ir_csn_152 size mismatch: Error(s) in loading state_dict for VideoResNet

[KhHTran]

I follow the installation guide to install VMZ locally. When I try to load the pretrained ir_csn_152 model to PyTorch:
from vmz.models import ir_csn_152
model = ir_csn_152(pretraining="ig65m_32frms")
I received an error saying that there is a size mismatch with fc.weight and fc.bias. I assumed this is a bug where the finetuned model was given the pretrained model weights or vice versa.

[ningbende]

same problem+1

[hoonkai]

Same problem here. Is this solved?

[KhHTran]

I'm not sure. But you can solve the issue pretty easily.

[GriesserP]

I'm not sure. But you can solve the issue pretty easily.

Hello, could you give the solution please ?

[KhHTran]

You load an empty model then load the state of the pre-trained model.
`url = 'https://github.com/bjuncek/VMZ/releases/download/test_models/irCSN_152_ft_kinetics_from_ig65m_f126851907.pth'
dic = torch.hub.load_state_dict_from_url(url, progress = True)

model = ir_csn_152()
model.load_state_dict(dic)`