Image Classification on DataComp 38 downstream tasks medium (test)
66.7AccuracyViTamin-L
Evaluation results
| Method | Links | |
|---|---|---|
| ViTamin-LEncoder Size(M)=333, Input Res.=224, Latency Enc.(ms)=38.1, Hardware=M1 Macbook Pro2024.12 | 66.7 | |
| ViT-L/14Encoder Size(M)=304, Input Res.=224, Latency Enc.(ms)=47.2, Hardware=M1 Macbook Pro2024.12 | 66.3 | |
| FastViTHDEncoder Size(M)=125, Input Res.=224, Latency Enc.(ms)=6.8, Hardware=M1 Macbook Pro2024.12 | 66.3 | |
| ConvNeXt-LEncoder Size(M)=200, Input Res.=320, Latency Enc.(ms)=34.4, Hardware=M1 Macbook Pro2024.12 | 63.9 | |
| DFN U HYPE U Ours (10%)Strategy=Data Union, Selection Ratio=10%, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 38.8 | |
| DFN U Ours (20%)Strategy=Data Union, Selection Ratio=20%, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 38.6 | |
| DFN U HYPEStrategy=Data Union, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 37.9 | |
| DFNStrategy=Data Filtering Network, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 37.3 | |
| DevilsStrategy=Filtering, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 37.1 | |
| T-MARSStrategy=Filtering, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 36.1 | |
| CLIPScoreStrategy=Filtering, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 32.8 | |
| No filteringStrategy=Baseline, Pre-training Dataset Scale=DataComp-medium (128M data), Evaluation Protocol=Zero-shot2024.05 | 25.8 |