BYOL-Explore Loss Lite

Problem

Compute the BYOL-style cosine loss 2 - 2 * cos(predicted, target). If either vector has zero norm, return 2.0 as the maximum loss.

Hint

The predictor is trained to match a slowly updated target embedding.

case 1

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

expected: 0

case 2

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

expected: 2

case 3

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

expected: 4.440892098500626e-16

case 4

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

expected: 2

Starter codePython

import math
def byol_explore_loss(predicted, target):
    # Your code here
    pass

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