With USI, we can reliably identify models that provide good speed-accuracy trade-off.
For those top models, we provide here weights from large-scale pretraining on ImageNet-21K. We recommended using the large-scale weights for transfer learning - they almost always provide superior results on transfer, compared to 1K weights.
| Backbone | 21K Single-label Pretraining weights | 21K Multi-label Pretraining weights | ImageNet-1K Accurcy [%] |
|---|---|---|---|
| TResNet-L | Link | Link | 83.9 |
| TResNet-M | Link | Link | 82.5 |
| ResNet50 | Link | Link | 81.0 |
| MobileNetV3_Large_100 | Link | N/A | 77.3 |
| LeViT-384 | Link | Link | 82.7 |
| LeViT-768 | Link | N/A | 84.2 |
| EdgeNeXt-S | N/A | N/A | 81.1 |