Lie group checks should be silent for internal operations

[luisenp]

When the outcome of something like g1.compose(g2) needs to be normalized, the current code throws a warning. However, this is overly informative for internal ops, and might result in an excessive/intrusive number of warnings. Better to keep these warnings only for when normalizing a user-given tensor at construction time.

[jacoblubecki]

I'd like to take a look at this, but from some cursory code-searching, it isn't immediately clear where the warning is occurring. Could you point me in the right direction and I can throw something together?

[luisenp]

Hi @jacoblubecki. Thanks for the offer to help!

These checks happen here. Which is called in the base Manifold class here.

The first option that comes to mind is to extend the _LieGroupCheckContext class to include something like cls.contexts.silent_normalization, which can then be passed in the call to Manifold._check_tensor(). Then in TheseusLayer.forward() you can disable/enable the silent option on enter/exit, w/o changing what the current value of _LieGroupCheckContext.check_lie_group is.

Happy to answer any questions.

[jacoblubecki]

Thank you, this is perfect. I'll start taking a look! In the meantime, can you assign the ticket to me?

[luisenp]

Assigned. Thanks!

[jacoblubecki]

I created a draft pull request -- the CONTRIBUTING.md instructions say to add myself as the assignee and to assign some specific reviewers, but it looks like I don't have permission to set assignees or reviewers on the draft PR...