tf_agents.metrics.py_metric.PyMetric

Defines the interface for metrics.

tf_agents.metrics.py_metric.PyMetric(
    name: Text, prefix: Text = 'Metrics'
)

prefix Prefix for the metric.
summary_op TF summary op for this metric.
summary_placeholder TF placeholder to be used for the result of this metric.

Methods

aggregate

View source

@staticmethod
aggregate(
    metrics: Sequence[tf_agents.metrics.py_metric.PyMetricType]
) -> tf_agents.typing.types.Float

Aggregates a list of metrics.

The default behaviour is to return the average of the metrics.

Args
metrics a list of metrics, of the same class.

Returns
The result of aggregating this metric.

log

View source

log()

reset

View source

@abc.abstractmethod
reset()

Resets internal stat gathering variables used to compute the metric.

result

View source

@abc.abstractmethod
result() -> Any

Evaluates the current value of the metric.

tf_summaries

View source

tf_summaries(
    train_step: tf_agents.typing.types.Int = None,
    step_metrics: Sequence[tf_agents.metrics.py_metric.MetricType] = ()
) -> tf.Operation

Build TF summary op and placeholder for this metric.

To execute the op, call py_metric.run_summaries.

Args
train_step Step counter for training iterations. If None, no metric is generated against the global step.
step_metrics Step values to plot as X axis in addition to global_step.

Returns
The summary op.

Raises
RuntimeError If this method has already been called (it can only be called once).
ValueError If any item in step_metrics is not of type PyMetric or tf_metric.TFStepMetric.

__call__

View source

__call__(
    *args
)

Method to update the metric contents.

To change the behavior of this function, override the call method.

Different subclasses might use this differently. For instance, the PyStepMetric takes in a trajectory, while the CounterMetric takes no parameters.

Args
*args See call method of subclass for specific arguments.