Can Each Source Be Trained Separately? - Unable to Train Quality Model

❓ Questions

@adefossez

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?

[Anjok07]

@Anjok07 These are really fantastic values that you have achieved here, well with the number of data in the dataset.

Could you somehow make your trained model available to the community? I think demucs is a great invention, and I would like it to be even better, because of a larger dataset. Would be really great if this could be done somehow.

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.

[Anjok07]

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