Keras weighted log loss

weighted_log_loss.py

def weighted_log_loss(yt, yp):
    '''Log loss that weights false positives or false negatives more.
    Punish the false negatives if you care about making sure all the neurons
    are found and don't mind some false positives. Vice versa for punishing
    the false positives. Concept taken from the UNet paper where they
    weighted boundary errors to get cleaner boundaries.'''

    emphasis = 'fn'
    assert emphasis in ['fn', 'fp']
    m = 2

    # Apply the multiplier to y_true.
    if emphasis == 'fn':
        # [0,1] -> [0,m].
        w = yt * m
    elif emphasis == 'fp':
        # [0,1] -> [-1,0] -> [1,0] -> [m-1,0] -> [m,1]
        w = ((yt - 1) * -1) * (m - 1) + 1

    a = yt * K.log(yp + K.epsilon())
    b = (1 - yt) * K.log(1 + K.epsilon() - yp)
    return -1 * K.mean(w * (a + b))
    # Below return statement would match the keras binary_crossentropy value.
    # return -1. * K.mean((a + b))