Mixup Loss from Logits

Problem

Mixup training loss for one example: lam * CE(logits, y_a) + (1 - lam) * CE(logits, y_b), using stable cross-entropy for each label.

The signature is mixup_loss(logits, y_a, y_b, lam).

Hint

The mixup coefficient blends the two one-hot cross-entropy terms.

case 1

in: [[1,2,3],2,0,0.7]

expected: 1.0076059644

case 2

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

expected: 0.4076059644

case 3

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

expected: 2.4076059644

case 4

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

expected: 0.6931471806

Starter codePython

import math
def mixup_loss(logits, y_a, y_b, lam):
    # Your code here
    pass

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