search

sp-uhh / sgmse

Score-based Generative Models (Diffusion Models) for Speech Enhancement and Dereverberation
MIT License
520 stars 76 forks source link

The data simulation is not reproducible.. #35

Closed LiChenda closed 11 months ago

LiChenda commented 11 months ago

Dear authors,

Thank you for open-sourcing this awesome work! I have run the data simulation script create_wsj0_reverb.py on two different servers, but the simulated audio files are different. So, how can I make sure that my simulated data is the same as that you used in the paper? np.random.seed(SEED) seems to not work as expected. Is it possible to fix the randomness by providing meta files of the intermediate simulation parameters? Or would you like to provide your file list, so I can make sure we are using the same data?

10 audio files (in the test folder) from the first simulation in server 1:

anechoic/440c0201_357_2.22_-11.9.wav
anechoic/440c0202_352_1.26_-6.1.wav
anechoic/440c0203_407_1.09_-4.1.wav
anechoic/440c0204_372_0.59_-1.3.wav
anechoic/440c0205_392_0.64_-1.4.wav
anechoic/440c0206_340_1.30_-7.3.wav
anechoic/440c0207_367_1.39_-11.6.wav
anechoic/440c0208_385_1.23_-9.0.wav
anechoic/440c0209_350_1.21_-16.7.wav

10 audio files (in the test folder) from the second simulation in server 2:

anechoic/440c0201_346_0.40_-9.1.wav
anechoic/440c0202_397_1.52_-15.0.wav
anechoic/440c0203_365_0.88_-12.8.wav
anechoic/440c0204_330_1.73_-12.4.wav
anechoic/440c0205_356_1.60_-9.0.wav
anechoic/440c0206_351_0.65_-8.4.wav
anechoic/440c0207_355_0.62_-10.3.wav
anechoic/440c0208_368_1.33_-12.0.wav
anechoic/440c0209_336_1.79_-17.0.wav
jmlemercier commented 11 months ago

Hi @LiChenda. np.random.seed() works as expected, it is just that we forgot to sort the wsj0 files when fetching them with glob(), hence the difference order on different machines. Just pushed a change with sorted(glob(...)) which fixes the behaviour.

LiChenda commented 11 months ago

Thank you @jmlemercier !