Image Classification on Tiny ImageNet (Acc@1, Acc@5, CV)
59.13Top-1 AccuracyHi-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Hi-MoEBackbone=Swin Transformer, Total Parameters=1B, Active Parameters=83M, Epochs=2002026.05 | 59.13 | 81.18 | 0.29 | |
| Loss-free balancingBackbone=Swin Transformer, Total Parameters=1B, Active Parameters=83M, Epochs=2002026.05 | 58.87 | 79.65 | 0.31 | |
| Vanilla GShard-styleBackbone=Swin Transformer, Total Parameters=1B, Active Parameters=83M, Epochs=2002026.05 | 58.61 | 80.39 | 0.33 | |
| MoGEBackbone=Swin Transformer, Total Parameters=1B, Active Parameters=83M, Epochs=2002026.05 | 58.61 | 80.92 | 0.3 |