Support model in_chans not equal to pre-trained weights in_chans

[keves1]

This PR addresses #2289.

I've made a good start on this and am creating this draft PR to get some feedback before continuing. I have created a load_pretrained function that will copy weights as done by timm, but which supports any number of in_chans of the weights, not just 3. So far I just implemented this new functionality in resnet50 and will expand that to other models if this is a good approach.

With this change, you can create a model like this, in this case the two channels will be copied into 4:

model = resnet50(weights=ResNet50_Weights.SENTINEL1_ALL_MOCO, in_chans=4)

I created a new file torchgeo/models/utils.py where I put the load_pretrained function since I didn't see any suitable existing file. Is there a better place for it?

[keves1]

@microsoft-github-policy-service agree

[adamjstewart]

I have created a load_pretrained function that will copy weights as done by timm, but which supports any number of in_chans of the weights, not just 3.

Why can't we use timm.models.helpers.load_pretrained instead of writing our own custom code?

[keves1]

Why can't we use timm.models.helpers.load_pretrained instead of writing our own custom code?

Because that function will only copy the weights to additional input channels if the first convolution layer of the weights has 3 input channels. Otherwise it raises a NotImplementedError exception and randomly initializes the weights. I linked to the timm implementation in my comment on issue #2289 .

[adamjstewart]

Will try to review next week, this week is quite busy. Apologies for the wait!