Research map / Architectures and efficiency / CNN design
Efficient Training of Visual Transformers with Small Datasets
Yahui Liu, Enver Sangineto, Wei Bi, Nicu Sebe, Bruno Lepri, Marco De Nadai
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 dense relative localization loss | Swin-T · CIFAR-100, 100 epochs, from scratch | Top-1 accuracy: 53.28 -> 66.23 | Improved |
| Add dense relative localization loss | Swin-T · Quickdraw, 100 epochs, from scratch | Top-1 accuracy: 24.08 -> 69.41 | Improved |
| Add dense relative localization loss | Swin-T · ImageNet-100, 100 epochs | Top-1 accuracy: 82.66 -> 83.95 | Improved |
| Add dense relative localization loss during fine-tuning | Swin-T · Infograph, 100 epochs, fine-tuning | Top-1 accuracy: 41.07 -> 46.10 | Improved |
| Add dense relative localization loss during pre-training | Swin-T · ADE20K validation, ImageNet-1K pre-training | mIoU: 43.87 -> 44.33 | Improved |
| Use signed target offsets instead of nonnegative offsets (ablation) | Swin-T dense relative localization · ImageNet-100, 100 epochs | Top-1 accuracy: 84.02 -> 83.14 | Hurt |
| Replace regression with cross-entropy classification (ablation) | Swin-T dense relative localization · ImageNet-100, 100 epochs | Top-1 accuracy: 84.02 -> 83.86 | Hurt |
| Use Gaussian-regularized localization loss (ablation) | Swin-T dense relative localization · ImageNet-100, 100 epochs | Top-1 accuracy: 84.02 -> 83.24 | Hurt |
| Apply localization loss at every transformer block (ablation) | Swin-T dense relative localization · ImageNet-100, 100 epochs | Top-1 accuracy: 82.76 -> 81.88 | Hurt |
| Use 32 rather than 64 embedding pairs per image (ablation) | Swin-T dense relative localization · CIFAR-100, 100 epochs | Top-1 accuracy: 66.23 -> 63.70 | Hurt |
| Use 128 rather than 64 embedding pairs per image (ablation) | Swin-T dense relative localization · CIFAR-100, 100 epochs | Top-1 accuracy: 66.23 -> 65.16 | Hurt |
| Use 256 rather than 64 embedding pairs per image (ablation) | Swin-T dense relative localization · CIFAR-100, 100 epochs | Top-1 accuracy: 66.23 -> 64.87 | Hurt |
Closest papers by meaning
- Visual Transformers: Where Do Transformers Really Belong in Vision Models? (ICCV 2021 · similarity 0.92)
- Vision Transformers provably learn spatial structure (NeurIPS 2022 · similarity 0.90)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.90)
- Training Vision Transformers with Only 2040 Images (ECCV 2022 · similarity 0.90)
- An Empirical Study of Training Self-Supervised Vision Transformers (ICCV 2021 · similarity 0.89)
- Do Vision Transformers See Like Convolutional Neural Networks? (NeurIPS 2021 · 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.
