How to use unet_learner with only 1 channel?

[Mike_Nelson (Michael Nelson)]

@BobMcDear

fastai’s SegmentationDataLoaders.from_label_func because it converts gray-scale pictures to RGB ones. To get the intended behaviour, you can implement a custom DataBlock that ensures the input is being loaded properly.

Thanks, I think that is what I need to look into doing. I did end up getting things working either by going with pure PyTorch or simply allowing the conversion to greyscale RGB, but since I would like to scale this, I would prefer to understand how to choose the correct number of channels (which may someday be 2 or 4, 3D, etc.) and load in pretrained weights for a given number of independent channels. Much more to learn.

@BobMcDear

Finally, for using the PyTorch UNet you have linked with fastai, you would wrap it in a Learner alongside your data to acquire fastai functionalities.

Yep, think I had that in the code above. Thanks.