Losses

Tips For New Losses

As mentioned in the paper's Eq. 11, one can create new loss functions for a multitude of vision problems. A good starting point is the Table 3 in the paper. The code should not break but some functions might be harder to train than others due to the nature of data and mathematical stability of the problem. We found that architecture has a change in end performance but not generally on the training procedure as long as exploding and vanishing gradients are taken care of.