Python port
srush opened this issue · comments
Hi there,
I really love the g9.js demos! I think they convey how gradients work in such a nice and intuitive manner.
Just for fun I was playing with the idea of running a Torch-like library in the browser. I ported over most of the g9 code to python and ran it with AD (as opposed to finite-differences) -> https://srush.github.io/g9py/ . Not sure why this is better, but it might let you run models with tons of parameters. Let me know if this is okay.
Cheers!
Sasha
Very cool! I'd be happy to add a link to this if you like.
It seems at first click that it's a bit less performant for the recursive demos - I haven't read through the code yet, but did you implement the pruning part that uses ctx.pure? Or does that not play nice with AD?
Right, yeah I forget why I didn't add that. Need to think about what it means for with torch-style AD.