Imbalance in training data

[J_Johnson]

What you’re doing is a semantic segmentation task. And it’s important to understand that this is effectively a classification task.

With that out of the way, we just need to calculate weights to pass into the cross_entropy_loss function. This post addresses how to do that with code:

How do I get the accuracy of an unbalanced dataset or segmentation task?

Here is a function for calculating the weight vector used by CrossEntropyLoss based on a set of targets with integers. def scaled_weights(targets): ''' Gets class weights :param targets: pass in the matrix of size (batch, classes) containing correct targets or labels as one_hot :return: returns a vector for weight rescaling for use in cross entropy loss function ''' N, C = targets.size(0), targets.size(1) # get vector of total for each class totals = ta…