allennlp.training.metric_tracker

class allennlp.training.metric_tracker.MetricTracker(patience: Optional[int] = None, metric_name: str = None, should_decrease: bool = None)[source]

Bases: object

This class tracks a metric during training for the dual purposes of early stopping and for knowing whether the current value is the best so far. It mimics the PyTorch state_dict / load_state_dict interface, so that it can be checkpointed along with your model and optimizer.

Some metrics improve by increasing; others by decreasing. Here you can either explicitly supply should_decrease, or you can provide a metric_name in which case “should decrease” is inferred from the first character, which must be “+” or “-“.

Parameters:
patience : int, optional (default = None)

If provided, then should_stop_early() returns True if we go this many epochs without seeing a new best value.

metric_name : str, optional (default = None)

If provided, it’s used to infer whether we expect the metric values to increase (if it starts with “+”) or decrease (if it starts with “-“). It’s an error if it doesn’t start with one of those. If it’s not provided, you should specify should_decrease instead.

should_decrease : str, optional (default = None)

If metric_name isn’t provided (in which case we can’t infer should_decrease), then you have to specify it here.

add_metric(metric: float) → None[source]

Record a new value of the metric and update the various things that depend on it.

add_metrics(metrics: Iterable[float]) → None[source]

Helper to add multiple metrics at once.

clear() → None[source]

Clears out the tracked metrics, but keeps the patience and should_decrease settings.

is_best_so_far() → bool[source]

Returns true if the most recent value of the metric is the best so far.

load_state_dict(state_dict: Dict[str, Any]) → None[source]

A Trainer can use this to hydrate a metric tracker from a serialized state.

should_stop_early() → bool[source]

Returns true if improvement has stopped for long enough.

state_dict() → Dict[str, Any][source]

A Trainer can use this to serialize the state of the metric tracker.