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

MetricsLogger.log_value(key: str | Tuple[str, ...], value: Any, *, reduce: str | None = 'mean', window: int | float | None = None, ema_coeff: float | None = None, percentiles: List[int] | bool = False, clear_on_reduce: bool = False, with_throughput: bool = False, throughput_ema_coeff: float | None = None, reduce_per_index_on_aggregate: bool = False) → None[source]#

Logs a new value under a (possibly nested) key to the logger.

from ray.rllib.utils.metrics.metrics_logger import MetricsLogger
from ray.rllib.utils.test_utils import check

logger = MetricsLogger(root=True)

# Log n simple float values under the "loss" key. By default, all logged
# values under that key are averaged, once `reduce()` is called.
logger.log_value("loss", 0.01, window=10)
logger.log_value("loss", 0.02, window=10)
logger.log_value("loss", 0.03, window=10)

# Peek at the current (reduced) value.
# Note that in the underlying structure, the internal values list still
# contains all logged values (0.01, 0.02, and 0.03).
check(logger.peek("loss"), 0.02)

# Log 10x (window size) the same value.
for _ in range(10):
    logger.log_value("loss", 0.05, window=10)
check(logger.peek("loss"), 0.05)

# Internals check (note that users should not be concerned with accessing
# these). Len should always be 10, since the underlying struct is a
# `deque(max_len=10)`.
check(len(logger.stats["loss"].values), 10)

# Only, when we call `reduce` does the underlying structure get "cleaned
# up". In this case, the list is shortened to 10 items (window size).
results = logger.reduce()
check(results, {"loss": 0.05})
check(len(logger.stats["loss"].values), 10)

# Log a value under a deeper nested key.
logger.log_value(("some", "nested", "key"), -1.0)
check(logger.peek(("some", "nested", "key")), -1.0)

# Log n values without reducing them (we want to just collect some items).
logger.log_value("some_items", 5.0, reduce=None)
logger.log_value("some_items", 6.0, reduce=None)
logger.log_value("some_items", 7.0, reduce=None)
# Peeking at these returns the full list of items (no reduction set up).
check(logger.peek("some_items"), [5.0, 6.0, 7.0])
# If you don't want the internal list to grow indefinitely, you should set
# `clear_on_reduce=True`:
logger.log_value("some_more_items", -5.0, reduce=None, clear_on_reduce=True)
logger.log_value("some_more_items", -6.0, reduce=None, clear_on_reduce=True)
logger.log_value("some_more_items", -7.0, reduce=None, clear_on_reduce=True )
# Peeking at these returns the full list of items (no reduction set up).
check(logger.peek("some_more_items"), [-5.0, -6.0, -7.0])
# Reducing everything (and return plain values, not `Stats` objects).
results = logger.reduce()
check(results, {
    "loss": 0.05,
    "some": {
        "nested": {
            "key": -1.0,
        },
    },
    "some_items": [5.0, 6.0, 7.0],  # reduce=None; list as-is
    "some_more_items": [-5.0, -6.0, -7.0],  # reduce=None; list as-is
})
# However, the `reduce()` call did empty the `some_more_items` list
# (b/c we set `clear_on_reduce=True`).
check(logger.peek("some_more_items"), [])
# ... but not the "some_items" list (b/c `clear_on_reduce=False`).
check(logger.peek("some_items"), [])
Parameters:

  • key – The key (or nested key-tuple) to log the value under.

  • value – The value to log. This should be a numeric value.

  • reduce – The reduction method to apply, once self.reduce() is called. If None, will collect all logged values under key in a list (and also return that list upon calling self.reduce()).

  • window – An optional window size to reduce over. If not None, then the reduction operation is only applied to the most recent window items, and - after reduction - the internal values list under key is shortened to hold at most window items (the most recent ones). Must be None if ema_coeff is provided. If None (and ema_coeff is None), reduction must not be “mean”.

  • ema_coeff – An optional EMA coefficient to use if reduce is “mean” and no window is provided. Note that if both window and ema_coeff are provided, an error is thrown. Also, if ema_coeff is provided, reduce must be “mean”. The reduction formula for EMA is: EMA(t1) = (1.0 - ema_coeff) * EMA(t0) + ema_coeff * new_value

  • percentiles – If reduce is None, we can compute the percentiles of the values list given by percentiles. Defaults to [0, 0.5, 0.75, 0.9, 0.95, 0.99, 1] if set to True. When using percentiles, a window must be provided. This window should be chosen carefully. RLlib computes exact percentiles and the computational complexity is O(m*n*log(n/m)) where n is the window size and m is the number of parallel metrics loggers involved (for example, m EnvRunners).

  • clear_on_reduce – If True, all values under key will be emptied after self.reduce() is called. Setting this to True is useful for cases, in which the internal values list would otherwise grow indefinitely, for example if reduce is None and there is no window provided.

  • with_throughput – Whether to track a throughput estimate together with this metric. This is only supported for reduce=sum and clear_on_reduce=False metrics (aka. “lifetime counts”). The Stats object under the logged key then keeps track of the time passed between two consecutive calls to reduce() and update its throughput estimate. The current throughput estimate of a key can be obtained through: <MetricsLogger>.peek(key, throughput=True).

  • throughput_ema_coeff – The EMA coefficient to use for throughput tracking. Only used if with_throughput=True. Defaults to 0.05 if with_throughput is True.

  • reduce_per_index_on_aggregate – If True, when merging Stats objects in parallel, we reduce incoming values per index such that the new value at index n will be the reduced value of all incoming values at index n. If False, when reducing n Stats, the first n merged values will be the reduced value of all incoming values at index 0, the next n merged values will be the reduced values of all incoming values at index 1, etc.