Cosine Schedule Mean Value

Problem

Compute the mean learning rate of a cosine annealing schedule over discrete steps.

For step = 0 through total inclusive evaluate lr(step) = min_lr + 0.5 * (base_lr - min_lr) * (1 + cos(pi * step / total)) and return the arithmetic mean of those total + 1 values. If total is not positive, return min_lr.

Hint

The discrete average of the cosine curve sits near the midpoint of the range.

Test cases

case 1

in: [0.1,0,4]

expected: 0.05

case 2

in: [0.1,0.01,10]

expected: 0.05500000000000001

case 3

in: [1,0,100]

expected: 0.49999999999999983

case 4

in: [0.1,0,0]

expected: 0

Starter codePython

def cosine_schedule_mean(base_lr, min_lr, total):
    # Your code here
    pass

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