Imbalanced classes in semantic segmentation

Discussion
1

Is it better to calculate class weights to re-scale loss for the whole dataset or per minibatch?

2

In my experience, it’s better to rescale the loss based on the entire training set. But I’d love to hear other opinions!