Domain Adaptation on ImageNet Robustness Sets
45.8Accuracy (ImageNet-Sketch)PACE
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| PACEBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K, Base Method=LoRA_mul + VPT_add2024.09 | 45.8 | 71.2 | 24.6 | 43.6 | |
| Linear ProbingBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 43.2 | 69.5 | 23.4 | 40.9 | |
| LoRA_mulBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 41.9 | 68.8 | 22.6 | 37.8 | |
| LoRA_mul + VPT_addBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 41.5 | 68.6 | 21.6 | 38.2 | |
| LoRA_add + VPT_addBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 41.4 | 68.3 | 20.6 | 36.9 | |
| VPT_addBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 40.4 | 68.7 | 22.4 | 38.4 | |
| LoRA_addBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 39.8 | 67.8 | 20.4 | 35.6 | |
| PACEBackbone=Swin-B, Pre-training=ImageNet-21K, Base Method=LoRA_mul + VPT_add2024.09 | 39.2 | 70.1 | 25.2 | 38 | |
| Linear ProbingBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 36.7 | 68.8 | 23.2 | 35.9 | |
| PACEBackbone=ViT-B, Pre-training=ImageNet-21K, Base Method=LoRA_mul + VPT_add2024.09 | 31.8 | 69.4 | 16.3 | 35.2 | |
| LoRA_mulBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 31.2 | 68.3 | 13.4 | 32.7 | |
| LoRA_mul + VPT_addBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 31.2 | 66.5 | 18.5 | 29.4 | |
| LoRA_mulBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 30.8 | 65.7 | 18.1 | 28.9 | |
| LoRA_addBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 30.7 | 65.7 | 16.8 | 28.9 | |
| LoRA_add + VPT_addBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 30.7 | 65.7 | 16.8 | 28.9 | |
| LoRA_mul + VPT_addBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 30.6 | 68.5 | 14.1 | 32.5 | |
| VPT_addBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 30.6 | 66.2 | 17.6 | 29.1 | |
| Full Fine-tuningBackbone=ViT-B, Pre-training=Laion2B-ImageNet-12K2024.09 | 29 | 56.1 | 8.1 | 27.9 | |
| LoRA_add + VPT_addBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 28.7 | 66.6 | 13.7 | 29.9 | |
| LoRA_addBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 27.1 | 64.8 | 13.6 | 25 | |
| VPT_addBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 27.1 | 65.9 | 11.5 | 26.7 | |
| Full Fine-tuningBackbone=Swin-B, Pre-training=ImageNet-21K2024.09 | 27 | 61.1 | 10.8 | 24.4 | |
| Full Fine-tuningBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 18.5 | 52.5 | 3.2 | 21.2 | |
| Linear ProbingBackbone=ViT-B, Pre-training=ImageNet-21K2024.09 | 14.4 | 60.8 | 9.4 | 25.6 |