-
Notifications
You must be signed in to change notification settings - Fork 640
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Backbone/model weights source should be more obvious to figure out. #543
Comments
I'd be happy to do both but perhaps can the core contributors validate/reject the idea? |
@jpcbertoldo, I'm not sure how easy Option-2 would be since It might also be an idea to deprecate the |
We also use torchvision for its |
ah yes, of course. |
Looks like it is also used in reverse distillation |
It's this
Okay, probably not the way to go then :)
It is used in many places apparently. If I got it right To me it sounds like a good idea to keep it. I guess option 1 (add some mention to |
I wouldn't be in favor of removing torchvision because it provides several useful features besides pretrained model weights. Also, If I remember correctly, many timm models actually obtain their pretrained weights directly from torchvision. Therefore I believe that in most cases getting the pretrained weights from torchvision would give us the same weights as when getting the weights from timm. So I feel the second option would add unnecessary complexity to the library. However I do agree that it would be a good idea to mention the source of the pre-trained weights either in the readme or documentation or both. |
My bad, forget about it. I just forgot that we already refactored I also agree with @djdameln. Removing |
Ok, so just some documentation update. |
(asking this here but i will create a proper issue if necessary) context: The
I believe there are two problems here (can you confirm?): a. the user cannot adjust b. there is no validation that the offset was incorrect, so the bug would be silent I think
|
Addressed with #576 |
I saw
torchvision
in the requirements and immediately assumed that the pre-trained model weights would come from there (I didn't knowtimm
until very recently).I saw they are actually coming from
timm
and realized it's not a well documented information.I have two suggestions:
(easy) mention it in the
readme.md
and add it to the docs.(maybe easy?) make the pre-trained weights source a parameter that would show up in the config file (it could come from
torchvision
for instance)The text was updated successfully, but these errors were encountered: