Image Classification on ImageNet and derived distribution shifts standard suite (test val)
88.3IN Accuracy (ref.)WiSE-FT
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| WiSE-FTBackbone=ALIGN, alpha=optimal2021.09 | 88.3 | 80.4 | 93.3 | 71.1 | 68.6 | 81 | 78.4 | 82.8 | |
| ALIGNProtocol=Standard fine-tuning2021.09 | 88.2 | 80.1 | 88.5 | 69.1 | 61 | 76.3 | 75 | 81.6 | |
| WiSE-FTBackbone=BASIC-L, alpha=0.52021.09 | 87.9 | 81.6 | 94.5 | 73.6 | 84.1 | 83.2 | 83.4 | 85.7 | |
| WiSE-FTBackbone=BASIC-L, alpha=optimal2021.09 | 87.9 | 82.1 | 96 | 76.5 | 84.9 | 86.5 | 85 | 86.2 | |
| BASIC-LProtocol=Standard fine-tuning2021.09 | 87.5 | 79.8 | 84.3 | 68 | 77.4 | 72.1 | 76.3 | 81.9 | |
| WiSE-FTBackbone=CLIP ViT-L/14@336px, alpha=optimal2021.09 | 87.1 | 79.5 | 90.3 | 65 | 72.1 | 81 | 77.4 | 81.9 | |
| CaRotBackbone=ViT-L/142026.05 | 87 | 79.3 | 88 | 62.7 | — | 72.7 | — | 75.6 | |
| WiSE-FTBackbone=CLIP ViT-L/14@336px, alpha=0.52021.09 | 86.8 | 79.5 | 89.4 | 64.7 | 71.1 | 79.9 | 76.9 | 81.8 | |
| SAE-FTBackbone=ViT-L/142026.05 | 86.5 | 78.6 | 88.8 | 63.5 | — | 73.1 | — | 76 | |
| StarFTBackbone=ViT-L/142026.05 | 86.4 | 78.9 | 88.7 | 63.2 | — | 73.8 | — | 76.2 | |
| WiSE-FTBackbone=ALIGN, alpha=0.52021.09 | 86.3 | 79.2 | 93 | 71.1 | 67.8 | 81 | 78.4 | 82.3 | |
| CAR-FTBackbone=ViT-L/142026.05 | 86.3 | 76.8 | 84.2 | 60 | — | 66.6 | — | 71.9 | |
| CLIP ViT-L/14@336pxProtocol=Standard fine-tuning2021.09 | 86.2 | 76.8 | 79.8 | 57.9 | 63.3 | 65.4 | 68.6 | 77.4 | |
| BASIC-MProtocol=Standard fine-tuning2021.09 | 86.2 | 77.8 | 84.9 | 64.3 | 75.3 | 63.7 | 73.2 | 79.7 | |
| WiSE-FTBackbone=BASIC-M, alpha=optimal2021.09 | 86.2 | 78.6 | 91.1 | 68.8 | 78 | 71.4 | 77.4 | 81.4 | |
| FLYPBackbone=ViT-L/142026.05 | 86.2 | 78.2 | 83.8 | 60.2 | — | 68.9 | — | 72.8 | |
| WiSE-FTBackbone=ViT-L/142026.05 | 86.1 | 78.1 | 88.5 | 63.6 | — | 72.9 | — | 75.8 | |
| WiSE-FTBackbone=BASIC-M, alpha=0.52021.09 | 85.6 | 78.5 | 90.2 | 68.6 | 78 | 71.1 | 77.3 | 81.4 | |
| BASIC-LProtocol=Zero-shot2021.09 | 85.6 | 80.5 | 95.7 | 76.2 | 82.3 | 85.7 | 84.1 | 84.8 | |
| WiSE-FTBackbone=JFT pre-trained ViT-H, alpha=optimal2021.09 | 85.4 | 77.8 | 89.3 | 64.5 | 66 | 66.6 | 72.5 | 78.6 | |
| JFT pre-trained ViT-HProtocol=Standard fine-tuning2021.09 | 85.1 | 77.6 | 84.9 | 62.8 | 63.1 | 60.8 | 69.8 | 77.6 | |
| FTBackbone=ViT-L/142026.05 | 84.7 | 75.3 | 75.4 | 54.4 | — | 55.7 | — | 65.2 | |
| WiSE-FTBackbone=JFT pre-trained ViT-H, alpha=0.52021.09 | 82.9 | 75.4 | 89.3 | 63.8 | 65.8 | 66.2 | 72.1 | 77.5 | |
| WiSE-FTBackbone=CLIP ViT-B/16, alpha=optimal2021.09 | 82.6 | 73.2 | 80.1 | 54.1 | 57.7 | 54.6 | 63.5 | 72.3 | |
| WiSE-FTBackbone=CLIP ViT-B/16, alpha=0.52021.09 | 81.7 | 72.8 | 78.7 | 53.9 | 57.3 | 52.2 | 63 | 72.3 | |
| BASIC-MProtocol=Zero-shot2021.09 | 81.4 | 74.1 | 90.6 | 67.4 | 73.5 | 66.7 | 74.5 | 78 | |
| CLIP ViT-B/16Protocol=Standard fine-tuning2021.09 | 81.3 | 70.9 | 65.6 | 46.3 | 49.6 | 36.7 | 53.8 | 67.5 | |
| CLIP ViT-L/14@336pxProtocol=Zero-shot2021.09 | 76.6 | 70.5 | 89 | 60.9 | 68.5 | 77.6 | 73.3 | 74.9 | |
| ALIGNProtocol=Zero-shot2021.09 | 76.4 | 70.1 | 92.1 | 67.9 | 67.2 | 75.9 | 74.6 | 75.5 | |
| Zero-shotBackbone=ViT-L/142026.05 | 75.6 | 69.9 | 87.9 | 59.6 | — | 70.8 | — | 72 | |
| JFT pre-trained ViT-HProtocol=Zero-shot2021.09 | 72.9 | 66.1 | 85.9 | 57 | 59.2 | 58.4 | 65.3 | 69.1 | |
| CLIP ViT-B/16Protocol=Zero-shot2021.09 | 68.3 | 61.9 | 77.6 | 48.2 | 53 | 49.8 | 58.1 | 63.2 |