Image Classification on ImageNet 1K (val) (Top-1, Throughput, Memory)
84.7Top-1 AccuracyLITv2-B
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| LITv2-BParam (M)=87, FLOPs (G)=39.7, Resolution=384x384, Batch size=64, GPU device=NVIDIA RTX 3090, Fine-tuned=true2022.05 | 84.7 | 198 | 35.8 | 4.6 | |
| Swin-BParam (M)=88, FLOPs (G)=47.1, Resolution=384x384, Batch size=64, GPU device=NVIDIA RTX 3090, Fine-tuned=true2022.05 | 84.5 | 142 | — | 6.1 | |
| ConvNext-BParam (M)=89, FLOPs (G)=15.4, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.8 | 469 | 16.9 | 2.9 | |
| Twins-SVT-LParam (M)=99, FLOPs (G)=14.8, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.7 | 440 | 13.7 | 3.1 | |
| LITv2-BParam (M)=87, FLOPs (G)=13.2, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.6 | 602 | 12.2 | 2.1 | |
| LITv1-BParam (M)=86, FLOPs (G)=15.0, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.4 | 444 | 16.4 | 2.1 | |
| LITv2-MParam (M)=49, FLOPs (G)=7.5, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.3 | 812 | 8.8 | 1.4 | |
| Swin-BParam (M)=88, FLOPs (G)=15.4, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.3 | 386 | 13.4 | 2.4 | |
| Twins-SVT-BParam (M)=56, FLOPs (G)=8.3, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.2 | 621 | 9.8 | 1.9 | |
| ConvNext-SParam (M)=50, FLOPs (G)=8.7, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83.1 | 639 | 12.3 | 1.8 | |
| DeiT-BParam (M)=86, FLOPs (G)=55.4, Resolution=384x384, Batch size=64, GPU device=NVIDIA RTX 3090, Fine-tuned=true2022.05 | 83.1 | 159 | 39.9 | 2.5 | |
| Swin-SParam (M)=50, FLOPs (G)=8.7, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83 | 582 | 9.7 | 1.7 | |
| LITv1-MParam (M)=48, FLOPs (G)=8.6, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 83 | 638 | 12 | 1.4 | |
| Focal-TinyParam (M)=29, FLOPs (G)=4.9, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 82.2 | 384 | 12.2 | 3.3 | |
| ConvNext-TiParam (M)=28, FLOPs (G)=4.5, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 82.1 | 1,079 | 8.3 | 1.7 | |
| LITv2-SParam (M)=28, FLOPs (G)=3.7, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 82 | 1,471 | 5.1 | 1.2 | |
| ResNet-152Param (M)=60, FLOPs (G)=11.6, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 3090, Training strategy=ResNet Strikes Back2022.05 | 82 | 512 | 13.4 | 2.9 | |
| CvT-13Param (M)=20, FLOPs (G)=4.5, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 81.6 | 947 | 6.1 | 1.5 | |
| LITv1-SParam (M)=27, FLOPs (G)=4.1, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 81.5 | 1,298 | 5.8 | 1.2 | |
| ResNet-101Param (M)=45, FLOPs (G)=7.9, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 3090, Training strategy=ResNet Strikes Back2022.05 | 81.5 | 722 | 10.5 | 3 | |
| Swin-TiParam (M)=28, FLOPs (G)=4.5, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 81.3 | 961 | 6.1 | 1.5 | |
| Twins-PCPVT-SParam (M)=24, FLOPs (G)=3.8, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 81.2 | 998 | 6.8 | 1.2 | |
| PVT-MParam (M)=44, FLOPs (G)=6.7, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 81.2 | 680 | 9.3 | 1.5 | |
| ResNet-50Param (M)=26, FLOPs (G)=4.1, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 3090, Training strategy=ResNet Strikes Back2022.05 | 80.4 | 1,279 | 7.9 | 2.8 | |
| PVT-SParam (M)=25, FLOPs (G)=3.8, Resolution=224x224, Batch size=64, GPU device=NVIDIA RTX 30902022.05 | 79.8 | 1,007 | 6.8 | 1.3 |