Tanh Backward

Problem

Backpropagate through tanh. For each pre-activation z[i], return dout[i] * (1 - tanh(z[i])^2).

The signature is tanh_grad(dout, z).

Hint

Compute tanh(z) once per element, then use 1 - tanh(z)^2.

case 1

in: [[1],[0]]

expected: [1]

case 2

in: [[1],[1]]

expected: [0.4199743416]

case 3

in: [[2,3],[-1,0.5]]

expected: [0.8399486832,2.3593431989]

case 4

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

expected: [0,0]

Starter codePython

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

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