Closed manu0586 closed 3 years ago
Hello thanks for reporting, I took down the model for now and will upload a corrected version.
Hi, has this issue solved? C:
Not yet, I'm training the corrected model and will release it when its ready. I will post the link here and close the issue when it's done
Training a model like this on LibriMix takes 5-7 days. Sorry for the inconvenience
Can we explain what was the problem for loading the previous one?
Le ven. 26 févr. 2021 à 12:19, JorisCos notifications@github.com a écrit :
Training a model like this on LibriMix takes 5-7 days. Sorry for the inconvenience
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/asteroid-team/asteroid/issues/450#issuecomment-786585978, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEND2HEXUDEV4DHAIWAGMATTA57VDANCNFSM4YCVHD6A .
The first Large-DCUNet-20
architecture wasn't implemented correctly we corrected it. I trained on a different Large-DCUNet-20
architecture before we corrected the current version. The correction ended up in small size changes in the last 2 decoders layers compared to the architecture that I used. This why we can't load the models size mismatch
asteroid version : 0.4.4
function call :
model = DCUNet.from_pretrained("JorisCos/DCUNet_Libri1Mix_enhsingle_16k")
Error :information from cached_model (via cached_download function + torch.load) model_args= {'architecture': 'Large-DCUNet-20', 'stft_kernel_size': 1024, 'stft_stride': 256, 'sample_rate': 16000.0, 'fix_length_mode': 'pad', 'n_src': 1}