Tutorials

[rkeane]

tried to train but got this error at the end

1 +- 03_bill_mix.mp3 +- 03_bill_inst.mp3
2 +- 04_fasc_mix.mp3 +- 04_fasc_inst.mp3
3 +- 01_amd_mix.mp3 +- 01_amd_inst.mp3
4 +- 02_beat_mix.mp3 +- 02_beat_inst.mp3
0%| | 0/4 [00:00 warnings.warn('PySoundFile failed. Trying audioread instead.')
C:\Users\Robert\AppData\Local\Programs\Python\Pyth on37\lib\site-packages\librosa\core\audio.py:161: UserWarning: PySoundFile failed. Trying audioread instead.
warnings.warn('PySoundFile failed. Trying audioread instead.')
100%|█████████████████████████████████████████████ ███████████████████████████████████████| 4/4 [01:53<00:00, 28.25s/it]
0it [00:00, ?it/s]
# epoch 0
* inner epoch 0
Traceback (most recent call last):
File "train.py", line 223, in
main()
File "train.py", line 194, in main
X_train, y_train, model, optimizer, args.batchsize, instance_loss)
File "train.py", line 75, in train_inner_epoch
return sum_loss / len(X_train)
ZeroDivisionError: division by zero

[Anjok]

tried to train but got this error at the end

1 +- 03_bill_mix.mp3 +- 03_bill_inst.mp3
2 +- 04_fasc_mix.mp3 +- 04_fasc_inst.mp3
3 +- 01_amd_mix.mp3 +- 01_amd_inst.mp3
4 +- 02_beat_mix.mp3 +- 02_beat_inst.mp3
0%| | 0/4 [00:00 warnings.warn('PySoundFile failed. Trying audioread instead.')
C:\Users\Robert\AppData\Local\Programs\Python\Pyth on37\lib\site-packages\librosa\core\audio.py:161: UserWarning: PySoundFile failed. Trying audioread instead.
warnings.warn('PySoundFile failed. Trying audioread instead.')
100%|█████████████████████████████████████████████ ███████████████████████████████████████| 4/4 [01:53<00:00, 28.25s/it]
0it [00:00, ?it/s]
# epoch 0
* inner epoch 0
Traceback (most recent call last):
File "train.py", line 223, in
main()
File "train.py", line 194, in main
X_train, y_train, model, optimizer, args.batchsize, instance_loss)
File "train.py", line 75, in train_inner_epoch
return sum_loss / len(X_train)
ZeroDivisionError: division by zero

This error is due to your training set being too small. You need a bare minimum of 15 pairs in order to start training. Also, if you're training from scratch like this you'll need at LEAST 50-75 pairs for it to be effective at all. Your training/validation numbers won't move with sets any lower than 50; You'll end up wasting your system resources and being sorely disappointed with your models' performance.

If you choose to train with a set between 15-50 pairs, just finetune one of the baseline models (commands in the main thread). I figured out how to train effectively with a GPU, so train with your GPU if you have one.

[rkeane]

GPU not much cop mate.....
so how does it work after you have trained it
does it recognize wether its a rock song etc

[Anjok]

A new model has been posted to the main page! Please make sure to use it with the new A.I. provided as it won't work with the old one.

[halofan253]

Hey Anjok! First of all thank you for this awesome AI, it works really well and does a great job separating the tracks.
But now I have a problem with the new model uploaded.
When I tried to run using GPU I get the following error:

Traceback (most recent call last):
File "inference.py", line 104, in
main()
File "inference.py", line 64, in main
pred = model.predict(X_window)
File "C:\Users\KennA\Documents\vocal-removerV2\lib\nets.py", line 79, in predict
h = self.full_band_net(self.bridge(h))
File "C:\Users\KennA\Documents\vocal-removerV2\lib\nets.py", line 34, in __call__
h = self.dec1(h, e1)
File "C:\Users\KennA\Documents\vocal-removerV2\lib\layers.py", line 79, in __call__
x = spec_utils.crop_center(x, skip)
File "C:\Users\KennA\Documents\vocal-removerV2\lib\spec_utils.py", line 20, in crop_center
return torch.cat([h1, h2], dim=1)
RuntimeError: CUDA out of memory. Tried to allocate 384.00 MiB (GPU 0; 2.00 GiB total capacity; 948.49 MiB already allocated; 308.74 MiB free; 137.51 MiB cached)

This didn't happened with the old version. There's a way to solve this? Because using CPU is reaaaally slow. Thank you!

