MixupMixin

class fgvc.core.training.mixup_mixin.MixupMixin(model: Module, trainloader: DataLoader | None = None, *args, mixup: float = 0.0, cutmix: float = 0.0, mixup_prob: float = 1.0, **kwargs)

Mixin class that adds LR scheduler functionality to the trainer class.

The SchedulerMixin supports PyTorch and timm schedulers.

Parameters:

• model – Pytorch neural network. MixupMixin uses it to get number of classes.

• trainloader – Pytorch dataloader with training data. MixupMixin uses it to get number of classes.

• mixup – Mixup alpha value, mixup is active if > 0.

• cutmix – Cutmix alpha value, cutmix is active if > 0.

• mixup_prob – Probability of applying mixup or cutmix per batch.

apply_mixup(imgs: Tensor, targs: Tensor) → Tuple[Tensor, Tensor]

Apply mixup or cutmix method if arguments mixup or cutmix were used in Trainer.