Can Each Source Be Trained Separately? - Unable to Train Quality Model
See original GitHub issue❓ Questions
Thank you for your help in https://github.com/facebookresearch/demucs/issues/221 !
I’m running into a new problem.
I was able to get training started by using the following dset settings:
dset:
wav: C:/Users/Anjok/Desktop/Demucs-Train/test <------ Dataset location
sources: ['vocals', 'non_vocals']
samplerate: 44100
channels: 2
epochs: 320
batch_size: 6 <--------- I'm using an RTX 3090 with 24GB's of V-RAM. This was the largest batch size I was able to use.
weights: [1.,1.] <--------- I had to decrease the number of weights from 4 to 2.
I was able to train a model, but it comes out very poorly and with much distortion. The dataset consists of 1400 tracks and I’ve been able to successfully train a strong vocal model that achieved a 9.703 SDR on AIcrowds testset using the KUIELAB-MDX-Net code, so I know there isn’t an issue with the dataset. I’m wondering if having to decrease the number of weights is causing this issue? I also tried training with the vocal stem and mixture only using the following settings, but the results were even worse.
dset:
wav: C:/Users/Anjok/Desktop/Demucs-Train/test
sources: ['vocals']
samplerate: 44100
channels: 2
epochs: 320
batch_size: 6
weights: [1.] <--------- Decreased the number of weights from 2 to 1.
I’m eager to train Demucs using our dataset because I know it’s solid, but I can’t seem to get a good model yet. What would you recommend? Perhaps I need to be using a different settings? I did a test with a small dataset consisting of 4 stem models as well, and the results were very poor. What exact settings were used to train mdx_extra?
Issue Analytics
- State:
- Created 2 years ago
- Comments:6 (1 by maintainers)
Top GitHub Comments
I plan on it! It’s going to take time though since the dataset is so big and each epoch takes about 9 hours. I recently had to restart the session from scratch due to a small mistake in my code. It’ll likely be a little over a month before I release it.
These models have been released and added to the UVR GUI. Check it out here! - https://github.com/Anjok07/ultimatevocalremovergui/releases/tag/v5.3.0