Calculation of mean for centering the features

[pradipto111]

According to the given implementation in classification_tasks/utils_eval.py, the features phi is being centered as follows -
Line 502 phi = phi - torch.mean(phi, axis=1, keepdims=True).

However, assuming phi has a dimension (D, f) where D is the dataset size, and f is the feature size. The mean on axis 1 turns out to have the dimension (D,1). But in my opinion, we want the mean feature to have the dimension of (1,f), hence the mean should be applied across the dataset i.e. axis=0.

Kindly look into it.

[max-andr]

Good point. Centering along axis=0 would be more appropriate. But since BatchNorm is used throughout the convolutional networks, I would assume that this doesn't have much influence. I will double check, though.