Standard error of CLIP-ViT experiments
korts-lew opened this issue · comments
Byounggyu Lew commented
Hello, I make this issue to ask standard errors of experiments using CLIP-ViT.
Unlike the results reported using ResNet-50 and RegNet w/ SWAG, there is no standard errors for the results of CLIP-ViT
Can I get standard errors for it? (stds of 3 runs in PACS, VLCS, OfficeHome, TerraInc, DomainNet)
If it is impossible, can you give me hyperparameters used for the experiments instead?
Thank you for your great work again
Junbum Cha commented
Here is the results with standard errors of CLIP-ViT:
Algorithm | PACS | VLCS | OfficeHome | TerraInc | DomainNet |
---|---|---|---|---|---|
ERM | 83.4 ± 0.5 | 75.9 ± 1.3 | 66.4 ± 0.5 | 35.3 ± 0.8 | 44.4 ± 0.6 |
MIRO | 95.6 ± 0.8 | 82.2 ± 0.3 | 82.5 ± 0.1 | 54.3 ± 0.4 | 54.0 ± 0.3 |
We omitted this due to lack of the space. Sorry for the inconvenience!
Byounggyu Lew (Korts) commented
Thank you for your kind reply !!