yxlllc/DDSP-SVC

size mismatch

Opened this issue · 4 comments

curious-ai-developer commented

Hi,

I tried training model 5 but got an error for training. Error is:
size mismatch for ddsp_model.window: copying a param with shape torch.size([1024]) from checkpoint, the shape in current model is torch.size([2048])
size mismatch for ddsp_model.unit2ctrl.denseout
Then I ran for model 4 and it works.
What is the problem with model 5?
What are the benefits of model 5?

yxlllc commented

Check whether the configuration file or pre-trained model used is correct

Y-Kitaro commented

I got the same error . You start training without using ‘model0.pt’ or try using this 'model0.pt'
https://github.com/yxlllc/DDSP-SVC/releases/tag/5.0

curious-ai-developer commented

I used pre trained model for training. I still have the problem as Y-Kitaro. Should I change configuration file? Which yaml? How?

Y-Kitaro commented

I started learning without changing the settings in 'config/diffusion-fast.yaml'.
I think your win_length setting or pre-trained model is wrong.
V5.0 'model_0.pt' has the same file name, but the contents are different.