image preprocessing about docta HOT 4 CLOSED

Hi

Thanks for the great catching and your detailed post! We have fixed this "double processing" in our latest commit.

Briefly speaking, in the file customize_img_folder.py, we removed the data preprocessing code when transform=None to avoid double preprocessing. In this case, Docta will simply append features by changing each image from <PIL.Image.Image> to the numpy array. Thus, the revised code will process & normalize the image data for only once when transform=None (default value).

Please feel free to let us know if you have any additional concerns!

Best,
Jiaheng

from docta.

Yes, the CIFAR data is not double-processed. The transforms in the CIFAR dataloader are left for future extensions. Thank you for the good catching.

from docta.

Hi @weijiaheng.
Thank you for your reply and for helping me clear my doubts.

I also went over again the CIFAR10 preprocessing cause I thought there was the same double processing issue also there.
But I missed the following:

and here self.data is not normalized yet (since the normalization is applied by the dataloader and not by the dataset)
and then it just undergoes the preprocessing of CLIP with:
model_embedding, _, preprocess = open_clip.create_model_and_transforms(self.cfg.embedding_model)
and
CustomizedDataset(..., preprocess = preprocess) in line:

Thank you once again for your help.
Have an amazing day.

from docta.

Thank you for the clarification @zwzhu-d and @weijiaheng.
It was extremely useful.

from docta.