CutMix

CutMix teaches a model not to treat every training example as a perfectly intact, isolated object. Instead, it creates a new example by cutting a rectangular region from one input and pasting it into another, then tells the model that both original labels are relevant.

How the mixed example is labelled
For two training examples, A and B, CutMix replaces a patch of A with a patch from B. Its target label becomes a weighted combination of their labels, where the weights come from the patch area—not from a guess about which content looks more important. If 30% of A’s area is replaced, the target assigns roughly 70% weight to A’s label and 30% to B’s label.

• Input: mostly A, with a rectangular patch from B.
• Target: λ · label(A) + (1 − λ) · label(B), where λ is the fraction of A left visible.
• Training only: at inference, the network receives normal, unmixed inputs.

Why it improves training
A high-capacity network can memorize superficial cues in clean training examples. CutMix disrupts that shortcut: the model must recognize evidence wherever it appears and distribute confidence between multiple sources. Unlike Mixup, which blends every input value together, CutMix preserves untouched regions exactly. In an image classifier, a pasted patch can retain sharp edges and realistic local texture, while still forcing the prediction to reflect both examples. This can improve robustness and reduce overconfident predictions.

Practical trade-offs
CutMix is applied per batch by pairing examples and sampling patch locations and sizes. It adds little model compute, but increases input-pipeline work and makes loss curves noisier because targets are deliberately soft. Very large patches can make labels poorly matched to semantic content—for example, a patch may cover background rather than the object—while very weak mixing provides little regularization. It is commonly used with convolutional networks and can be implemented before a PyTorch model’s forward pass, alongside standard augmentations.