Sigmoid Backward

Problem

Backpropagate through a sigmoid. For each pre-activation z[i], compute s = sigmoid(z[i]) and return dout[i] * s * (1 - s).

The signature is sigmoid_grad(dout, z). Use a numerically stable sigmoid.

Hint

The sigmoid derivative is s * (1 - s); multiply by the upstream gradient.

case 1

in: [[1],[0]]

expected: [0.25]

case 2

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

expected: [0.25,0.5]

case 3

in: [[2],[1]]

expected: [0.3932238665]

case 4

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

expected: [0.1966119332]

case 5

in: [[1,1],[-1000,1000]]

expected: [0,0]

Starter codePython

import math
def sigmoid_grad(dout, z):
    # Your code here
    pass

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