kd.metrics.SkipIfMissing#

class kauldron.metrics.SkipIfMissing(metric: kauldron.metrics.base.Metric)[source]

Bases: kauldron.metrics.base.Metric

Skip this metric if any of the keys are missing.

This can be useful for example for metrics that are only defined for a subset of the datasets, or for metrics of gradients that would fail during evaluation.

Usage:

cfg.train_metrics[“optional_metric”] = kd.metrics.SkipIfMissing(: kd.metrics.Norm(tensor=”grads.encoder”)

)

metric

The metric to skip if any of its kontext-keys are missing.

Type:: kauldron.metrics.base.Metric

metric: kauldron.metrics.base.Metric

class State( state: base_state.State | None = None, *, parent: _MetricT = _EMPTY_TYPE.EMPTY, )[source]

Bases: kauldron.metrics.base_state.State

State for SkipIfMissing.

state: base_state.State | None = None

classmethod empty() → kauldron.metrics.base.SkipIfMissing.State[source]: Returns an empty instance (i.e. .merge(State.empty()) is a no-op).

merge( other: kauldron.metrics.base.SkipIfMissing.State, ) → kauldron.metrics.base.SkipIfMissing.State[source]

Returns a new state that is the accumulation of self and other.

Parameters:: other – A State whose intermediate values should be accumulated onto the values of self.
Returns:: A new State that accumulates the value from both self and other.

compute() → PyTree[Any][source]: Computes final metrics from intermediate values.

replace(**updates): Returns a new object replacing the specified fields with new values.

get_state(

**kwargs,

) → kauldron.metrics.base.Metric.State[source]

get_state_from_context( context: Any, ) → kauldron.metrics.base.Metric.State[source]

empty() → kauldron.metrics.base.Metric.State[source]