Image Classification on iNaturalist (val)
73.1Top-1 AccuracyViT-B-16 (Semantic Softmax)
Evaluation Results
| Method | Links | |
|---|---|---|
| ViT-B-16 (Semantic Softmax)Architecture=ViT-B-16, Pretraining=Our ImageNet-21K (Semantic Softmax)2021.04 | 73.1 | |
| ViT-B-16Architecture=ViT-B-16, Pretraining=Official ImageNet-21K2021.04 | 71.7 | |
| Mixer-B-16 (Semantic Softmax)Architecture=Mixer-B-16, Pretraining=Our ImageNet-21K (Semantic Softmax)2021.04 | 66.6 | |
| Mixer-B-16Architecture=Mixer-B-16, Pretraining=Official ImageNet-21K2021.04 | 62.2 | |
| Full fine-tuningFine-tuning regime=Full2022.03 | 54 | |
| Learnable Memory TokensMemory cells=52022.03 | 50.1 | |
| Learnable Memory TokensMemory cells=102022.03 | 50 | |
| Learnable Memory TokensMemory cells=202022.03 | 50 | |
| Learnable Memory TokensMemory cells=12022.03 | 49.4 | |
| Head + Class token fine-tuningFine-tuning regime=Head + Class2022.03 | 48.7 | |
| Head-only fine-tuningFine-tuning regime=Head Only2022.03 | 46.3 | |
| SupervisedBackbone=ResNet-50, Parameters=25.6M, Evaluation Setup=[19]2019.12 | 45.4 | |
| PIRLBackbone=ResNet-50, Parameters=25.6M, Evaluation Setup=[19]2019.12 | 34.1 | |
| NPID++Backbone=ResNet-50, Parameters=25.6M, Evaluation Setup=[19]2019.12 | 32.4 | |
| RotationBackbone=ResNet-50, Parameters=25.6M, Evaluation Setup=[19]2019.12 | 23 | |
| JigsawBackbone=ResNet-50, Parameters=25.6M, Evaluation Setup=[19]2019.12 | 21.3 |