Batch Padding Waste Ratio

Problem

Compute the fraction of padding tokens in a padded batch: 1 - sum(lengths) / (num_sequences * max_length).

The signature is padding_waste_ratio(lengths, max_length). Return a float.

Hint

Length bucketing reduces this waste by grouping similar-length sequences.

case 1

in: [[10,20],20]

expected: 0.25

case 2

in: [[20,20],20]

expected: 0

case 3

in: [[0],10]

expected: 1

case 4

in: [[5],10]

expected: 0.5

Starter codePython

def padding_waste_ratio(lengths, max_length):
    # Your code here
    pass

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