pierreablin / landing Goto Github PK

Any updates on this?

Thanks!

Hi @pierreablin,

Updating the code of RotoGrad, I started using your library, and during testing I've found that the rotation matrices occasionally end up with nan parameters.

Debugging the code, I traced the bug to line 48 of the _safe_step_size function:

Specifically, the distance to the manifold is bigger than epsilon (0.5) at times, and so the square root yields a nan value. Is this a bug, or am I missing any step? I project the orthogonal matrix at initialization as in the snippet code so that it starts within the manifold.

Thanks for the library!

Hi @pierreablin, thank you for outstanding job! Recently, I have met a problem when using your code. In your code, before we use the landing algorithm to optimize the parameters on orthogonal manifolds, we need to project the parameters to make them orthogonal. But, I have found that when we initialize the optimzer which is based on the landing algorithm, the optimzer function will also project the parameters to be optimized to the orthogonal manifolds.(which can be found on the lines 142-144 of the optimzer.py) So, it looks like a repetition?