Research map / Architectures and efficiency / Attention and transformers
AutoFormer: Searching Transformers for Visual Recognition
Minghao Chen, Houwen Peng, Jianlong Fu, Haibin Ling
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 |
|---|---|---|---|
| use weight entanglement instead of classical weight sharing, with evolution search (ablation) | AutoFormer-S supernet · ImageNet val | top-1 accuracy (inherited): 71.5% -> 81.7% | Improved |
| use weight entanglement instead of classical weight sharing, with evolution search (ablation) | AutoFormer-S · ImageNet val | top-1 accuracy (retrain): 80.4% -> 81.7% | Improved |
| use weight entanglement instead of classical weight sharing, with random search (ablation) | AutoFormer-S · ImageNet val | top-1 accuracy (inherited): 69.7% -> 81.3% | Improved |
| use evolution search instead of random search with weight entanglement (ablation) | AutoFormer-S · ImageNet val | top-1 accuracy (inherited): 81.3% -> 81.7% | Improved |
| use supernet-based search instead of plain random search (ablation) | AutoFormer-S · ImageNet val | top-1 accuracy (retrain): 79.4% -> 81.7% | Improved |
| finetune searched subnet 40 epochs after inheriting weights | AutoFormer-T · ImageNet val | top-1 accuracy: 74.7% -> 74.9% | No clear effect |
| finetune searched subnet 40 epochs after inheriting weights | AutoFormer-B · ImageNet val | top-1 accuracy: 82.4% -> 82.6% | No clear effect |
| retrain searched subnet from scratch for 300 epochs | AutoFormer-S · ImageNet val | top-1 accuracy: 81.7% -> 81.7% | No clear effect |
| add knowledge distillation from RegNetY-32GF teacher | AutoFormer-T/S/B · ImageNet val | top-1 accuracy: 74.7%/81.7%/82.4% -> 75.7%/82.4%/82.9% | Improved |
| fine-tune at 384 resolution | AutoFormer-S · ImageNet val | top-1 accuracy: 81.7% -> 83.4 | Improved |
Later papers that compared with this method
1 later accepted paper reports AutoFormer-tiny as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Searching for BurgerFormer with Micro-Meso-Macro Space Design (ICML 2022 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Masked Distillation Advances Self-Supervised Transformer Architecture Search (ICLR 2024 · similarity 0.90)
- ViTAS: Vision Transformer Architecture Search (ECCV 2022 · similarity 0.89)
- Super Vision Transformer (IJCV 2023 · similarity 0.89)
- Searching the Search Space of Vision Transformer (NeurIPS 2021 · similarity 0.89)
- Auto-Prox: Training-Free Vision Transformer Architecture Search via Automatic Proxy Discovery (AAAI 2024 · similarity 0.89)
- A Survey on Vision Transformer (TPAMI 2022 · 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.
