luz_callback_keep_best_model: Keep the best model

In mlverse/torchlight: Higher Level 'API' for 'torch'

Keep the best model

Description

Each epoch, if there's improvement in the monitored metric we serialize the model weights to a temp file. When training is done, we reload weights from the best model.

Usage

luz_callback_keep_best_model(
  monitor = "valid_loss",
  mode = "min",
  min_delta = 0
)

Arguments

monitor	A string in the format ⁠<set>_<metric>⁠ where ⁠<set>⁠ can be 'train' or 'valid' and ⁠<metric>⁠ can be the abbreviation of any metric that you are tracking during training. The metric name is case insensitive.
mode	Specifies the direction that is considered an improvement. By default 'min' is used. Can also be 'max' (higher is better) and 'zero' (closer to zero is better).
min_delta	Minimum improvement to reset the patience counter.

See Also

Other luz_callbacks: luz_callback_auto_resume(), luz_callback_csv_logger(), luz_callback_early_stopping(), luz_callback_interrupt(), luz_callback_lr_scheduler(), luz_callback_metrics(), luz_callback_mixed_precision(), luz_callback_mixup(), luz_callback_model_checkpoint(), luz_callback_profile(), luz_callback_progress(), luz_callback_resume_from_checkpoint(), luz_callback_train_valid(), luz_callback()

Examples

cb <- luz_callback_keep_best_model()

mlverse/torchlight documentation built on Sept. 19, 2024, 11:22 p.m.