Does rvc support training with cpu?

[c469591]

Hello, may I ask if RVC supports training with CPU? Because I don't have a separate graphics card, and I encountered an error in the second step of training, which showed "slow_conv2d_cpu" not implemented for 'Half'. Thank you!

[kalomaze]

CPU inference of most AI models is already pretty slow as is. I don't think there is any popular machine learning software that has CPU training as an option. It would probably take weeks to get a decent result.

[fumiama]

Maybe add later.

[justinjohn0306]

@c469591 use google colab

[fumiama]

Yes, that's also a solution.

[c469591]

I still hope this will be added to the new version. Thanks!

[Sovenok-Hacker]

I still hope this will be added to the new version. Thanks!

Yes. I have the same problem. I want to train a model but I don`t have a GPU.

[RonFusaishi]

considering google now banning RVC code, i think something like this is better than nothing.

[Sovenok-Hacker]

considering google now banning RVC code, i think something like this is better than nothing.

Google is nuts

[RonFusaishi]

Google is nuts

google and greed are one of the same thing unfortunately, you can still run it but only if you have a colab pro.

I personally blame the influx of zoomers abusing the damn thing just to spam youtube with their degenerate garbage, i saw this coming from a mile away.

[yapvoonyee]

I would like a cpu only version because I have an intel GPU.

[fumiama]

I would like a cpu only version because I have an intel GPU.

You can try the dml or wait #278

[loboere]

I have a GPU and I want to infer with a CPU, because my GPU gives cuda errors, is there any parameter so that RVC does not detect the GPU?

[LoloCodes]

so can we use cpus to train or not?

[treya-lin]

It can. At least worked when tested on my linux server. My server actually has gpu on it. To disable gpu, I added os.environ["CUDA_VISIBLE_DEVICES"] = "-1" before importing torch in infer/modules/train/train.py

from infer.lib.train import utils

hps = utils.get_hparams()
# os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
os.environ["CUDA_VISIBLE_DEVICES"] = "-1" # set this to "-1" if you wanna test cpu training on a server with gpu. Training with cpu does work, but it is extremely slow.

As for the speed, according to my very brief experience, for an epoch which costs 20s on a 1080 Ti GPU, it took about 2min on cpu.

[Axis4s]

It can. At least worked when tested on my linux server. My server actually has gpu on it. To disable gpu, I added os.environ["CUDA_VISIBLE_DEVICES"] = "-1" before importing torch in infer/modules/train/train.py

from infer.lib.train import utils

hps = utils.get_hparams()
# os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
os.environ["CUDA_VISIBLE_DEVICES"] = "-1" # set this to "-1" if you wanna test cpu training on a server with gpu. Training with cpu does work, but it is extremely slow.

As for the speed, according to my very brief experience, for an epoch which costs 20s on a 1080 Ti GPU, it took about 2min on cpu.

slow but still useful to ppl who wanna make models

[Axis4s]

I would like a cpu only version because I have an intel GPU.

You can try the dml or wait #278

what is 'dml'?

[thefellaguy]

I'm know I'm late to the game, but I've trained on CPU only with no issues. I let it run on all night on 6 cores and it gets the job done in 8-10 hours on harvest. Converting audio is about a 1:2 ratio. a 10 audio clip tales about 20 seconds to convert.