Callback reference

Included callbacks

FluxTraining.jl comes with many callbacks included. Some of them are added to Learner by default, here marked with a *.

Callback	Description
`Metrics`*	Tracks loss and additional metrics on a per-step and per-epoch base
`Recorder`*	Records training stats like number of steps and epochs
`ProgressPrinter`*	Prints a progress bar for the current epoch during training
`MetricsPrinter`*	Prints out metrics after every epoch
`StopOnNaNLoss`*	Stops training early if a step loss is `NaN`
`SanityCheck`	Performs sanity checks on data, model and loss before training
`ToGPU`	Trains using a CUDA GPU if available
`Checkpointer`	Saves the model after every epoch
`EarlyStopping`	Stops training early when a criterion is met
`Scheduler`	Schedules hyperparameters
`LogMetrics`	Logs metrics to a logging backend
`LogHyperParams`	Logs hyperparameters to a logging backend
`LogVisualization`	Logs visualization to a logging backend
`LogHistograms`	Logs model weight histograms to a logging backend

To construct a learner without default callbacks, pass usedefaultcallbacks=false when constructing it.

For working with callbacks on an existing Learner, see:

There are also some utilities for creating callbacks:

CustomCallback to quickly hook a function into an event
throttle to run a callback only after every n events or every t seconds

Extension API

The following types and functions can be used to create custom callbacks. Read the custom callbacks guide for more context.

Backlinks

The following pages link back here:

Getting started, How to use callbacks, News, Training an image classifier, Training loop