf_env for evaluations. Otherwise this method mirrors compute directly.

Args
metrics	List of metrics to compute.
environment	tf_environment instance.
policy	tf_policy instance used to step the environment.
num_episodes	Number of episodes to compute the metrics over.
train_step	An optional step to write summaries against.
summary_writer	An optional writer for generating metric summaries.
summary_prefix	An optional prefix scope for metric summaries.
use_function	Option to enable use of tf.function when collecting the metrics.

Returns
A dictionary of results {metric_name: metric_value}