Image Classification on ImageNet-1k 2012 (val)
84.4Top-1 AccuracyViT-B/8
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ViT-B/8Arch.=ViT-B/8, Res.=224, SS PT=NNCLR, Reg. / Extra Data=DeIT Aug.2021.04 | 84.4 | — | — | |
| ViT-B/16Arch.=ViT-B/16, Res.=384, Reg. / Extra Data=JFT-300M2021.04 | 84.2 | — | — | |
| NesT-BArch. base=Transformer local-attention, #Params=68M, Image size=224, Training initialization=random initialization2021.05 | 83.8 | — | — | |
| ViT-B/16Arch.=ViT-B/16, Res.=384, SS PT=NNCLR, Reg. / Extra Data=DeIT Aug.2021.04 | 83.7 | — | — | |
| Swin-BArch. base=Transformer local-attention, #Params=88M, Image size=224, Training initialization=random initialization2021.05 | 83.3 | — | — | |
| NesT-SArch. base=Transformer local-attention, #Params=38M, Image size=224, Training initialization=random initialization2021.05 | 83.3 | — | — | |
| ViT-B/16Arch.=ViT-B/16, Res.=384, Reg. / Extra Data=DeIT Aug.2021.04 | 83.1 | — | — | |
| ViT-B/8Arch.=ViT-B/8, Res.=224, Reg. / Extra Data=DeIT Aug.2021.04 | 83.1 | — | — | |
| Swin-SArch. base=Transformer local-attention, #Params=50M, Image size=224, Training initialization=random initialization2021.05 | 83 | — | — | |
| RegNetY-16GArch. base=Convolutional, #Params=84M, Image size=224, Training initialization=random initialization2021.05 | 82.9 | — | — | |
| ViT-S/16Arch.=ViT-S/16, Res.=224, SS PT=NNCLR, Reg. / Extra Data=DeIT Aug.2021.04 | 82.7 | — | — | |
| ViT-B/16Arch.=ViT-B/16, Res.=224, SS PT=NNCLR, Reg. / Extra Data=DeIT Aug.2021.04 | 82.5 | — | — | |
| ViT-B/16Arch.=ViT-B/16, Res.=224, Reg. / Extra Data=DeIT Aug.2021.04 | 81.8 | — | — | |
| DeiT-BArch. base=Transformer full-attention, #Params=86M, Image size=224, Training initialization=random initialization2021.05 | 81.8 | — | — | |
| NesT-TArch. base=Transformer local-attention, #Params=17M, Image size=224, Training initialization=random initialization2021.05 | 81.5 | — | — | |
| Swin-TArch. base=Transformer local-attention, #Params=29M, Image size=224, Training initialization=random initialization2021.05 | 81.3 | — | — | |
| RegNetY-4GArch. base=Convolutional, #Params=21M, Image size=224, Training initialization=random initialization2021.05 | 80 | — | — | |
| ViT-S/16Arch.=ViT-S/16, Res.=224, Reg. / Extra Data=DeIT Aug.2021.04 | 79.8 | — | — | |
| DeiT-SArch. base=Transformer full-attention, #Params=22M, Image size=224, Training initialization=random initialization2021.05 | 79.8 | — | — | |
| ResNet50Arch.=ResNet50, Res.=224, SS PT=NNCLR2021.04 | 79.1 | — | — | |
| ResNet50Arch.=ResNet50, Res.=224, Reg. / Extra Data=JFT-300M2021.04 | 79 | — | — | |
| ViT-B/16Arch.=ViT-B/16, Res.=3842021.04 | 77.9 | — | — | |
| ViT-B/16Arch. base=Transformer full-attention, #Params=86M, Image size=384, Training initialization=random initialization2021.05 | 77.9 | — | — | |
| ResNet50Arch.=ResNet50, Res.=2242021.04 | 76.2 | — | — | |
| ResNet-50Arch. base=Convolutional, #Params=25M, Image size=224, Training initialization=random initialization2021.05 | 76.2 | — | — | |
| RegNetNumber of Layers=50, Params=31.3M, FLOPs=5.12G2021.01 | — | 23.43 | 6.93 | |
| ResNetNumber of Layers=502021.01 | — | 24.7 | 7.8 | |
| ResNet*Number of Layers=50, Evaluation Protocol=re-implementation by our experiments, Params=26.6M, FLOPs=4.14G2021.01 | — | 24.81 | 7.78 |