Zero-shot Classification on ImageNet 1k V2 R A (test)
86.3Top-1 Acc (Test)COCA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| COCAArchitecture Image=ViT-g, Architecture Text=T-g, Batch size=65k, Examples seen=32.8B, Parameters per token=2.1B, Training objective=non-contrastive2022.06 | 86.3 | 80.7 | 96.5 | 90.2 | |
| BASICArchitecture Image=CoAtNet-7, Architecture Text=T-H, Pretrained=true, Batch size=65k, Examples seen=19.7B PT + 32.8B, Parameters per token=3B2022.06 | 85.7 | 80.6 | 95.7 | 85.6 | |
| LITArchitecture Image=ViT-g, Architecture Text=T-g, Pretrained=true, Batch size=32k, Examples seen=25.8B PT + 18.2B, Parameters per token=2.1B2022.06 | 84.5 | 78.7 | 93.9 | 79.4 | |
| LIMoEArchitecture=H/14, Model structure=One-tower, Batch size=21k, Examples seen=23.3B, Parameters per token=675M2022.06 | 84.1 | 77.7 | 94.9 | 78.7 | |
| ALIGNArchitecture Image=EffNet-L2, Architecture Text=T-L, Pretrained Text=true, Batch size=16k, Examples seen=19.8B, Parameters per token=~ 820M2022.06 | 76.4 | 70.1 | 92.2 | 75.8 | |
| CLIPArchitecture Image=ViT-L/14, Architecture Text=T-B, Uses FixRes=true, Batch size=32k, Examples seen=12.8B, Parameters per token=~ 400M2022.06 | 76.2 | 70.1 | 88.9 | 77.2 |