Using pretrained ResNet with MedicalNET weights

[Meddebma]

Hi,
I am trying to make a classification for 3d data using a pretrained Resnet with MedicalNet weights, when I start to train, I get this error:
Error(s) in loading state_dict for ResNet:
size mismatch for conv1.weight: copying a param with shape torch.Size([64, 1, 7, 7, 7]) from checkpoint, the shape in current model is torch.Size([64, 3, 7, 7, 7]).

Could you help me find out where the mismatch come from?
Here is the Notebook:
https://github.com/Meddebma/pyradiomics/blob/master/Classification_HBI_outcome.ipynb

Thanks!

[Nic-Ma]

Hi @yiheng-wang-nv ,

Could you please help take a look at this ticket?

Thanks in advance.

[yiheng-wang-nv]

HI @Meddebma , the default number of input channels of monai's resnet is 3 (https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/nets/resnet.py#L163), but in the pretrained weights it was 1. If you need to use that pretrained weights, you may need to load the weights except the conv1 layer.
You can check here for reference

[Hogandr]

Hey @Meddebma,
can you share your working code?
Thanks in advance