trainer

The trainer config group sets the flags for the lightning.Trainer used by mml The full documentation can be found here.

default_trainer

_target_

default: lightning.Trainer

• the trainer is instantiated by the scheduler through create_trainer()

• any kwargs accepted by Trainer can be provided as trainer.KWARG=VALUE

• callbacks are determined separately through callbacks (see there for details on checkpointing)

• the experiment logger is determined through logging

• see Trainer in the lightning documentation

benchmark

default: True

• deactivate if image sizes change over time or to reduce memory consumption

precision

default: 16-mixed

• deactivate if not supported by hardware or high precision is required

min_epochs

default: 10

• will block “early stopping” and similar from interrupting the training until this number of epochs is reached

max_epochs

default: 50

• will stop training once this number of epochs is reached

enable_model_summary

default: true

• prints a model summary at the beginning of each fitting

num_sanity_val_steps

default: 0

• as it is set to zero, no sanity check is performed to reduce time

max_time

default: null

• will stop training once the given training duration is reached

accelerator

default: auto

• determine the hardware accelerator, “auto” will choose depending on available hardware

devices

default: 1

• number of hardware devices, currently mml is not yet optimized for multi-GPU usage