Research map / Architectures and efficiency / Attention and transformers
Pruning Self-Attentions Into Convolutional Layers in Single Path
Haoyu He, Jianfei Cai, Jing Liu, Zizheng Pan, Jing Zhang, Dacheng Tao, Bohan Zhuang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| single-path instead of multi-path search (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 70.5 -> 70.7 | Improved |
| single-path instead of multi-path search (ablation) | SPViT-DeiT-S · ImageNet-1k val | Top-1 accuracy: 78.0 -> 78.3 | Improved |
| add MSA pruning to FFN-only pruning (ablation) | SPViT-DeiT-B · ImageNet-1k val | Top-1 accuracy: 81.3 -> 81.5 | Improved |
| add hard-label distillation with RegNetY-16GF teacher (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 70.7 -> 73.2 | Improved |
| add hard-label distillation (ablation) | SPViT-DeiT-B · ImageNet-1k val | Top-1 accuracy: 81.5 -> 82.4 | Improved |
| use uncompressed DeiT-Ti as teacher (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 70.7 -> 71.1 | No clear effect |
| use CNN teacher RegNetY-1.6GF instead of DeiT-Ti teacher (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 71.1 -> 72.8 | Improved |
| use higher capacity CNN teacher RegNetY-16GF (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 72.8 -> 73.2 | Improved |
| CNN teacher RegNetY-1.6GF instead of DeiT-Ti teacher (ablation) | UVC-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 71.3 -> 70.5 | Hurt |
| search MSA-to-conv choice instead of random sampling (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 70.2 -> 70.7 | Improved |
| search FFN ratios instead of random sampling (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 70.4 -> 70.6 | Improved |
| prune attention into convolution instead of random head pruning (ablation) | SPViT-DeiT-Ti · ImageNet-1k val | Top-1 accuracy: 69.6 -> 70.7 | Improved |
Closest papers by meaning
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.91)
- MaxViT: Multi-axis Vision Transformer (ECCV 2022 · similarity 0.89)
- LPViT: Low-Power Semi-structured Pruning for Vision Transformers (ECCV 2024 · similarity 0.89)
- Data-independent Module-aware Pruning for Hierarchical Vision Transformers (ICLR 2024 · similarity 0.89)
- FasterViT: Fast Vision Transformers with Hierarchical Attention (ICLR 2024 · similarity 0.89)
- You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
