training.md

Training loop API reference

The training loop API centers around the abstract type Phase and the function step!. To implement a custom training, you need to

Usage

• fit for n epochs of supervised training and validation using fit!(learner, n)
• train for an epoch using epoch!(learner, phase, dataiter)

Extending

• subtype Phase
• implement step!

You can optionally

• overwrite default epoch! implementation
• implement phasedataiter to define which data iterator should be used when epoch! is called without one.
• create custom Callback and Events with event handlers that dispatch on your Phase subtype.

Control flow

Inside callback handlers and step! implementations, you can throw CancelFittingException to stop the training and CancelEpochException and CancelStepException to skip the current epoch or step.