amanteur / bandsplitrnn-pytorch Goto Github PK

Hey thank you for your amazing work! I was wondering if there is a pretrained model available for this, and if there ever will be, thank you!

also, how does processing time compare to Demus all things being equal?

Thanks your great implementation. Could you reproduce the vocal SDR performance (uSDR=10.04 dB) ? Could you provide your pretrain model ? Thanks

Hello, in the class MaskEstimationModule, should we add an extra line of code to permute the tensor before reshaping that? Otherwise the meaning of the dimensions may not match?

Installing numpy 1.23.5 needs python version 3.10, but the pytorch version only supports 3.7 and below.

Hi,
Congrats for all of these, excellent work!
Maybe you can find some spare time to update everything to Python 3.11
It's not nice to have 3.10 and 3.11 on the same machine.

python inference.py -i back.mp3 -o vocalsonly -t vocals -c .\saved_models\vocals\vocals_upd.pt gives me a huge error.

Traceback (most recent call last): File "C:\Users\Nicolas Hinds\Downloads\Compressed\BandSplitRNN-Pytorch-main\src\inference.py", line 112, in <module> main(args) File "C:\Users\Nicolas Hinds\Downloads\Compressed\BandSplitRNN-Pytorch-main\src\inference.py", line 65, in main program = InferenceProgram(**args) File "C:\Users\Nicolas Hinds\Downloads\Compressed\BandSplitRNN-Pytorch-main\src\inference.py", line 46, in __init__ self.sep = Separator(self.cfg, self.ckpt_path) File "C:\Users\Nicolas Hinds\Downloads\Compressed\BandSplitRNN-Pytorch-main\src\separator.py", line 26, in __init__ self.model = self.initialize_modules() File "C:\Users\Nicolas Hinds\Downloads\Compressed\BandSplitRNN-Pytorch-main\src\separator.py", line 62, in initialize_modules _ = model.load_state_dict(state_dict, strict=True) File "C:\Users\Nicolas Hinds\Downloads\Compressed\BandSplitRNN-Pytorch-main\venv\lib\site-packages\torch\nn\modules\module.py", line 1671, in load_state_dict raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format( RuntimeError: Error(s) in loading state_dict for BandSplitRNN: Missing key(s) in state_dict: "bandsplit.layernorms.0.weight", "bandsplit.layernorms.0.bias",
etc

renaming vocals_upd to vocals.pt and doing python inference.py -i back.mp3 -o vocalsonly -t vocals has the same effect.

Hi, I'm having some problems with my training. It was in the data augmentation. I got the following error :
File "/home3/will/BandSplitRNN-Pytorch-main/src/data/dataset.py", line 161, in mix_segments
mix_segment += segment_to_add
RuntimeError: The size of tensor a (263432) must match the size of tensor b (264600) at non-singleton dimension 1

The length of data in dataloader does not match the value from prepare_dataset. Is there a problem with SAD? Or is there something that needs to be changed when preparing the dataset?

Hi, I would like to ask how the results of this model run are so different from the original paper, is it because there is no data augmentation?

Hi, I'm having some problems with my training. I didn't change the contents of the config, but I get the following error :
TypeError: on_before_optimizer_step() missing 1 required positional argument: 'optimizer_idx'

May I ask what is the problem that causes the error in trainer.fit?

Hello.

I want to train this model with not MUSDB18 dataset.

But In prepare_dataset.py, It can only be used in Musdb18 Dataset.

So I want to know, how to train without Musdb18 dataset.