Integrated Gradients Step

Problem

Evaluate one integrand term of integrated gradients: (x - baseline) * grad, where grad is the gradient of the model output at one interpolation point between the baseline and the input.

The signature is integrated_gradients_step(x, baseline, grad). Return the element-wise product.

Hint

The full attribution averages this term over many points along the path from baseline to input.

Test cases

case 1

in: [[1,2],[0,0],[0.5,0.5]]

expected: [0.5,1]

case 2

in: [[1,1],[1,1],[5,5]]

expected: [0,0]

case 3

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

expected: [8,0]

case 4

in: [[0.5],[0],[1]]

expected: [0.5]

Starter codePython

def integrated_gradients_step(x, baseline, grad):
    # Your code here
    pass

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