Transfer learning with black/white images

How does the fastai library deal with training on e.g. MNIST data with transfer learning in lesson 1? MNIST data is monochrome, so there’s only one channel. But the transfer learning weights are for three channels (RGB). I guess there are a few ways to deal with that, but specifically how does the fastai library deal with that?

I remember seeing some posts about transfer learning with 2, 4 (so not 3) etc.channels. But there was no code recipes.