TrainingCallback

class TrainingCallback

Bases: object

An interface for training callbacks.

Initialize the callback.

Attributes Summary

loss	The loss, accessed via the training loop.
model	The model, accessed via the training loop.
optimizer	The optimizer, accessed via the training loop.
result_tracker	The result tracker, accessed via the training loop.
training_loop	The training loop.

Methods Summary

on_batch(epoch, batch, batch_loss, **kwargs)	Call for training batches.
post_batch(epoch, batch, **kwargs)	Call for training batches.
post_epoch(epoch, epoch_loss, **kwargs)	Call after epoch.
post_train(losses, **kwargs)	Call after training.
pre_batch(**kwargs)	Call before training batch.
pre_step(**kwargs)	Call before the optimizer's step.
register_training_loop(training_loop)	Register the training loop.

Attributes Documentation

loss

The loss, accessed via the training loop.

model

The model, accessed via the training loop.

optimizer

The optimizer, accessed via the training loop.

result_tracker

The result tracker, accessed via the training loop.

training_loop

The training loop.

Methods Documentation

on_batch(epoch: int, batch, batch_loss: float, **kwargs: Any) → None

Call for training batches.

Parameters:

• epoch (int)

• batch_loss (float)

• kwargs (Any)

Return type:

None

post_batch(epoch: int, batch, **kwargs: Any) → None

Call for training batches.

Parameters:

• epoch (int)

• kwargs (Any)

Return type:

None

post_epoch(epoch: int, epoch_loss: float, **kwargs: Any) → None

Call after epoch.

Parameters:

• epoch (int)

• epoch_loss (float)

• kwargs (Any)

Return type:

None

post_train(losses: list[float], **kwargs: Any) → None

Call after training.

Parameters:

• losses (list[float])

• kwargs (Any)

Return type:

None

pre_batch(**kwargs: Any) → None

Call before training batch.

Parameters:

kwargs (Any)

Return type:

None

pre_step(**kwargs: Any) → None

Call before the optimizer’s step.

Parameters:

kwargs (Any)

Return type:

None

register_training_loop(training_loop: TrainingLoop) → None

Register the training loop.

Parameters:

training_loop (TrainingLoop)

Return type:

None