Check issue #359 for a detailed description of the problem and PR #402 for a reproducing it
I'm not sure about how this solves the problem, but seems to work. I'll keep as a draft until I understand what's going on
UPDATE: @RainerKuemmerle agrees that this change makes sense. He believes that what happens is that this epsilon affects the numeric evaluation of the gradient at some point.
Also we are using the change at Rapyuta w/o issues, so I change to ready for review
Check issue #359 for a detailed description of the problem and PR #402 for a reproducing it
I'm not sure about how this solves the problem, but seems to work. I'll keep as a draft until I understand what's going on
UPDATE: @RainerKuemmerle agrees that this change makes sense. He believes that what happens is that this epsilon affects the numeric evaluation of the gradient at some point. Also we are using the change at Rapyuta w/o issues, so I change to ready for review