[Anjok]

Hey Anjok! First of all thank you for this awesome AI, it works really well and does a great job separating the tracks.
But now I have a problem with the new model uploaded.
When I tried to run using GPU I get the following error:

This didn't happened with the old version. There's a way to solve this? Because using CPU is reaaaally slow. Thank you!

You're welcome! I'm glad you've enjoyed it!

To answer your question, this new model is bigger and has more layers, so is requires more V-RAM. Your GPU might not have enough memory for this one sadly :(

[ChrisCall]

I made it to the conversion step and then got an error I can't figure out;

C:\Users\xxxx\Documents\vocal-remover>python inference.py --input Daredevil.mp3 --gpu 0
C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\librosa\util\decorators.py:9: NumbaDeprecationWarning: An import was requested from a module that has moved location.
Import requested from: 'numba.decorators', please update to use 'numba.core.decorators' or pin to Numba version 0.48.0. This alias will not be present in Numba version 0.50.0.
from numba.decorators import jit as optional_jit
C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\librosa\util\decorators.py:9: NumbaDeprecationWarning: An import was requested from a module that has moved location.
Import of 'jit' requested from: 'numba.decorators', please update to use 'numba.core.decorators' or pin to Numba version 0.48.0. This alias will not be present in Numba version 0.50.0.
from numba.decorators import jit as optional_jit
loading model... done
C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\librosa\core\audio.py:161: UserWarning: PySoundFile failed. Trying audioread instead.
warnings.warn('PySoundFile failed. Trying audioread instead.')
loading wave source... Traceback (most recent call last):
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\librosa\core\audio.py", line 129, in load
with sf.SoundFile(path) as sf_desc:
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\soundfile.py", line 629, in __init__
self._file = self._open(file, mode_int, closefd)
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\soundfile.py", line 1184, in _open
"Error opening {0!r}: ".format(self.name))
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\soundfile.py", line 1357, in _error_check
raise RuntimeError(prefix + _ffi.string(err_str).decode('utf-8', 'replace'))
RuntimeError: Error opening 'Daredevil.mp3': File contains data in an unknown format.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "inference.py", line 104, in
main()
File "inference.py", line 39, in main
args.input, args.sr, False, dtype=np.float32, res_type='kaiser_fast')
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\librosa\core\audio.py", line 162, in load
y, sr_native = __audioread_load(path, offset, duration, dtype)
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\librosa\core\audio.py", line 186, in __audioread_load
with audioread.audio_open(path) as input_file:
File "C:\Users\xxxx\AppData\Local\Programs\Python\Python 37\lib\site-packages\audioread\__init__.py", line 116, in audio_open
raise NoBackendError()
audioread.exceptions.NoBackendError

[djtayz]

I made it to the conversion step and then got an error I can't figure out;

...

Happened to me too. I wrote pip3 install numba==0.48.0 (or change it to 0.50.0) and I think that fixed it for me .

[ChrisCall]

Yep, that fixed it! Thanks :) I also needed to use wav files (mp3's won't work, at least not on my system).

I threw a couple of the toughest conversions that I know of at it to see. On the whole, the primary AI over in the other topic is superior, at least presently, in that this one has more trace vocal across the tracks, like a lingering echo instead of the static the other one gives ... but both are in the same ballpark, which is way ahead of any other program.

One fascinating thing though, is that this one actually seems to handle some things BETTER than the other.. though I'd need to do more testing to see. I'd say this is the exception, not the rule ... but... For example, Steven Wilson - Blackest Eyes it does a poorer job of the verse sections, but does a superior job on the bridge. On Smashing Pumpkins - JellyBelly it does a poorer job on the overall vocals, since there is trace bleed here, and the other AI eliminates it entirely, ... but on certain parts of the song, the other AI completely fails to remove any vocals at all, and this one does not do a perfect job by any stretch, but it does a noteworthy better job on those parts.

I've had very little time to test, and I know that there are more builds to come (which I look forward to), just very early observations. Genre specific models fascinate me as well ... what if that one part in a Rock song converts better with a Pop oriented model and can be sliced in with the rest of the song converted with the Rock model to create a complete product? I'm already getting that vibe just comparing this model with the other one. Even failing that, more diverse coverage of quality results is likely.

[Anjok]

Genre specific models fascinate me as well ... what if that one part in a Rock song converts better with a Pop oriented model and can be sliced in with the rest of the song converted with the Rock model to create a complete product? I'm already getting that vibe just comparing this model with the other one. Even failing that, more diverse coverage of quality results is likely.

This is actually something I'm testing now! I had to give my PC a break from training for a bit because I didn't want to burn it out.. I'm almost done building a new one that going to be 100x more powerful. Once I get my remaining parts in the mail, I'm going to start training aggressively with new settings and different batch sizes.