I have used keras-contrib implementation of Instance Normalization Layer with default parameters (axis = None) for the image-to-image translation network. I have just seen that axis parameter should be -1 (channel) not None as it is by default. Which type of normalization I have done with axis=None?
I have good results with both options, although there are some differences in translations.
I think it should be corrected in the implementation as it is not uniform for other normalization layers in keras as well as in other frameworks.
I have used keras-contrib implementation of Instance Normalization Layer with default parameters (axis = None) for the image-to-image translation network. I have just seen that axis parameter should be -1 (channel) not None as it is by default. Which type of normalization I have done with axis=None?
I have good results with both options, although there are some differences in translations.
I think it should be corrected in the implementation as it is not uniform for other normalization layers in keras as well as in other frameworks.