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
SanityCheck *	Performs sanity checks on data, model and loss before training
StopOnNaNLoss	Stops training early if a step loss is NaN
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

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

And for working with callbacks on an existing Learner :

• setcallbacks!

• addcallback!

• getcallback

• replacecallback!

• removecallback!

Extension API

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

• Callback

• stateaccess

• runafter

• resolveconflict