ray.rllib.utils.metrics.metrics_logger.MetricsLogger.peek#

MetricsLogger.peek(key: str | Tuple[str, ...], *, default: Any | None = None, throughput: bool = False) → Any[source]#

Returns the (reduced) value(s) found under the given key or key sequence.

If key only reaches to a nested dict deeper in self, that sub-dictionary’s entire values are returned as a (nested) dict with its leafs being the reduced peek values.

Note that calling this method does NOT cause an actual underlying value list reduction, even though reduced values are being returned. It’ll keep all internal structures as-is.

logger = MetricsLogger()
ema = 0.01

# Log some (EMA reduced) values.
key = ("some", "nested", "key", "sequence")
logger.log_value(key, 2.0, ema_coeff=ema)
logger.log_value(key, 3.0)

# Expected reduced value:
expected_reduced = (1.0 - ema) * 2.0 + ema * 3.0

# Peek at the (reduced) value under `key`.
check(logger.peek(key), expected_reduced)

# Peek at the (reduced) nested struct under ("some", "nested").
check(
    logger.peek(("some", "nested")),
    {"key": {"sequence": expected_reduced}},
)

# Log some more, check again.
logger.log_value(key, 4.0)
expected_reduced = (1.0 - ema) * expected_reduced + ema * 4.0
check(logger.peek(key), expected_reduced)
Parameters:

  • key – The key/key sequence of the sub-structure of self, whose (reduced) values to return.

  • default – An optional default value in case key cannot be found in self. If default is not provided and key cannot be found, throws a KeyError.

  • throughput – Whether to return the current throughput estimate instead of the actual (reduced) value.

Returns:

The (reduced) values of the (possibly nested) sub-structure found under the given key or key sequence.

Raises:

KeyError – If key cannot be found AND default is not provided.