Temperature Scaling

Problem

Apply temperature scaling to logits: return the probabilities softmax(logits / temperature).

If logits is empty or temperature <= 0, return [].

Hint

Divide every logit by the temperature before applying a numerically stable softmax.

case 1

in: [[2,0],0.5]

expected: [0.9820137900379085,0.017986209962091555]

case 2

in: [[2,0],10]

expected: [0.549833997312478,0.4501660026875221]

case 3

in: [[1,1],5]

expected: [0.5,0.5]

case 4

in: [[2,0],1]

expected: [0.8807970779778823,0.11920292202211755]

Starter codePython

import math
def temperature_scaled_softmax(logits, temperature):
    # Your code here
    pass

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