Gradient Noise Injection

Problem

Add reproducible Gaussian noise to a gradient vector.

Call random.seed(seed), then add an independent random.gauss(0.0, std) draw to every component. Return the noisy gradient as a list. An empty input returns [].

Hint

Reseeding with the same seed reproduces the exact same noise draws.

Test cases

case 1

in: [[1,2,3],0.1,0]

expected: [1.0941715404680665,1.860342189529885,2.932028555192158]

case 2

in: [[0.5,-0.5],0.2,7]

expected: [0.4488239423104799,-0.3977136974966972]

case 3

in: [[0],1,42]

expected: [-0.14409032957792836]

case 4

in: [[],0.1,3]

expected: []

Starter codePython

def add_gradient_noise(grads, std, seed):
    # Your code here
    pass

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