How can I change the number of channels of a pretrained model?

[aykutcayir34]

I want to use a single channel resnet-18 model in flashlightning image classifier. Is there a convenient way to change the number of channels of a pre-trained model in flashlightning?

[ethanwharris]

Hi @aykutcayir34 you should be able to do this:

model = ImageClassifier(
    backbone="resnet18",
    backbone_kwargs={"in_chans": 1},
    ...
)

Hope that helps 😃

[aykutcayir34]

flash version is '0.5.0rc0'

[aykutcayir34]

Your solution worked for efficientnet, but I could not run the solution for resnet-18 :/

[ethanwharris]

Hi @aykutcayir34 thanks for the extra information! This is definitely a bug on our side. The efficientnet comes from timm underneath which has a way to handle pretrained weights with different input channels. Our resnet code is a little different so we need to add support for this there. I've opened #741 so that you can track the progress of this issue. In the meantime, the efficient net and any resenet variants like resnext or other models should all work. The resnet18 should work ok if you set pretrained to false.

Hope that helps 😃

[aykutcayir34]

Thank you so much. You helped me a lot at speed light ;)

[ethanwharris]

@aykutcayir34 you're welcome 😃 This PR should fix your issue once it's merged: #743