Unable to train models

[5909c]

I understand that this isn't really an bug, more how the software is written, but I run on AMD hardware, both CPU and GPU, and there's no way for me to run the Cuda code that RVC uses, the web UI doesn't pick up even my CPU, if there is any kind of workaround or fix for this, I'd be very grateful, because I'm currently unable to train vocies at all.

[ns-studios]

Same here! I can generate indexes and features, but not the pth file

[Tommy2003-star]

Same here! I can generate indexes and features, but not the pth file

This is exactly the problem I'm facing. Can you guys help us with this problem?

[TheTrustedComputer]

Me too! I have an all-AMD system running Linux, and training requires a modern NVIDIA GPU, which I don't have and can't afford. However, I do have a very old GTX 550 Ti lying around, but I don't think it has a version of CUDA that RVC needs.

Please support training on non-NVIDIA cards as much as you can. I'm dying to train my own models and test them against a variety of audio samples!

[HardikJain02]

Same here! I can generate indexes and features, but not the pth file

Did you get it fixed?

[ns-studios]

No. I had to use a cloud computer option that had a good enough Nvidia GPU.

Did you get it fixed?

[HardikJain02]

What causes this problem exactly?

[ns-studios]

It seems to fail silently at this: use gpus:
runtime\python.exe train_nsf_sim_cache_sid_load_pretrain.py -e myvoice -sr 40k -f0 1 -bs 1 -te 20 -se 5 -pg pretrained/f0G40k.pth -pd pretrained/f0D40k.pth -l 0 -c 0 -sw 0 -v v1
If I give it GPU 0-1: use gpus: 0-1
runtime\python.exe train_nsf_sim_cache_sid_load_pretrain.py -e myvoice -sr 40k -f0 1 -bs 1 -g 0-1 -te 20 -se 5 -pg pretrained/f0G40k.pth -pd pretrained/f0D40k.pth -l 0 -c 0 -sw 0 -v v1

[PKLR7]

No. I had to use a cloud computer option that had a good enough Nvidia GPU.

Did you get it fixed?

Can you elaborate on how you did that

[ns-studios]

@PKLR7 I rented a machine from a vps provider. The latest version of rvc I tried (RVC0813AMD_Intel), actually let me train using my regular computer (i7 6th gen, Intel HD Graphics 520, and AMD Radeon R7 M360), although looking at task manager, the Python process seemed not to use any GPU power/dedicated memory during the model training part of the process, for some reason, which may explain why it took about an hour per epoch. It did seem to use it for feature extraction, though. However, this version doesn't work at all for rendering; it just immediately fails with any of the algorithms when I use model inference.

[github-actions[bot]]

This issue was closed because it has been inactive for 15 days since being marked as stale.