4

I am implementing a code for semantic segmentation using Keras and I wrote my loss function as in the paper "Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations" (link: https://arxiv.org/abs/1707.03237) to balance each class. My data are organized as (bacth_size, ImDim1, ImDim2, Nclasses). My loss function is:

eps = 1e-3

def dice(y_true, y_pred):
    weights = 1./K.sum(y_true, axis=[0,1,2])
    weights = weights/K.sum(weights)
    num = K.sum(weights*K.sum(y_true*y_pred, axis=[0,1,2]))
    den = K.sum(weights*K.sum(y_true+y_pred, axis=[0,1,2]))
    return 2.*(num+eps)/(den+eps)

def dice_loss(y_true, y_pred):
    return 1-dice(y_true, y_pred)

Doing in this way, that looks correct to me, the loss function returns nan and I do not get why!?

Marco Domenico Cirillo
  • 43
  • 4
  • have you tried to add ```eps``` in the first equation? If all ```y_true``` are 0, ```weights``` should be NaN. – anki Sep 18 '19 at 08:00

0 Answers0