Image Classification on ImageNet-1K (val) (Top-1 Accuracy and MHSA Cost)
83.4Top-1 AccuracySparsif-LV-S
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Sparsif-LV-STok. keep rate=1, Att. keep rate=0.5, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 83.4 | 339.7 | |
| LV-VIT-STok. keep rate=1, Att. keep rate=1, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 83.3 | 476.9 | |
| Sparsif-LV-STok. keep rate=1, Att. keep rate=0.25, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 83.3 | 221.7 | |
| EVIT-LV-STok. keep rate=0.7, Att. keep rate=1, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 83 | 256 | |
| DynViT-LV-STok. keep rate=0.7, Att. keep rate=1, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 83 | 256 | |
| Sparsif-LV-STok. keep rate=1, Att. keep rate=0.1, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 82.8 | 149.5 | |
| EVIT-LV-STok. keep rate=0.5, Att. keep rate=1, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 82.5 | 198.8 | |
| DynViT-LV-STok. keep rate=0.5, Att. keep rate=1, Input resolution=224 x 224, Backbone=LV-ViT-S2023.03 | 82 | 198.8 | |
| DeiT-STok. keep rate=1, Att. keep rate=1, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 79.8 | 357.7 | |
| EVITTok. keep rate=0.7, Att. keep rate=1, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 79.5 | 193.1 | |
| Sparsif-EVITTok. keep rate=0.7, Att. keep rate=0.25, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 79.5 | 113.3 | |
| DynamicViTTok. keep rate=0.7, Att. keep rate=1, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 79.3 | 193.1 | |
| SparsifinerTok. keep rate=0.7, Att. keep rate=0.25, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 79.3 | 113.3 | |
| Sparsif-EVITTok. keep rate=0.5, Att. keep rate=0.25, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 78.7 | 86.6 | |
| EVITTok. keep rate=0.5, Att. keep rate=1, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 78.5 | 149.1 | |
| SparsifinerTok. keep rate=0.5, Att. keep rate=0.25, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 78.4 | 86.6 | |
| DynamicViTTok. keep rate=0.5, Att. keep rate=1, Input resolution=224 x 224, Backbone=DeiT-S2023.03 | 77.3 | 149.1 |