CLIP Learned Logit Scale

Problem

Compute the CLIP logit scale from its learned parameter.

CLIP parameterizes the temperature as a scalar and uses logit_scale = min(exp(raw_scale), max_scale), where the cap keeps the similarity logits from becoming too sharp (100 in the original implementation). Return the scale rounded to 4 decimal places.

Test cases

case 1

in: [0,100]

expected: 1

case 2

in: [2.302585,100]

expected: 10

case 3

in: [5,100]

expected: 100

case 4

in: [-4.60517,100]

expected: 0.01

Starter codePython

def clip_logit_scale(raw_scale, max_scale):
    # Your code here
    pass

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