Loss

class liesel.optim.Loss(*args, **kwargs)

Bases: Protocol

Protocol for optimizer losses.

OptimEngine is intentionally agnostic about the concrete loss type. Any object satisfying this protocol can be optimized: it must expose the split used for training and validation, provide initial parameter positions, compute training and validation losses, and provide gradients for optimizer updates.

Notes

The standard built-in optimizer calls value_and_grad() so the objective value used for monitoring and its gradient share one stochastic evaluation. Custom losses can inherit from LossMixin to get value_and_grad() and grad() automatically.

Methods

grad(params, carry)	Returns the gradient of loss_train_batched() with respect to params.
loss_monitor(params, carry)	Computes the complete validation monitoring loss at params.
loss_train(params, carry)	Computes the full-data training loss at params.
loss_train_batched(params, carry)	Computes the training loss for the current mini-batch.
position(position_keys)	Extracts an initial optimizer position.
value_and_grad(params, carry)	Returns (loss_train_batched(params, carry), grad).

Attributes

split

Train/validation/test split used by this loss.