Image Classification on ImageNet-1K (val) (Accuracy and FLOPs Saving)
13.5FLOPs Saving (%)AdaViT
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| AdaViTBackbone=CLIP ViT-B/32, Precision=INT8, Evaluation Protocol=Zero-shot2026.05 | 13.5 | 59.84 | 1.12 | |
| LRA-EEBackbone=CLIP ViT-B/32, Precision=INT8, Evaluation Protocol=Zero-shot2026.05 | 13.38 | 61.16 | 2.44 | |
| PABEEBackbone=CLIP ViT-B/32, Precision=INT8, Evaluation Protocol=Zero-shot2026.05 | 3.64 | 61.82 | 3.1 | |
| Baseline (INT8 full-depth)Backbone=CLIP ViT-B/32, Precision=INT8, Evaluation Protocol=Zero-shot2026.05 | 0 | 58.72 | — | |
| CLS-EEBackbone=CLIP ViT-B/32, Precision=INT8, Evaluation Protocol=Zero-shot2026.05 | 0 | 58.72 | 0 |