Image Classification on ImageNet-1k 1.0 (val) (Top-1 and Top-5 Accuracy)
81.2Top-1 AccDeRy(50, 10)-FT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DeRy(50, 10)-FTParams (M)=40.41, FLOPs (G)=6.43, Training Epochs=100, Training Protocol=FULL-TUNING, Resolution=224x2242022.10 | 81.2 | 95.6 | |
| ELABackbone=DeiT-S2025.03 | 81 | 95.8 | |
| Swin-SParams (M)=49.61, FLOPs (G)=8.52, Training Epochs=100, Training Protocol=FULL-TUNING, Resolution=224x2242022.10 | 80.8 | 95.7 | |
| MRLA-LBackbone=DeiT-S2025.03 | 80.7 | 95.3 | |
| ResNet-50Backbone=ResNet-50, Context=Comparison with DeiT-S tier2025.03 | 79.9 | 95 | |
| DeRy(30, 6)-FTParams (M)=24.89, FLOPs (G)=4.47, Training Epochs=100, Training Protocol=FULL-TUNING, Resolution=224x2242022.10 | 79.6 | 94.8 | |
| ResNet-101Params (M)=44.55, FLOPs (G)=7.85, Training Epochs=100, Training Protocol=FULL-TUNING, Resolution=224x2242022.10 | 79 | 94.5 | |
| MRLA-LBackbone=PVTv2-B12025.03 | 78.9 | 94.9 | |
| ELABackbone=PVTv2-B12025.03 | 78.9 | 94.7 | |
| Swin-TParams (M)=28.29, FLOPs (G)=4.36, Training Epochs=100, Training Protocol=FULL-TUNING, Resolution=224x2242022.10 | 78.3 | 94.6 | |
| ResNet-50Backbone=ResNet-50, Context=Comparison with PVTv2-B1 tier2025.03 | 78.3 | 94.3 | |
| ResNet-50Params (M)=25.56, FLOPs (G)=4.12, Training Epochs=100, Training Protocol=FULL-TUNING, Resolution=224x2242022.10 | 76.8 | 93.3 | |
| MoCo-v3 + SERView=2+4 view, Protocol=Fine-tune2026.03 | 75.02 | 92.12 | |
| MoCo-v3 + E-SSLView=2+4 view, Protocol=Fine-tune2026.03 | 75 | 92.36 | |
| MoCo-v3 + SERView=2 view, Protocol=Fine-tune2026.03 | 74.33 | 91.79 | |
| MoCo-v3 + EquiModView=3 view, Protocol=Fine-tune2026.03 | 74.08 | 91.75 | |
| MoCo-v3 + STLView=2 view, Protocol=Fine-tune2026.03 | 73.9 | 91.49 | |
| MoCo-v3View=2 view, Protocol=Fine-tune2026.03 | 73.83 | 91.91 | |
| MoCo-v3 + AugSelfView=2 view, Protocol=Fine-tune2026.03 | 73.5 | 91.67 | |
| ELABackbone=DeiT-Ti2025.03 | 73.4 | 92.1 | |
| MRLA-LBackbone=DeiT-Ti2025.03 | 73 | 91.7 | |
| ResNet-50Backbone=ResNet-50, Context=Comparison with DeiT-Ti tier2025.03 | 72.6 | 91.1 | |
| MoCo-v3 + SERView=2+4 view, Protocol=MLP2026.03 | 70.99 | 89.72 | |
| ELABackbone=PVTv2-B02025.03 | 70.8 | 90.3 | |
| Hybrid Π-NetsImage Size=224x224, # par (M)=11.962023.03 | 70.74 | 89.548 | |
| MRLA-LBackbone=PVTv2-B02025.03 | 70.6 | 90 | |
| R-PolyNetsImage Size=224x224, # par (M)=12.382023.03 | 70.228 | 89.39 | |
| D-PolyNetsImage Size=224x224, # par (M)=11.362023.03 | 70.09 | 89.424 | |
| ResNet-50Backbone=ResNet-50, Context=Comparison with PVTv2-B0 tier2025.03 | 70 | 89.7 | |
| ResNet18Image Size=224x224, # par (M)=11.692023.03 | 69.758 | 89.078 | |
| MoCo-v3 + E-SSLView=2+4 view, Protocol=MLP2026.03 | 68.45 | 88.31 | |
| MoCo-v3 + EquiModView=3 view, Protocol=MLP2026.03 | 68.33 | 88.5 | |
| MoCo-v3 + SERView=2 view, Protocol=MLP2026.03 | 68.04 | 88.37 | |
| MoCo-v3View=2 view, Protocol=MLP2026.03 | 67.84 | 88.37 | |
| MoCo-v3 + STLView=2 view, Protocol=MLP2026.03 | 65.74 | 86.32 | |
| MoCo-v3 + SERView=2+4 view, Protocol=20-NN2026.03 | 65.32 | — | |
| Π-NetsImage Size=224x224, # par (M)=12.382023.03 | 65.28 | 85.958 | |
| MoCo-v3 + E-SSLView=2+4 view, Protocol=20-NN2026.03 | 64.56 | — | |
| MoCo-v3 + AugSelfView=2 view, Protocol=MLP2026.03 | 63.24 | 85.86 | |
| MoCo-v3 + SERView=2 view, Protocol=20-NN2026.03 | 61.64 | — | |
| MoCo-v3View=2 view, Protocol=20-NN2026.03 | 61.56 | — | |
| MoCo-v3 + AugSelfView=2 view, Protocol=20-NN2026.03 | 60.63 | — | |
| MoCo-v3 + EquiModView=3 view, Protocol=20-NN2026.03 | 58.28 | — | |
| MoCo-v3 + STLView=2 view, Protocol=20-NN2026.03 | 57.34 | — | |
| ResNet18 without activationsImage Size=224x224, # par (M)=11.692023.03 | 20.536 | 39.986 |