Focal Loss (Multiclass)

Problem

Compute the multiclass focal loss for one example: FL = -(1 - p_t)^gamma * log(p_t), where p_t is the stable-softmax probability of the target class.

The signature is multiclass_focal_loss(logits, target, gamma). gamma = 0 recovers plain cross-entropy.

Hint

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

case 1

in: [[1,2,3],2,0]

expected: 0.4076059644

case 2

in: [[1,2,3],2,2]

expected: 0.045677799

case 3

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

expected: 0.7324081924

case 4

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

expected: 0.045677799

Starter codePython

import math
def multiclass_focal_loss(logits, target, gamma):
    # Your code here
    pass

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