diffwave-sashimi icon indicating copy to clipboard operation
diffwave-sashimi copied to clipboard

Published 20 hours ago • verified publisher icon albertfgu

Regarding conditional generation on SC09

Open lizaitang opened this issue 2 years ago • 3 comments

Hello, thank you so much for your work, do we have conditional model checkpoint for SC)( currently?

lizaitang avatar Dec 01 '22 03:12 lizaitang

We didn't train any conditional models. SC09 is a dataset for unconditional generation.

albertfgu avatar Dec 02 '22 06:12 albertfgu

Thanks, besides, during the generation process, I encounter a problem that "size mismatch for c_layers.4.layer.D: copying a param with shape torch.Size([512]) from checkpoint, the shape in current model is torch.Size([1, 512])". How to solve it?

lizaitang avatar Dec 12 '22 04:12 lizaitang

Sorry I missed this comment. I'm not sure why you got this error if using this repo at HEAD, because I checked that everything loaded. Note that, according to the README, you need to go to a specific checkpoints branch to load pre-trained models.

If you pasted in a later version of S4 then it's possible to get these errors because of small changes in the model. One could just load and modify the weights to resolve this.

albertfgu avatar Apr 03 '23 18:04 albertfgu