but the feature maps in a P4M are not equivariant when testing an image and its 90 rotation version

[shuchao1212]

Dear Adam,

Thank you so much for your pytorch version of Group CNNs.

with your codes, I just wonder that the features maps in a P4M are not equivariant when testing an image and its 90 rotation version. 
 For example, a test image A ==> trained-well G-CNNs==> an outputed P4M feature maps M(8 ones); similarly, the rotated 90 version of A ==> the same G-CNNs==> an output P4M feature maps M'(8 ones). 
 Although I can find 8 pairs of feature maps from M and M' respectively, which are roughly very very similar, there exists minor differences any pair of them. As follows, the max value and min value of a feature map is compared:

 can you give some suggestions, Adam?

Thanks a lot