Table of Contents

Background

This is an extended design discussion for callback design in training fit loop of the Gluon Fit API design. Callback is a powerful tool, we can use it to deliver useful features during training for users. For example, metric update, validation, logging, and saving a model periodically.

In this document, we will discuss what functionality we should provide, how to manage shared/independent states in callbacks, how to pass arguments in callbacks, and what function should be implemented as callbacks.

Note: callbacks in the Gluon Fit-API/Estimator Design are called event handlers to avoid confusion of modules under mx.callbcaks (used for mxnet modules only)

...