Hi! I was poking around yesterday evening to add some kind of “interpretability layer” to my satellite imagery project (check it out here if you haven’t yet).
On the subject of interpretability, I loved this distill.pub article. However, there doesn’t seem to be the equivalent library to the one they used, Lucid for PyTorch. I’m still on the fence but working on a subset of something similar for PyTorch/fast.ai sounds like a worthwhile project.
Would this be of interest to any of you?
[EDIT]: forget it 
current discussion of implementations and initiatives by fastai students can be found here