- Notifications
You must be signed in to change notification settings - Fork4
chenjianyi/fastsag
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
This is a PyTorch/GPU implementation of the IJCAI 2024 paperFastSAG: Towards Fast Non-Autoregressive Singing Accompaniment Generation.Demo page can be found atdemo.
@article{chen2024fastsag, title={FastSAG: Towards Fast Non-Autoregressive Singing Accompaniment Generation}, author={Chen, Jianyi and Xue, Wei and Tan, Xu and Ye, Zhen and Liu, Qifeng and Guo, Yike}, journal={arXiv preprint arXiv:2405.07682}, year={2024}}- Download this code:
git clone https://github.com/chenjianyi/fastsag/cd fastsagDownload fastsag checkpoint fromhereand put all weights in fastsag/weights
BigvGAN checkpoints could be downloaded fromBigvGAN. The checkpoints we used is "bigvgan_24khz_100band". I upgrade BigvGAN to BigvGAN-v2, and the checkpoints would be downloaded automatically.
MERT pretrained checkpoints would be downloaded automatically from huggingface. Please make sure your sever could access huggingface.
- Source seperation:
cd preprocessingpython3 demucs_processing.py # you may need to change root_dir and out_dir in this file- cliping to 10s and filtering salient clips
python3 clip_to_10s.py # change src_root and des_root for your datasetcd ../sde_diffusionpython3 train.py --data_dir YOUR_TRAIN_DATA --data_dir_testset YOUR_TEST_DATA --results_folder RESULTSpython3 generate.py --ckpt TRAINED_MODEL --data_dir DATA_DIR --result_dir OUTPUTAbout
FastSAG: Towards Fast Non-Autoregressive Singing Accompaniment Generation
Topics
Resources
Stars
Watchers
Forks
Packages0
No packages published