Label Smoothing

Problem

Build smoothed one-hot targets: the true class gets 1 - epsilon and every other class gets epsilon / (num_classes - 1).

Return the list of rows; empty input returns [].

Hint

Redistribute epsilon mass equally to the non-target classes.

case 1

in: [[0,2],3,0.3]

expected: [[0.7,0.15,0.15],[0.15,0.15,0.7]]

case 2

in: [[1],2,0.2]

expected: [[0.2,0.8]]

case 3

in: [[0],1,0.1]

expected: [[0.9]]

case 4

in: [[],3,0.1]

expected: []

Starter codePython

def label_smoothing(y_true, num_classes, epsilon):
    # Your code here
    pass

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