VGGBaseline

class torch_uncertainty.baselines.classification.VGGBaseline(num_classes, in_channels, loss, version, arch, style='imagenet', num_estimators=1, dropout_rate=0.0, last_layer_dropout=False, optim_recipe=None, mixup_params=None, groups=1, alpha=None, gamma=1, ood_criterion='msp', log_plots=False, save_in_csv=False, calibration_set='val', eval_ood=False, eval_shift=False, eval_grouping_loss=False)

VGG backbone baseline for classification providing support for various versions and architectures.

Parameters:

• num_classes (int) – Number of classes to predict.

• in_channels (int) – Number of input channels.

• loss (nn.Module) – Training loss.

• optim_recipe (Any) – optimization recipe, corresponds to what expect the LightningModule.configure_optimizers() method.

• version (str) –

  Determines which VGG version to use:

  • "std": original VGG

  • "mc-dropout": Monte Carlo Dropout VGG

  • "packed": Packed-Ensembles VGG

• arch (int) –

  Determines which VGG architecture to use:

  • 11: VGG-11

  • 13: VGG-13

  • 16: VGG-16

  • 19: VGG-19

• style (str, optional) – Which VGG style to use. Defaults to

• imagenet. –

• num_estimators (int, optional) – Number of estimators in the ensemble. Only used if version is either "packed", "batched" or "masked" Defaults to None.

• dropout_rate (float, optional) – Dropout rate. Defaults to 0.0.

• mixup_params (dict, optional) – Mixup parameters. Can include mixtype, mixmode, dist_sim, kernel_tau_max, kernel_tau_std, mixup_alpha, and cutmix_alpha. If None, no augmentations. Defaults to None.

• last_layer_dropout (bool) – whether to apply dropout to the last layer only.

• groups (int, optional) – Number of groups in convolutions. Defaults to 1.

• alpha (float, optional) – Expansion factor affecting the width of the estimators. Only used if version is "packed". Defaults to None.

• gamma (int, optional) – Number of groups within each estimator. Only used if version is "packed" and scales with groups. Defaults to 1s.

• ood_criterion (str, optional) – OOD criterion. Defaults to "msp". MSP is the maximum softmax probability, logit is the maximum logit, entropy is the entropy of the mean prediction, mi is the mutual information of the ensemble and vr is the variation ratio of the ensemble.

• log_plots (bool, optional) – Indicates whether to log the plots or not. Defaults to False.

• save_in_csv (bool, optional) – Indicates whether to save the results in a csv file or not. Defaults to False.

• calibration_set (Callable, optional) – Calibration set. Defaults to None.

• eval_ood (bool, optional) – Indicates whether to evaluate the OOD detection or not. Defaults to False.

• eval_shift (bool) – Whether to evaluate on shifted data. Defaults to False.

• eval_grouping_loss (bool, optional) – Indicates whether to evaluate the grouping loss or not. Defaults to False.

Raises:

ValueError – If version is not either "std", "packed", "batched" or "masked".

Returns:

VGG baseline ready for training and evaluation.

Return type:

LightningModule