RVC / Onnx usage, lack of information + lack of documentation. about ailia-models HOT 4 CLOSED

If you want to use rvc v2, please use ‘—version 2’ option for onnx inference.
You can use official onnx converter of rvc webui.
We are supporing with/without f0, and rvc v1/v2.
You can get onnx2prototxt.py from ailia SDK pacakage.
ailia SDK 1.2.16 supports opset <=17.

from ailia-models.

Also you can omit prototxt file by set None instead “xxx.onnx.prototxt” to ailia.Net constructor.

from ailia-models.

RVC version : 58370b048c580a34d33be4fde402bf847a9e3622 2023年7月26日 19:53:31

Training settings:

ONNX conversion:

Generate prototxt:

python3 onnx2prototxt.py input.onnx

Inference :

python3 rvc.py -m input.onnx --f0_method crepe_tiny --f0 1 --tgt_sr 48000 --version 2 

from ailia-models.

If you want to use rvc v2, please use ‘—version 2’ option for onnx inference. You can use official onnx converter of rvc webui. We are supporing with/without f0, and rvc v1/v2. You can get onnx2prototxt.py from ailia SDK pacakage. ailia SDK 1.2.16 supports opset <=17.

Well tbf, it's pointless, moving over onnx anyway. The current exporting method isn't dynamic, it's static and honestly, for that little of speed boost, losing accuracy / quality is def for a minus. For true potential of pytorch -> onnx in terms of rvc, we'd need dynamo exporting, sadly devs don't really care much about further rvc development, so it seems recently. Tho, thanks for info anyway, might come in handy in future, maybe.

from ailia-models.