asmushetzel commented on issue #11271: nd.topk regression with nan values URL: https://github.com/apache/incubator-mxnet/issues/11271#issuecomment-397763186 @piiswrong I searched a bit through the documentation and the only place where I could find something about NaN handling was in the nansum/nanprod operators (which are specifically designed for handling NaN). All the rest is undefined. Which is kind of o.k. as having NaN in any array is likely a severe problem (see all the issues in MxNet where people got suddenly an NaN during training and see this as a bug). While I agree that it would be desirable to standardize NaN handling, I think - whatever we do, it must be documented. Otherwise we don't have a written convention and things may change erratically all the time. - Labeling such issues as "bug" is not correct. It can only be a bug if it violates a defined behavior - Whatever we do, it should be consistent across all comparison operators (max, min, broadcast_greater, .....). Which could mean that we won't find any consistent solution that does not alter the behavior of at least one of such operators. - If we can't come up with consistent handling, then NaN should be declared has having undefined behaviour whenever encountered in comparisons in MXNet
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
