MediumCalculusca-285

Multiclass Hinge Gradient

Problem

For the multiclass hinge loss with scores s and correct label index y, the loss is max_j (1 + s_j - s_y) over j != y, capped at zero.

The gradient with respect to s is +1 at the best violating class, -1 at the correct class (when a violation exists), and zeros elsewhere. Given scores and the label index, return the gradient list.

Hint

Only the single highest-scoring violator receives the positive gradient.

Test cases

case 1

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

expected: [0,0,0]

case 2

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

expected: [-1,1,0]

case 3

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

expected: [1,-1,0]

Starter codePython

def multiclass_hinge_gradient(scores, label):
    # Your code here
    pass

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