Image Classification on ImageNet-1K 1.0 (val) (Top-1, Top-5, and Inference Efficiency Metrics)
88.94Top-1 AccToaSt
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| ToaStBackbone=ViT-MAE-Large, Hardware=H100 GPU, Batch size=1282026.02 | 88.94 | 97.95 | 38.5 | 37.5 | 527 | 1.51 | |
| ToaStBackbone=ViT-MAE-Huge, Hardware=H100 GPU, Batch size=1282026.02 | 88.52 | 98.29 | 101.4 | 39.4 | 206.21 | 1.59 | |
| BaselineBackbone=ViT-MAE-Huge, Hardware=H100 GPU, Batch size=1282026.02 | 86.88 | 98.07 | 167.4 | — | 129.7 | 1 | |
| DiffRateBackbone=ViT-MAE-Huge, Hardware=H100 GPU, Batch size=1282026.02 | 86.65 | 97.88 | 103.4 | 38.2 | 202.9 | 1.56 | |
| ToMeBackbone=ViT-MAE-Huge, r=5, Hardware=H100 GPU, Batch size=1282026.02 | 86.28 | 97.88 | 113.9 | 31.9 | 185.9 | 1.43 | |
| BaselineBackbone=ViT-MAE-Large, Hardware=H100 GPU, Batch size=1282026.02 | 85.96 | 97.55 | 61.6 | — | 349 | 1 | |
| DiffRateBackbone=ViT-MAE-Large, Hardware=H100 GPU, Batch size=1282026.02 | 85.66 | 97.44 | 42.3 | 31.3 | 474.7 | 1.36 | |
| ToaStBackbone=Swin-Base, Hardware=H100 GPU, Batch size=1282026.02 | 85.21 | 96.5 | 8.8 | 42.7 | 1,408.6 | 1.28 | |
| ToaStBackbone=DeiT-Base, Hardware=H100 GPU, Batch size=1282026.02 | 84.82 | 97.1 | 10.7 | 39.2 | 1,690.9 | 1.51 | |
| ToaStBackbone=Swin-Small, Hardware=H100 GPU, Batch size=1282026.02 | 84.65 | 96.8 | 5.4 | 38.2 | 1,909.5 | 1.24 | |
| ToMeBackbone=ViT-MAE-Large, r=6, Hardware=H100 GPU, Batch size=1282026.02 | 84.58 | 97.12 | 38.5 | 37.5 | 523.2 | 1.5 | |
| ToaStBackbone=ViT-MAE-Base, Hardware=H100 GPU, Batch size=1282026.02 | 84.13 | 96.39 | 11 | 37.5 | 1,692.6 | 1.48 | |
| BaselineBackbone=ViT-MAE-Base, Hardware=H100 GPU, Batch size=1282026.02 | 83.75 | 96.54 | 17.6 | — | 1,140.2 | 1 | |
| BaselineBackbone=Swin-Base, Hardware=H100 GPU, Batch size=1282026.02 | 83.5 | 96.5 | 15.4 | — | 1,100.1 | 1 | |
| ToaStBackbone=DeiT-Small, Hardware=H100 GPU, Batch size=1282026.02 | 83.4 | 96.97 | 2.5 | 45.7 | 4,783.3 | 2.07 | |
| BaselineBackbone=Swin-Small, Hardware=H100 GPU, Batch size=1282026.02 | 83.2 | 96.2 | 8.7 | — | 1,534.4 | 1 | |
| STViT-RBackbone=Swin-Base, Hardware=H100 GPU, Batch size=1282026.02 | 83.2 | 96.4 | 10.3 | 33.1 | 1,206.2 | 1.1 | |
| DiffRateBackbone=ViT-MAE-Base, Hardware=H100 GPU, Batch size=1282026.02 | 82.9 | 96.14 | 11.5 | 34.7 | 1,552.8 | 1.36 | |
| STViT-RBackbone=Swin-Small, Hardware=H100 GPU, Batch size=1282026.02 | 82.6 | 96.07 | 5.8 | 33.3 | 1,646.6 | 1.07 | |
| ToMeBackbone=ViT-MAE-Base, r=13, Hardware=H100 GPU, Batch size=1282026.02 | 81.87 | 96.02 | 10.4 | 40.9 | 1,783.3 | 1.56 | |
| BaselineBackbone=DeiT-Base, Hardware=H100 GPU, Batch size=1282026.02 | 81.8 | 95.6 | 17.6 | — | 1,122.9 | 1 | |
| ToaStBackbone=Swin-Tiny, Hardware=H100 GPU, Batch size=1282026.02 | 81.76 | 95.7 | 3.1 | 31.3 | 2,705.8 | 1.04 | |
| DiffRateBackbone=DeiT-Base, Hardware=H100 GPU, Batch size=1282026.02 | 81.51 | 95.4 | 11.5 | 34.7 | 1,553.9 | 1.38 | |
| BaselineBackbone=Swin-Tiny, Hardware=H100 GPU, Batch size=1282026.02 | 81.2 | 95.5 | 4.5 | — | 2,610.9 | 1 | |
