Min-SNR Gamma Weight

Problem

Compute the min-SNR loss weight: min(SNR, gamma) / SNR, which caps the emphasis on high-SNR (easy) timesteps.

The signature is min_snr_weight(alpha_bar_t, gamma=5.0).

Hint

The cap prevents easy low-noise steps from dominating the training gradient.

case 1

in: [0.5,5]

expected: 1

case 2

in: [0.9,5]

expected: 0.5555555556

case 3

in: [0.25,5]

expected: 1

case 4

in: [0.99,5]

expected: 0.0505050505

Starter codePython

def min_snr_weight(alpha_bar_t, gamma=5.0):
    # Your code here
    pass

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