The feature normalization is necessary?

[roomo7time]

https://github.com/yaohungt/Barlow-Twins-HSIC/blob/a30baba4d2d3dcdf85a4ccebe57902f0d827c1ed/model.py#L31

Hi, your code is very helpful and I want to firstly appreciate the code share.

I have a question on whether this feature normalization is necessary (to make the cifar10 performance to about 92% accuracy).

The original Barlow Twins does not contain this step. On the other hand, they rather define all linear layers in the projector with no bias.