Masked Attention Alignment for Data-Free ViT Quantization

Arxiv CS.CV (computer vision)

The paper introduces Masked Attention Alignment, a data-free quantization method for Vision Transformers that synthesizes samples without accessing real data. It leverages selective coupling of decoupled informative regions to generate effective synthetic data.

vision-transformer

Arxiv CS.CV (computer vision)·Biao QianYang WangYong WuJungong Han

7 days ago