Closed tomakorea closed 3 weeks ago
I was seeing this same issue whenever I tried to run Applio on Paperspace machines which use Linux. The feature extraction is slow to the point of being unusable.
So I was considering using the main RVC branch for that pre-process which is super fast using RMVPE GPU but I'm not 100% sure it's compatible with Applio...
We added the feature to set the custom cores you want to use in pitch extraction try a higher value. Commit: https://github.com/IAHispano/Applio/commit/5e302727f2b844e33503afb6cdba3a1945c76527
We added the feature to set the custom cores you want to use in pitch extraction try a higher value. Commit: 5e30272
I gave it a go, but even when giving it all 8 CPU cores of the Linux machine it still extracts incredibly slowly
But the normal rmvpe(not rmvpe gpu) it's also slow in the main branch?
But the normal rmvpe(not rmvpe gpu) it's also slow in the main branch?
Would it be possible to add RMVPE GPU for Applio in that case?
right now, no
I'm using Applio 3.2.0 but this behavior was there on previous versions too. I tested on both Windows 11 and Ubuntu with the same hardware, for pitch extraction, Windows is much faster for doing the RMVPE pitch extraction process before starting the training. For example, on a large dataset, I gave up after 4 hours of pitch extraction on Linux on an i7 8700k, while windows did it in about 6 to 10 minutes. However, usually training is a bit faster on Linux, maybe because of the windows overhead.
Steps to Reproduce
Expected Behavior I expected to take the same amount of time as on Windows. Also, it's not really convenient to switch OS to do only this job since the filelist.txt has all the audio files path, so it means it's necessary to edit this file and replace all the path if switching OS are necessary to save time.
Desktop Details: