Augmentation & regularization

What

The training pipeline can apply torchvision v2 transforms, MixUp and CutMix, label smoothing, dropout, weight decay, and early stopping. All of them are declared in YAML; the training loop receives tensors and never builds a transform.

Why

Augmentation and regularization are the difference between memorizing a small training set and learning a decision boundary that transfers. They are also easy to apply inconsistently. Declaring them in one place, and recording the resolved config in every run, makes experiments comparable.

Config keys

Knob	Config key	Notes
Train transforms	data.transforms.train	ordered list of named transforms
Eval transforms	data.transforms.eval	usually resize, tensor, normalize
MixUp	training.mixup_alpha	0.0 disables
CutMix	training.cutmix_alpha	0.0 disables
Label smoothing	training.loss.label_smoothing	[0, 1)
Dropout	model.params.dropout	model-specific
Weight decay	training.optimizer.weight_decay	optimizer-specific
Early stopping	training.early_stopping	monitor, mode, patience

Supported transform names include RandomHorizontalFlip, RandomVerticalFlip, RandomRotation, RandomCrop, RandomResizedCrop, ColorJitter, GaussianBlur, RandAugment, AutoAugment, RandomErasing, ToTensor, and Normalize. Normalization can use auto to take the run's statistics.

Implementation

data.transforms builds a callable per sample from the ordered config list. training.losses handles label smoothing and MixUp/CutMix; training.callbacks handles early stopping. The reference CIFAR baseline uses horizontal flip, random crop with padding 4, mild color jitter, and random erasing.

Evidence

• The CIFAR-10 scratch baseline reaches 92.25% test accuracy with the reference augmentation family (see the CIFAR-10 case study).
• On the reduced 8-epoch ablation budget, augmentation off reached 71.06% test accuracy versus 67.92% with the crop/flip pipeline. That is a measured outcome on one seed and a small budget, not a general recommendation.
• The optical degradation demo deliberately does not use flips or crops: motion-blur direction and vignetting position are part of the label.

Caveat

More augmentation is not automatically better, and it can destroy orientation-sensitive labels. The ablations report paired differences on one reduced budget, not universal recommendations. The CIFAR baseline is a reproduction, not a state-of-the-art result.