Implementation of the regularization loss

[adosar]

Hello and congrats for this nice repo!

I was looking in the implementation of the regularization loss and I don't know if it matches the original paper.

In the feature_transform_regularizer:

loss = torch.mean(torch.norm(torch.bmm(trans, trans.transpose(2,1)) - I, dim=(1,2)))

Why not:

loss = torch.mean(torch.norm(torch.bmm(trans, trans.transpose(2,1)) - I, dim=(1,2)).pow(2))

In the original paper they use: $L_\text{reg} = ||I - AA^T||^2$

[shan-zhong]

I found that the regularization loss has only been used for trans_feat, and the loss was simply added to the total loss. Why the regularization loss of trans has been ignored. And I think it may be better if the regularization loss is only used to update the parameter in T-net, rather than the whole net.