Closed yasserben closed 4 months ago
Hi @yasserben,
Thank you for the message and highlighting your concerns.
Regarding your questions, kindly note the following:
Could you clarify whether I'm misunderstanding the intended use, or if there might be a bug ?
Yes I checked it and it looks like there was a bug. I have removed the cross-dataset condition in the data_magager.py
, and now the model is training as expected. The repository has been now updated.
Additionally, there appears to be a typo in the cross dataset training command in TRAIN.md, where
promptsrc
should likely beprotext
Thanks for pointing this out, I have fixed the typos.
Kindly let me know if you have anything else to discuss.
Kind regards! Muhammad Uzair
Hi @muzairkhattak
Thank for your answers, everything is clear now :+1:
Best regards, Yasser
Hello :smiley:
I am experiencing an issue with the
cfg.TRAINER.PROTEXT.CROSS_DATASET
variable during cross-dataset training on ImageNet. This setting is supposed to enforce text-only training, using ImageNet class names and excluding image data. However, the code seems to contradict this:https://github.com/muzairkhattak/ProText/blob/32237a035fbfb71ed733b9c947b973259f923696/Dassl.pytorch/dassl/data/data_manager.py#L28-L30
This suggests that text data loads only when cfg.
TRAINER.PROTEXT.CROSS_DATASET
isFalse
, which seems incorrect for cross-dataset settings where no images should be loaded. As a result, training crashes at the following line as it tries to process text data that is not present in the batch : https://github.com/muzairkhattak/ProText/blob/32237a035fbfb71ed733b9c947b973259f923696/trainers/protext.py#L293-L294Could you clarify whether I'm misunderstanding the intended use, or if there might be a bug ?
Additionally, there appears to be a typo in the cross dataset training command in TRAIN.md, where
promptsrc
should likely beprotext
Thank you very much ! :wink: