Variable-size input

I wonder what the preferred way of dealing with variable-size inputs? Say, I’d like to train on batches of different sizes with different spatial dimensions (height and width).

Thanks!

It’s best to preprocess your dataset so that everything is the same size.