| ToMeBackbone=DeiT-Base, Hardware=H100 GPU, Batch size=1282026.02 | 80.59 | 94.83 | 11.5 | 34.7 | 1,628.4 | 1.45 | |
| BaselineBackbone=DeiT-Small, Hardware=H100 GPU, Batch size=1282026.02 | 79.82 | 94.95 | 4.6 | — | 2,313.2 | 1 | |
| DiffRateBackbone=DeiT-Small, Hardware=H100 GPU, Batch size=1282026.02 | 79.56 | 94.8 | 2.9 | 37 | 2,808.1 | 1.21 | |
| ToMeBackbone=DeiT-Small, Hardware=H100 GPU, Batch size=1282026.02 | 79.35 | 94.65 | 2.7 | 41.3 | 2,737.1 | 1.18 | |
| ResNet-101 + DLA-LBackbone=ResNet-101, Params(M)=47.8, FLOPs(G)=8.12026.06 | 78.9 | 94.5 | — | — | — | — | |
| ResNet-101 + KCLABackbone=ResNet-101, Params(M)=45.0, FLOPs(G)=8.02026.06 | 78.8 | 94.4 | — | — | — | — | |
| ResNet-101 + AABackbone=ResNet-101, Params(M)=47.6, FLOPs(G)=8.62026.06 | 78.7 | 94.4 | — | — | — | — | |
| ResNet-101 + ECABackbone=ResNet-101, Params(M)=44.5, FLOPs(G)=7.82026.06 | 78.7 | 94.3 | — | — | — | — | |
| ResNet-101 + MRLA-LBackbone=ResNet-101, Params(M)=44.9, FLOPs(G)=7.92026.06 | 78.7 | 94.4 | — | — | — | — | |
| ResNet-101 + CBAMBackbone=ResNet-101, Params(M)=49.3, FLOPs(G)=7.92026.06 | 78.5 | 94.3 | — | — | — | — | |
| ResNet-101 + RLAgBackbone=ResNet-101, Params(M)=45.0, FLOPs(G)=8.42026.06 | 78.5 | 94.2 | — | — | — | — | |
| ResNet-50 + DLA-LBackbone=ResNet-50, Params(M)=27.2, FLOPs(G)=4.32026.06 | 78 | 94 | — | — | — | — | |
| ResNet-50 + KCLABackbone=ResNet-50, Params(M)=25.8, FLOPs(G)=4.22026.06 | 77.8 | 93.8 | — | — | — | — | |
| ResNet-50 + AABackbone=ResNet-50, Params(M)=27.1, FLOPs(G)=4.52026.06 | 77.7 | 93.8 | — | — | — | — | |
| ResNet-50 + MRLA-LBackbone=ResNet-50, Params(M)=25.7, FLOPs(G)=4.22026.06 | 77.7 | 93.8 | — | — | — | — | |
| ResNet-101 + SEBackbone=ResNet-101, Params(M)=49.3, FLOPs(G)=7.82026.06 | 77.6 | 93.9 | — | — | — | — | |
| ResNet-50 + ECABackbone=ResNet-50, Params(M)=25.6, FLOPs(G)=4.12026.06 | 77.5 | 93.7 | — | — | — | — | |
| ResNet-101Backbone=ResNet-101, Params(M)=44.5, FLOPs(G)=7.82026.06 | 77.4 | 93.5 | — | — | — | — | |
| ResNet-50 + CBAMBackbone=ResNet-50, Params(M)=28.1, FLOPs(G)=4.22026.06 | 77.3 | 93.7 | — | — | — | — | |
| ResNet-50 + DIANetBackbone=ResNet-50, Params(M)=28.42026.06 | 77.2 | — | — | — | — | — | |
| ResNet-50 + RLAgBackbone=ResNet-50, Params(M)=25.9, FLOPs(G)=4.52026.06 | 77.2 | 93.4 | — | — | — | — | |
| ResNet-50 + SEBackbone=ResNet-50, Params(M)=28.1, FLOPs(G)=4.12026.06 | 76.7 | 93.4 | — | — | — | — | |
| ResNet-50Backbone=ResNet-50, Params(M)=25.6, FLOPs(G)=4.12026.06 | 76.1 | 92.9 | — | — | — | — | |
| ToaStBackbone=DeiT-Tiny, Hardware=H100 GPU, Batch size=1282026.02 | 74.25 | 92.65 | 0.76 | 38.5 | 4,249.7 | 2.03 | |
| BaselineBackbone=DeiT-Tiny, Hardware=H100 GPU, Batch size=1282026.02 | 72.2 | 91.1 | 1.3 | — | 2,090.9 | 1 | |
| DiffRateBackbone=DeiT-Tiny, Hardware=H100 GPU, Batch size=1282026.02 | 71.78 | 90.87 | 0.9 | 30.8 | 2,422.5 | 1.16 | |
| ToMeBackbone=DeiT-Tiny, Hardware=H100 GPU, Batch size=1282026.02 | 71.25 | 90.74 | 0.7 | 46.2 | 2,484.6 | 1.19 |