Interfaces

Core interface

Functionality

Core interface

First you need to define a custom LearningMethod (and possibly a Task).

Then implement the following functions by dispatching on the method.

Required functions	Description
`encode`	Encodes a sample containing input and target so they can be used for a training step
`encodeinput`	Encodes an input so that it can be fed to a model
`encodetarget`	Encodes a target so that it can be compared to a model output using a loss function
`decodeŷ`	Decodes the output of a model into a target

Optional functions	Description
`shouldbatch`	Defines whether the model for a method should take batches of
	encoded inputs. The default is `true`.

What is the context argument?

We often want to apply data augmentation during training but not during validation or inference. You can dispatch on context to define different behavior for the different situations. See Context to see the available options.
When should I implement encode vs. encodetarget?

See encode.