Activation Int8 Calibration

Problem

Calibrate an activation clipping threshold with the nearest-rank percentile: sort the observed activations and take index ceil(percentile / 100 * n) - 1, clamped to the valid range.

The signature is activation_calibration(activations, percentile). Return the threshold value.

Hint

Percentile clipping removes extreme outliers that would waste int8 dynamic range.

Test cases

case 1

in: [[1,2,3,4,5],90]

expected: 5

case 2

in: [[0.1,0.2],50]

expected: 0.1

case 3

in: [[10],99]

expected: 10

case 4

in: [[-1,0,1],50]

expected: 0

Starter codePython

import math
def activation_calibration(activations, percentile):
    # Your code here
    pass

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