Research map / Architectures and efficiency / Attention and transformers
Tokens-to-Token ViT: Training Vision Transformers From Scratch on ImageNet
Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis E.H. Tay, Jiashi Feng, Shuicheng Yan
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 |
|---|---|---|---|
| add T2T module (progressive tokenization) (ablation) | T2T-ViT-14 backbone · ImageNet-1k val (from scratch) | top-1 accuracy (%): 79.5 -> 81.5 | Improved |
| use Transformer layer instead of Performer in T2T module (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 81.5 -> 81.7 | Improved |
| use T2T soft-split module instead of 3 convolution layers (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 80.8 -> 81.5 | Improved |
| use deep-narrow instead of shallow-wide backbone (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 78.8 -> 81.5 | Improved |
| use deep-narrow (16 layers, 384 dim) instead of baseline (ablation) | ViT-S/16 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 78.1 -> 79.0 | Improved |
| use shallow-wide (4 layers, 1024 dim) structure (ablation) | ViT-S/16 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 78.1 -> 69.9 | Hurt |
| add dense connection (ablation) | ViT-S/16 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 78.1 -> 76.8 | Hurt |
| add dense connection (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 81.5 -> 80.6 | Hurt |
| add SE channel attention (ablation) | ViT-S/16 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 78.1 -> 78.4 | Improved |
| add SE channel attention (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 81.5 -> 81.6 | No clear effect |
| use 32 heads (ResNeXt structure) (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 81.5 -> 81.5 | No clear effect |
| add Ghost operations (ablation) | T2T-ViT-14 · ImageNet-1k val (from scratch) | top-1 accuracy (%): 81.5 -> 79.5 | Hurt |
Closest papers by meaning
- EViT: Expediting Vision Transformers via Token Reorganizations (ICLR 2022 · similarity 0.93)
- Token Cropr: Faster ViTs for Quite a Few Tasks (CVPR 2025 · similarity 0.91)
- ImagePiece: Content-aware Re-tokenization for Efficient Image Recognition (AAAI 2025 · similarity 0.91)
- Learning to Merge Tokens via Decoupled Embedding for Efficient Vision Transformers (NeurIPS 2024 · similarity 0.90)
- Slicing Vision Transformer for Flexible Inference (NeurIPS 2024 · similarity 0.89)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient 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.
