Attention Weight Sparsity Fraction

Problem

Measure attention sparsity as the fraction of weights strictly below a threshold.

The signature is attention_sparsity_fraction(weights, threshold). Count entries with p < threshold across all rows and divide by the total number of entries.

Hint

The comparison is strict, so a weight exactly equal to the threshold is not counted.

Test cases

case 1

in: [[[0.5,0.5]],0.1]

expected: 0

case 2

in: [[[0.05,0.95],[0,1]],0.1]

expected: 0.5

case 3

in: [[[1]],1]

expected: 0

case 4

in: [[[0,0],[0,0]],0.5]

expected: 1

Starter codePython

def attention_sparsity_fraction(weights, threshold):
    # Your code here
    pass

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