Research map / Architectures and efficiency / Neural architecture search
Searching the Search Space of Vision Transformer
Minghao Chen, Kan Wu, Bolin Ni, Houwen Peng, Bei Liu, Jianlong Fu, Hongyang Chao, 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 |
|---|---|---|---|
| evolve search space from A to B (one step) (ablation) | ViT search space · ImageNet-1k, top searched arch | top-1 accuracy (%): 65.5 -> 78.0 | Improved |
| evolve search space from B to C (two steps) (ablation) | ViT search space · ImageNet-1k, top searched arch | top-1 accuracy (%): 78.0 -> 82.1 | Improved |
| evolve search space from A to C (ablation) | ViT search space · ImageNet-1k, top searched arch | top-1 accuracy (%): 65.5 -> 82.1 | Improved |
| use inherited supernet weights vs finetuning or retraining from scratch (ablation) | sampled subnets (10 random) · ImageNet-1k | top-1 accuracy | No clear effect |
| replace Swin-T with S3-T | backbone · ImageNet-1k | top-1 accuracy (%): 81.3 -> 82.1 | Improved |
| replace Swin-S with S3-S | backbone · ImageNet-1k | top-1 accuracy (%): 83.0 -> 83.7 | Improved |
| replace Swin-B with S3-B | backbone · ImageNet-1k | top-1 accuracy (%): 83.3 -> 84.0 | Improved |
| increase input resolution from 224 to 384 for S3-S | S3-S · ImageNet-1k | top-1 accuracy (%): 83.7 -> 84.5 | Improved |
| increase input resolution from 224 to 384 for S3-B | S3-B · ImageNet-1k | top-1 accuracy (%): 84.0 -> 84.7 | Improved |
| replace Swin-S with S3-S as detection backbone | Cascade Mask R-CNN 3x · COCO | box AP: 51.8 -> 52.6 | Improved |
| replace Swin-T with S3-T as detection backbone | Cascade Mask R-CNN 3x · COCO | box AP: 50.4 -> 50.5 | No clear effect |
| replace Swin-S with S3-S | UperNet · ADE20K | mIoU (%): 47.60 -> 48.04 | Improved |
Closest papers by meaning
- AutoFormer: Searching Transformers for Visual Recognition (ICCV 2021 · similarity 0.89)
- GLiT: Neural Architecture Search for Global and Local Image Transformer (ICCV 2021 · similarity 0.88)
- Training-Free Transformer Architecture Search (CVPR 2022 · similarity 0.87)
- Auto-Prox: Training-Free Vision Transformer Architecture Search via Automatic Proxy Discovery (AAAI 2024 · similarity 0.86)
- Vision Transformer Neural Architecture Search for Out-of-Distribution Generalization: Benchmark and Insights (NeurIPS 2024 · similarity 0.84)
- einspace: Searching for Neural Architectures from Fundamental Operations (NeurIPS 2024 · similarity 0.83)
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.
