xnn.common.config.schema.TrainerConfig#

class xnn.common.config.schema.TrainerConfig(loggers=<factory>, log_every_n_epochs=1, log_every_n_steps=0, step_average='batch', metrics=<factory>, monitor='val/loss', mode='min', val_every_n_epochs=1, save_every_n_epochs=0, early_stopping_patience=0, early_stopping_min_delta=0.0, early_stopping_threshold=None, progress_bar=False)[source]#

Bases: object

Logging, error metrics, checkpointing and stopping of a training run.

Variables:

  • loggers (list of str) – Where the metrics go: any of console, csv (metrics.csv in output_dir) and tensorboard (output_dir/tensorboard). Defaults to ["console"].

  • log_every_n_epochs (int) – Log the epoch metrics every n epochs (and after the last). Defaults to 1.

  • log_every_n_steps (int) – When > 0, also log the training values every n optimizer steps. Defaults to 0.

  • step_average (str) – What a step value is: "batch" (that step’s batch) or "running" (the epoch so far). Defaults to "batch".

  • metrics (list of str) – Error metrics computed besides the loss terms, from mae and rmse, for every labelled property (energy per atom, forces, stress, dipole, polarizability, charges). Defaults to both.

  • monitor (str) – The logged value that picks best.pt, drives the plateau scheduler and early stopping, e.g. val/forces_mae. A val/ monitor falls back to train/ without a validation set. Defaults to "val/loss".

  • mode (str) – "min" or "max": whether a lower or a higher monitored value is better. Defaults to "min".

  • val_every_n_epochs (int) – Validate every n epochs (and after the last). Defaults to 1.

  • save_every_n_epochs (int) – When > 0, also keep epoch_<n>.pt every n epochs. Defaults to 0.

  • early_stopping_patience (int) – Stop after this many validations without improvement of the monitored value; 0 (default) never stops early.

  • early_stopping_min_delta (float) – The least change that counts as an improvement. Defaults to 0.0.

  • early_stopping_threshold (float, optional) – Stop as soon as the monitored value is better than this. Defaults to None.

  • progress_bar (bool) – Show a progress bar over each epoch’s batches. Defaults to False.

Parameters:

  • loggers (list[str])

  • log_every_n_epochs (int)

  • log_every_n_steps (int)

  • step_average (str)

  • metrics (list[str])

  • monitor (str)

  • mode (str)

  • val_every_n_epochs (int)

  • save_every_n_epochs (int)

  • early_stopping_patience (int)

  • early_stopping_min_delta (float)

  • early_stopping_threshold (float | None)

  • progress_bar (bool)

validate()[source]#

Normalize the list options and reject unknown names.

Return type:

None