Pixel-based models?

[LeanderK]

The readme states:

For the moment, I removed MNIST and CIFAR10 because I have plans to add pixel-based models later.

what do you mean by that? What's missing?
I might be able to implement some stuff.

[ikostrikov]

At some point I was going to add models with convolutions coupling layers.