Sigmoid Loss

A contrastive learning loss function that computes positive/negative probabilities independently for each pair rather than using global softmax.