cap_pos and cap_neg

[berkuva]

When working with imbalanced data, it may be possible that a batch of the training data may not have a positive or negative sample (depending on in which way the imbalance is). That causes cap_pos or cap_neg to become 0, and epoch_pos or epoch_neg can't be calculated. Would it be reasonable to do something like to following?

cap_pos = max(1, epoch_pos.shape[0])
cap_neg = max(1, epoch_neg.shape[0])