Something wrong with the Decoder.

[bukhalmae145]

I adjusted Hubert Korean to Grad-SVC and trained the model. But the exported audio file sounds weird and the generated decoder image looks weird.

[yeorinhieut]

At the very least, you need to provide what you fixed and how you fixed it so we can figure out what the problem is. In most cases, it's a user-created problem.

[MaxMax2016]

@bukhalmae145 I think your diffusion is not trained.

full_epochs: 500
fast_epochs: 100

diffusion just starts training after fast_epochs are finished.

[MaxMax2016]

@bukhalmae145 I recommend that you find a partner who knows deep learning to help you, and he is at your side.

[bukhalmae145]

At the very least, you need to provide what you fixed and how you fixed it so we can figure out what the problem is. In most cases, it's a user-created problem.

https://github.com/PlayVoice/so-vits-svc-5.0/issues/130

위 이슈 내용이랑 똑같이 수정해서 훈련은 되는데 결과가 요상하네요.. 생성된 wav 파일에서 핑크노이즈 같은 것만 들리네요. 저는 이쪽 분야에 대해서 문외한이고 github문화도 잘 모릅니다.. 혹시 더 이상 여쭤보는게 실례라면 그만하겠습니다..

[bukhalmae145]

@bukhalmae145 I think your diffusion is not trained.

full_epochs: 500
fast_epochs: 100

diffusion just starts training after fast_epochs are finished.

I just found out that leaving blank in the pretrain: line from base.yaml causes this issue. But I remember that you told me that I should leave it blank from the previous discussion. Is there any way to build pretrained model based on Korean Hubert? https://github.com/PlayVoice/so-vits-svc-5.0/issues/130

[bukhalmae145]

@bukhalmae145 I think your diffusion is not trained.

full_epochs: 500
fast_epochs: 100

diffusion just starts training after fast_epochs are finished.

I found out that it shows /AppleInternal/Library/BuildRoots/495c257e-668e-11ee-93ce-926038f30c31/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShaders/MPSNDArray/Kernels/MPSNDArrayConvolutionA14.mm:4352: failed assertiondestination kernel width and filter kernel width mismatch' [1] 4044 abort python gvc_trainer.py

~/Music/Grad-SVC-20230930 ....... ABRT | took 7s | Grad-SVC py | at 22:57:12

/Library/Frameworks/Python.framework/Versions/3.8/lib/python3.8/multiprocessing/resource_tracker.py:216: UserWarning: resource_tracker: There appear to be 3` error when it starts to diff train on mac no matter what Hubert Model I use.

[MaxMax2016]

It's hard for me to identify your problem，sorry.

[bukhalmae145]

It's hard for me to identify your problem，sorry.

I appreciate your kindness!! But is there any way to making pretrained model based on Korean Hubert Model to replace the original gvc.pretrained.pth?

[bukhalmae145]

It's hard for me to identify your problem，sorry.

I finally found out that line 38 of ./grad_extend/train.py hps.grad.dec_dim, hps.grad.beta_min, hps.grad.beta_max, hps.grad.pe_scale).to('mps') causes this issue. It works fine with the cpu but I think MPS device is not compatible with the code.

[bukhalmae145]

It's hard for me to identify your problem，sorry.

Other losses (prior_loss, mel_loss, etc..) are working fine with loss.backward() but only diff_loss makes error with using loss.backward(). (from ./grad_extend/train.py)