NSF by 王鑫
This project is Pytorch re-implementation of NSF models. For more information on NSF models, please visit https://nii-yamagishilab.github.io/samples-nsf/
| - DATA: folder to store data
|
| - cyc-noise-nsf-4: cyclic-noise hn-sinc-NSF
|
| - hn-nsf: harmonic-plus-noise NSF
|
| - hn-sinc-nsf-9: harmonic-plus-noise NSF with trainable sinc filter
|
| - hn-sinc-nsf-10: hn-sinc-nsf-9 with the BLSTM in condition module replaced by CNNs
|
| - hn-sinc-nsf-hifigan: hn-sinc-nsf 9 + hifi-gan discriminator
cd hn-nsf
source ../../../env.sh
bash 00_demo.sh
This script will download the CMU database and pre-extracted features. It then generates samples using a pre-trained model and the pre-extracted features. It finally trains a new model on the data.
Pre-trained models are either included in __pre-trained or downloaded through 00_demo.sh.
This may take a few days or more. You may consider run 00_demo.sh in the background.
-
Interactive tutorial: An interactive tutorial on NSF is in ../../tutorials/s1_demostration_hn-nsf.ipynb Please follow the README.md there to use that tutorial
-
To accelerate training: The default script uses torch.backends.cudnn.deterministic= True and torch.backends.cudnn.benchmark = False for reproducibility. https://pytorch.org/docs/stable/notes/randomness.html
If you want to accelerate training, add options to the command line in 00_demo.sh
python main.py --num-workers 10 --cudnn-deterministic-toggle --cudnn-benchmark-toggle
This will set torch.backends.cudnn.deterministic=False and torch.backends.cudnn.benchmark = True
-
To set batch size > 1: python main.py --num-wokers 10 --batch-size N
If you have any question, contact the author