anirudh2290 commented on issue #14830: Use env var to enforce safe accumulation in ReduceAxesCompute URL: https://github.com/apache/incubator-mxnet/pull/14830#issuecomment-493146937 Adding new operator to your model is different from switching to safe accumulation on your exact model. Also, if you can point me to an env var doc which affects more than one operator and makes some claims about accuracy or performance that may be misleading I am happy to revisit it. The doc says this : "If this variable is set, the accumulation will enter the safe mode, meaning accumulation is done in a data type of higher precision than the input data type, leading to more accurate results with a possible performance loss and backward compatibility loss." Its natural for user to "wonder more accuracy results of what ? Which operators support safe accumulation? This is a global switch so I am assuming this should impact accuracy of many of our operators. So if I turn it on accuracy of my model may get better or worst case remain the same not worsen". This is not the case and I want us to clarify it. I don't think I am making any unreasonable ask here (Just add a simple note: "This switch may not necessarily improve accuracy of your overall model and in certain cases can make it worse"). By the way your switch currently doesn't control the softmax operator safe accumulation, which is mentioned in the issue on Naveen's PR and probably that is why you are not able to reproduce the issue. I am sure in the future this switch will control softmax operator too.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to 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
