LossApply: Selectively apply loss to parameters

class hypercoil.loss.scheme.LossApply(loss: Callable, apply: Callable = <function unpack_or_noop>)

Callable loss function wrapper that composes the loss with a selector or other pretransformation.

Parameters:

losscallable

Loss function to apply.

applycallable

Callable that pretransforms the input to the loss. This can be used as a selector, so that different entries in a LossScheme are applied to different combinations of weights, inputs, and outputs. One use case is filtering the inputs to a LossScheme and forwarding only the ones relevant to the loss function at hand.

Attributes:

nu

Methods

__call__(*pparams, key, **params)	Evaluate the loss applied to the output of the apply operation.
apply()	For an argument sequence, unpack if it's a singleton sequence, and otherwise return the sequence.

cfg
get_loss
losses
step