anirudhacharya commented on issue #14262: Fix NaN value comparisons in relu,
max and min ops
URL: https://github.com/apache/incubator-mxnet/pull/14262#issuecomment-471102695
Run Mode: Before --> After ( time in ms)
'Whole CPU run: '- 0.843163 --> 0.864071
'Forward CPU r
anirudhacharya commented on issue #14262: Fix NaN value comparisons in relu,
max and min ops
URL: https://github.com/apache/incubator-mxnet/pull/14262#issuecomment-470760203
I modilied relu grad to also propagate `NaN` values. As discussed above,
since `NaN` does not exist in the domain of
anirudhacharya commented on issue #14262: Fix NaN value comparisons in relu,
max and min ops
URL: https://github.com/apache/incubator-mxnet/pull/14262#issuecomment-470359489
> Could we add a new operator to check whether there are nan? It is used
when debug.
> When nan appears in the mo
anirudhacharya commented on issue #14262: Fix NaN value comparisons in relu,
max and min ops
URL: https://github.com/apache/incubator-mxnet/pull/14262#issuecomment-470358848
@szha What you say makes sense, nan is not a number and hence not in the
realm of comparison. So any occurrence eith
anirudhacharya commented on issue #14262: Fix NaN value comparisons in relu,
max and min ops
URL: https://github.com/apache/incubator-mxnet/pull/14262#issuecomment-470349303
> @anirudhacharya I'm not sure if relu grad should act like that. As a
sanity check, consider if nan is larger than
