Normalized Focal Loss

Problem

Compute focal loss over a probability map normalized by the number of positives.

Sum -((1 - p_t)^gamma) * log(p_t) over all pixels, where p_t is prob for target 1 and 1 - prob otherwise, clamping p_t to at least 1e-12. Divide by max(1, num_positives) and round to 4 decimal places.

Test cases

case 1

in: [[[0.9,0.1]],[[1,0]],2,1]

expected: 0.0021

case 2

in: [[[0.5]],[[1]],0,2]

expected: 0.3466

case 3

in: [[[0.5]],[[1]],0,0]

expected: 0.6931

Starter codePython

def normalized_focal_loss(prob, target, gamma, num_positives):
    # Your code here
    pass

3 test cases · run your code against them in the editor