I work in deep learning, and I'm very dubious of the benefit of terse code. In my experience 90% of novel deep learning models can be coded with base layer classes from tf/pytorch/keras + maybe a few hundred lines of 'new' code. For this code I feel like you will avoid mistakes and make things so much easier for the reader (probably yourself in 6 months), if you modularise, write it as clearly as possible and explain what each part is doing.

I have used code from the fast.ai codebase and personally I find it horrendous to work with - I find the way they structure and name classes consfusing, they use wildcard imports everywhere, they use tiny variable names for everything - it's all extremely reader-unfriendly, which for an educational tool seems completely bizarre.

I don't quite enjoy "explicit mathematically dense" code, it leads to bad habits, see item 4: Python's syntax makes it all too easy to write single-line expressions that are overly complicated and difficult to read.

If you really need to pack a lot of math, just define it in a module (e.g. tensorflow does that nicely).

I've had an intern who came from MATLAB. Convincing them to stop packing numerical constants in the middle of every expression was unexpectedly difficult. "But it's shorter that way!"