search

berenslab / retinal-rl

Testing theories about retinal coding in reinforcement learning environments
GNU Affero General Public License v3.0
0 stars 1 forks source link

Flatten objectives #16

Closed alex404 closed 1 week ago

alex404 commented 1 week ago

Flattened objective hierarchy so that it's no longer "Goal -> Objective -> Loss" and just "Objective -> Loss". Losses now contain a list of target circuits and weights. Overall seems to be a bit more performant because we're not doing as many loops over the AD graph. Configs updated accordingly.

Resolves #8

alex404 commented 1 week ago

Overall I think for meaningful review we should strive for "smaller" and more concise pull requests, as quite a lot has changed here

Okay I'll try and keep that in mind. You'll probably not going to like my next pull request though...

alex404 commented 1 week ago

If you'd prefer to have smaller merges (which I think can make sense) maybe I'd reactivate the project interface. Then basically we keep it as a big list of todos. When a checklist item gets mature enough, we turn it into an issue, and then address it with a merge. Would you prefer that strategy?

Otherwise my current strategy is making thematically organized todos under issues, and addressing them in bulk. Currently I'm working on #11, which again will have a large number of changes.