Ross Wightman

heh, also just noticed this :) curious what impact it may have as I doubt it was intentional

@chuong98 yeah, that's wrong, but also there is an issue wrt to checking for None vs isinstance(nn.Identity()) .. need look through it more closely

@haideraltahan I don't think there is a problem here, looks more likely to be an issue with how model is being used/wrapped in whatever this use case is. Can do...

@holderhe at this point I don't have the time to prioritize models like this, if someone created a fully functional and clean PR I'd consider, but it's a bit of...

well the flag is in there, but it's causing a problem #966 with metaclip checkpoints so fun fun, this is no longer an issue but I've created other ones

@amyeroberts @f-fuchs oh fudge, I see the issue, I used sets to resolve the indices and they don't keep the order, unlike dicts (really wish they'd updated sets when they...

@gau-nernst I have thought about this, and yeah it's not done because to do it at weight load time where it's done now, it's more complexity than I'd like, especially...

@gau-nernst on PR #2225 there is first pass at implementing a set_input_size fn... currently should mostly work for models from vision_transformer.py, vision_transformer_hybrid.py, and swin_transformer.py (v1)

working for mentioned models and v2 swin now too, can add to other models like eva/beit vits but will do so by request.

@ioangatop if you want classifier weights loaded into feature extraction wrapped models, you need to load weights as 'pretrained' so that they are loaded before the model is mutated. See...