search

JieShibo / PETL-ViT

[ICCV 2023 & AAAI 2023] Binary Adapters & FacT, [Tech report] Convpass
MIT License
168 stars 8 forks source link

Some questions about SwinTrans edition #4

Closed 923714892 closed 2 years ago

923714892 commented 2 years ago

I try to use the SwinTrans convpass of the MLP version, but the dimension is always wrong. If dim and Vit_Dim parameter exchange will result in a very large number of parameters, which is far from consistent with the paper. I hope to get an answer. Thank you very much

JieShibo commented 2 years ago

Hi, I just checked and ran the code, and there seems to be no problem. Would you mind giving us more information about the error? To use the Convpass for Swin, just need to replace https://github.com/JieShibo/PETL-ViT/blob/3b25e37793a2ab97cbf7d06494edb2e222e561d0/vtab/train.py#L66 with 

    state_dict = torch.load('../swin_base_patch4_window7_224_22k.pth')
    model = create_model('swin_base_patch4_window7_224_in22k', drop_path_rate=0.1)
    model.load_state_dict(state_dict['model'])