Created by: jczaja

Hi,

There are operators like LRN which are having one of an outputs (MidOut) that is no use for user , but it contains intermediate data that are speeding computation of LRN grad operator. In unit tests, (test_check_output) this MidOut is also verified if it contains valid / expected data. Problem is that what is inside MidOut is implementation specific. oneDNN LRN will produce MidOut with diffrent kind of data, because LRN oneDNN grad op is diffrent in implementation to Paddle CPU LRN grad. So unit test may fail due to diffrent kind of data being put into MidOut. So my question is : is it possible to skip comparing MidOut tensor in Unit tests of LRN oneDNN ? If positive then how can I do that?