Improvements regarding applying box constraints

[sapo17]

Feedback from David and notes from the thesis

Might be worth having a look at how ADAM usually handles box constraints (or if that is commonly done anywhere) - here it sounds like you basically force the parameters onto the contraints by projection in between ADAM steps. (You can always say this is good enough for our purposes.)
Feb 21, 2023 4:16 PM

You: Unfortunately could not find anything about this in Mitsuba.
Morever, they also seemed to suggest doing it this way. For example: https://mitsuba.readthedocs.io/en/stable/src/inverse_rendering/gradient_based_opt.html#Optimization
Feb 22, 2023 9:02 PM

Ha ok, well if everyone does it ... ;-) ... and it seems to work just fine.
Although I currently don't recall any paper that analysed this in more detail (and this is probably not the right place to do that either). My feeling is that one should also project the gradient when projecting the value to the constraint set, but that really is a question for some other work.
Feb 23, 2023 9:40 AM