Reference Model Drift Check

Problem

Measure drift of a policy from its reference model.

Return [mean(|ref_log_probs[i] - log_probs[i]|), drift > max_drift]. An empty batch returns [0.0, False].

Hint

Monitoring drift catches a policy that has silently wandered away from its reference.

case 1

in: [[-1,-2],[-1,-2],0.5]

expected: [0,false]

case 2

in: [[0],[-1],0.5]

expected: [1,true]

case 3

in: [[-1,-3],[-2,-2],0.5]

expected: [1,true]

case 4

in: [[],[],0.5]

expected: [0,false]

Starter codePython

def reference_drift(log_probs, ref_log_probs, max_drift):
    # Your code here
    pass

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