Dzmitry Bahdanau
Dzmitry Bahdanau
Sorry for the late reply. If you are still interested, can you please post an code example that reproduces the issue? On Thu, 1 Dec 2016 at 17:56 varunkumar3618 wrote:...
See discussion in #343.
I want to avoid updating but raise an exception or warning. The new step rule just hides the NaNs.
While this is true that a temp file would do the job, I feel it much more naturally to handle these things in the algorithm. This way when you get...
Is the `NanGuardMode` as fast as `fast_run`?
I did not mean "exactly as fast" as `fast_run`, I meant that I would not like to break optimizations or use less optimization compared to `fast_run`.
Anyway, I think I know a super easy solution! We should simply add a shared variable to `RemoveNotFinite` which would be 1 if there were infinite elements in the step...
By choosing the right scaling constant the parameters that would contain NaNs will be not corrupted. Yours is a good point, that other parameters might be damaged. I think we...
Okay, then this ticket becomes a CCW level refactoring of `RemoveNotFinite`.
By the way, documentation `RemoveNotFinite` seems misleading to me. It is claimed that "parameters are replaced with scaled version", which is in the end true, but the thing is that...