medkit.training.trainer_config#

Classes:

TrainerConfig(output_dir[, learning_rate, ...])

Trainer configuration
class TrainerConfig(output_dir, learning_rate=1e-05, nb_training_epochs=3, dataloader_nb_workers=0, batch_size=1, seed=None, gradient_accumulation_steps=1, do_metrics_in_training=False, metric_to_track_lr='loss', checkpoint_period=1, checkpoint_metric='loss', minimize_checkpoint_metric=True)[source]#

Trainer configuration

Parameters

  • output_dir (str) – The output directory where the checkpoint will be saved.

  • learning_rate (float) – The initial learning rate.

  • nb_training_epochs (int) – Total number of training/evaluation epochs to do.

  • dataloader_nb_workers (int) – Number of subprocess for the data loading. The default value is 0, the data will be loaded in the main process. If this config is for a HuggingFace model, do not change this value.

  • batch_size (int) – Number of samples per batch to load.

  • seed (Optional[int]) – Random seed to use with PyTorch and numpy. It should be set to ensure reproducibility between experiments.

  • gradient_accumulation_steps (int) – Number of steps to accumulate gradient before performing an optimization step.

  • do_metrics_in_training (bool) – By default, only the custom metrics are computed using eval_data. If set to True, the custom metrics are computed also using training_data.

  • metric_to_track_lr (str) – Name of the eval metric to be tracked for updating the learning rate. By default, eval loss is tracked.

  • checkpoint_period (int) – How often, in number of epochs, should we save a checkpoint. Use 0 to only save last checkpoint.

  • checkpoint_metric (str) – Name of the eval metric to be tracked for selecting the best checkpoint. By default, eval loss is tracked.

  • minimize_checkpoint_metric (bool) – If True, the checkpoint with the lowest metric value will be selected as best, otherwise the checkpoint with the highest metric value.