gradhep / neos

Upstream optimisation for downstream inference
BSD 3-Clause "New" or "Revised" License
69 stars 5 forks source link

Update to new upcoming `fax` API #16

Open gehring opened 4 years ago

gehring commented 4 years ago

We've made some pretty big changes to fax which will break neos. We'll be updating pypi with a new release soon but I wanted to offer to help update to the new simpler api (assuming this project is still in development). I was going to go straight for a PR but I noticed that the neos.fit module has some issues that I don't think I can properly fix without a bit of help.

The current implementation in neos.fit never allows the adam optimizer to update its state. Specifically, it always uses freshly initialized internal parameters for its scaling and momentum terms. If someone can help me better understand what constrained_bestfit_minimized is doing, I can submit a PR to both fix the bug and update to the new API.

lukasheinrich commented 4 years ago

Hi @gehring thanks so much for the offer. Yes this project is very much still active. Is there a branch on which we can preview the new API?

gehring commented 4 years ago

Yes, you should be able to see the new api on the master branch.

gehring commented 4 years ago

Oh, I should have point out that we've also updated the readme with an example which includes using a non-differentiable solver.