tune

The tune config group allows to make use of the lightning.Tuner. A scheduler may call lightning_tune() before starting a Trainer.fit run. This will (depending on the configuration) allow to modify two hyperparmeters:

  • the batch size (to maximize usage of VRAM)

  • the learning rate (to stabilize learning)

default

lr
default: true

  • activate the learning rate tuning

  • happens before the actual fitting procedure through lightning_tune()

bs
default: false

  • activate the batch size tuning

  • happens before the actual fitting procedure through lightning_tune()

lr_kwargs
default: {}

  • any additional kwargs to the learning rate finder

  • see lr_find

bs_kwargs
default: {}