Very cool! Thanks.

Agreed. My git fu was weak in 2010.

For sure!

Thanks. I left out free()'s to get the thing working fast, but you’re right, they should be in there.

Nod, I’m leaving out early stopping so all versions are comparable.

Noted, thanks. I may swing back around and add some free() calls.

UPDATE: I’ve added an optimized numpy version that is as fast (actually slightly) faster than the naive c version. Yay!

Good point.

I used to struggle in the same way with kaggle competitions. Whole parts of my pipeline would change, but wanted/needed to preserve each pipeline independently for ad hoc experimentation. I used parallel dirs in the git repo.

The ideas is to see the progression as the code is optimized away for readability toward performance.

You’d prefer the files to be named meaningfully? Or you’d prefer to use “git diff” to see changes and have a single copy?

The risk is a new version has a good idea, but is slower. It should not be the canonical version, just a stepping stone.