facebookresearch / facebookresearch/MetaCLIP
I find a bug in your source code.
Open
- Dominant language
- Python
- Stars
- 1.9k
- Forks
- 80
- PR merge metrics
- No merged PRs in 30d
Description
In your default setting, "gpu_trans" is True. but training preprocessor uses "gpu_trans=True" and eval preprocessor uses "gpu_trans=False".
So data' range is different from training phrase and eval
By the way, model.forward has no operation to normalize data range for training phrase, so is is a big bug
Contributor guide
Assessment
This issue has not been assessed yet.