Focal Loss

Problem

Compute the mean binary focal loss: -alpha_t * (1 - p_t)^gamma * log(p_t), where p_t = p for y = 1 else 1 - p, and alpha_t = alpha for y = 1 else 1 - alpha.

Probabilities are clipped to [1e-15, 1 - 1e-15]. Empty input returns 0.0.

Hint

The (1 - p_t)^gamma factor down-weights easy, well-classified examples.

Test cases

case 1

in: [[1,0],[0.9,0.9],2,0.25]

expected: 0.6995419226415137

case 2

in: [[1,1],[0.6,0.6],0,0.5]

expected: 0.25541281188299536

case 3

in: [[0,1],[0.3,0.3],1,0.5]

expected: 0.23744586155244374

case 4

in: [[],[],2,0.25]

expected: 0

Starter codePython

import math
def focal_loss(y_true, y_pred, gamma, alpha):
    # Your code here
    pass

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