composer.callbacks

Note

To write a custom callback, see composer.Callback.

Composer contains built-in callbacks, which can be added via the --callbacks CLI flag:

python examples/run_mosaic_trainer.py -f my_model.yaml --callbacks lr_monitor grad_monitor

Callbacks

Benchmarker	Fast-forward the training loop to record throughput for specific epochs and/or steps.
GradMonitor	Logs the L2 norm.
LRMonitor	Logs the learning rate.
SpeedMonitor	Logs the training throughput.
TorchProfiler	Profile the execution using torch.profiler.profile.

Callback Hyperparameters

BenchmarkerHparams	Benchmarker hyperparameters.
GradMonitorHparams	GradMonitor hyperparamters.
LRMonitorHparams	LRMonitor hyperparameters.
SpeedMonitorHparams	SpeedMonitor hyperparameters.
TorchProfilerHparams	TorchProfiler hyperparameters.