Linear Batch Scaling Rule

Problem

Apply the linear batch scaling rule for learning rates.

Return base_lr * new_batch / base_batch so the learning rate grows in proportion to the batch size. If base_batch is not positive, return base_lr unchanged.

Hint

Doubling the batch doubles the learning rate under this rule.

Test cases

case 1

in: [0.1,32,128]

expected: 0.4

case 2

in: [0.1,32,32]

expected: 0.1

case 3

in: [0.2,64,16]

expected: 0.05

case 4

in: [0.1,0,5]

expected: 0.1

Starter codePython

def linear_scaling_lr(base_lr, base_batch, new_batch):
    # Your code here
    pass

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