weights = weights / np.sum(weights) * int(hparams['data_version'])

[nikenj]

the weights normalize function is not in the paper, why in the code need normalized ? and why to
multiply the number of classes? ths .

[richardaecn]

The weight normalization is mentioned in the 3rd paragraph of "4. Class-Balanced Loss".

The vanilla loss could be viewed as having weight of 1 for all classes. This simple normalization strategy ensures the loss value is in the similar scale as the vanilla loss.