About the error reported after the pre-trained model is loaded

[irinaxu]

Thank you very much for sharing!! I used the ckpt shared by the project to try and test, but an error occurred. parameter settings for parser_ in (baseparser, parser): parser.add_argument("--testdir", type=str, default = '/reverb/', help="Directory containing your corrupted files to enhance.") parser.add_argument("--enhanced_dir", type=str, default = '/reverbout/', help="Where to write your cleaned files.") parser.addargument("--ckpt", type=str,default = 'logs/epoch=237-pesq=0.00.ckpt') parser.add_argument("--mode",default="regen", choices=["score-only", "denoiser-only", "regen-freeze-denoiser", "regen-joint-training"])

parser_.add_argument("--corrector", type=str, choices=("ald", "langevin", "none"), default="ald", help="Corrector class for the PC sampler.")
parser_.add_argument("--corrector-steps", type=int, default=1, help="Number of corrector steps")
parser_.add_argument("--snr", type=float, default=0.5, help="SNR value for (annealed) Langevin dynamics.")
parser_.add_argument("--N", type=int, default=50, help="Number of reverse steps")

error：

[jmlemercier]

Dear Irina, thanks for reporting: I had indeed changed one line in ncsnpp.py to remove some modules which are useless when using the NCSN++ backbone in discriminative mode, but that messed up the checkpoint compatibility (I just reverted, because I could not fix it easily as the ncsnpp class is taken from Song's code and is very messy, will consider rewriting it in the future)

Best, Jm