Comment by nwoli

One elegant approach for this I’ve found is this https://github.com/mit-han-lab/gan-compression They basically train an “all in one” network from which you can extract small or large models afterwards (with optional additional finetuning to improve the selected channel size combinations)