Softmax Denominator From Logsumexp

Problem

Compute the log-sum-exp denominator of each score row with the stable max-subtraction trick: m + log(sum(exp(x - m))).

The signature is softmax_logsumexp(scores). Return one value per row; this is the logarithm of the softmax denominator.

Hint

Subtracting the row maximum prevents overflow while leaving the log-sum-exp value unchanged.

case 1

in: [[[0,0]]]

expected: [0.6931471806]

case 2

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

expected: [3.4076059644]

case 3

in: [[[1000,1000]]]

expected: [1000.6931471806]

case 4

in: [[[-1,-1,-1]]]

expected: [0.0986122887]

case 5

in: [[[5]]]

expected: [5]

Starter codePython

import math
def softmax_logsumexp(scores):
    # Your code here
    pass

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