What it does

This repo implements Attention Rollout and Gradient Attention Rollout for Vision Transformers. Feed it an image and it generates a heatmap showing which patches the model attended to. The gradient variant goes further: it multiplies attention weights by class-specific gradients, masking out negative contributions so you see only the attention that actually drove a particular classification decision.

The interesting bit

The authors didn’t just port a paper. They found empirically that the standard “average across heads” recipe from the original Attention Rollout paper is often worse than taking the minimum or maximum attention value, especially when you also discard the weakest 90% of activations. It’s a small, honest tweak that makes the visualizations noticeably sharper.

Key highlights

• Two methods: vanilla Attention Rollout (class-agnostic) and Gradient Attention Rollout (class-specific)
• Three head fusion strategies: mean, min, max — configurable per-run
• discard_ratio parameter filters low-attention noise layer by layer
• Works out of the box with torch.hub models (default: DeiT-Tiny)
• Command-line tool plus clean Python API for dropping into notebooks

Caveats

• “Attention flow is work in progress” — one of the three referenced methods isn’t implemented yet
• Only requires timm, but the repo itself is a thin wrapper; you’ll need to bring your own model if you stray from the DeiT default

Verdict

Worth a look if you’re debugging ViT behavior or writing a paper that needs interpretability baselines. Skip it if you need explanations for CNNs or a fully packaged library — this is research code with a narrow, useful scope.