Reproduce “Scaling Vision Transformers Without Training: Prune-and-Merge Token Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vatsal, Soumil Malhotra, Priyanshu Chahal, Sanjeev Kumar
Prune-and-Merge ViT reduces inference work in Vision Transformers by pruning patches selected through class-token attention, then merging similar survivors. On ImageNet-100, ViT-Base reaches 236.85 images per second at composite keep rate 0.25, versus 78.61 uncompressed, on one NVIDIA T4 GPU. At that rate, merging-heavy allocation improves accuracy over balanced and pruning-heavy allocations, but results vary by model and prediction stability declines.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.