MusiConGen

This is the official implementation of paper: "MusiConGen: Rhythm and chord control for Transformer-based text-to-music generation" in Proc. Int. Society for Music Information Retrieval Conf. (ISMIR), 2024.

MusiConGen is based on pretrained Musicgen with additional controls: Rhythm and Chords. The project contains inference, training code and training data (youtube list).

Arxiv Paper | Demo

Installation

MusiConGen requires Python 3.9 and PyTorch 2.0.0. You can run:

pip install -r requirements.txt

We also recommend having ffmpeg installed, either through your system or Anaconda:

sudo apt-get install ffmpeg
# Or if you are using Anaconda or Miniconda
conda install 'ffmpeg<5' -c  conda-forge

Model

The model is based on the pretrained MusicGen-melody(1.5B). For infernece, GPU with VRAM greater than 12GB is recommended. For training, GPU with VRAM greater than 24GB is recommended.

Inference

First, the model weight is at link. Move the model weight compression_state_dict.bin and state_dict.bin to directory audiocraft/ckpt/musicongen.

One can simply run inference script with the command to generate music with chord and rhythm condition:

cd audiocraft
python generate_chord_beat.py

Training

Training Data

The training data is provided as json format in 5_genre_songs_list.json. The listed suffixes are for youtube links.

Data Preprocessing

Before training, one should put audio data in audiocraft/dataset/$DIR_OF_YOUR_DATA$/full. And then run the preprocessing step by step:

cd preproc

1. demixing tracks

To remove the vocal stem from the track, we use Demucs. In main.py, change path_rootdir to your directory and ext_src to the audio extention of your dataset ('mp3' or 'wav').

cd 0_demix
python main.py

2. beat/downbeat detection and cropping

To extract beat and down beat of songs, you can use BeatNet or Madmom as the beat extrctor. For Beatnet user, change path_rootdir to your directory in main_beat_nn.py. For Madmom user, change path_rootdir to your directory in main_beat.py.

Then accroding to the extracted beat and downbeat, each song is cropped into clips in main_crop.py. path_rootdir should also be changed to your dataset directory.

The last stage is to filter out the clips with low volumn. path_rootdir should be changed to clip directory.

cd 1_beats-crop
python main_beat.py
python main_crop.py
python main_filter.py

3. chord extraction

To extract chord progression, we use BTC-ISMIR2019. The root_dir in main.py should be changed to your clips data directory.

cd 2_chord/BTC-ISMIR19
python main.py

4. tags/description labeling (optional)

For dataset crawled from website(e.g. youtube), the description of each song can be obtrained from crawled informaiton crawl_info.json(you can change the file name in 3_1_ytjsons2tags/main.py). We use the title of youtube song as description. The root_dir in main.py should be changed to your clips data directory.

cd 3_1_ytjsons2tags
python main.py

For dataset without information to describe, you can use Essentia to extract instrument and genre.

cd 3_tags/essentia
python main.py

After json files are created, run dump_jsonl.py to generate jsonl file in training directory.

Training stage

The training weight of MusiConGen is at link. Please place it into the directory MusiConGen/audiocraft/training_weights/xps/musicongen.

Before training, you should set your username in environment variable

export env USER=$YOUR_USER_NAME

If using single gpu to finetune, you can use the following command:

dora run solver=musicgen/single_finetune \
    conditioner=chord2music_inattn.yaml \
    continue_from=//sig/musicongen \ 
    compression_model_checkpoint=//pretrained/facebook/encodec_32khz \
    model/lm/model_scale=medium dset=audio/example \
    transformer_lm.n_q=4 transformer_lm.card=2048

the continue_from argument can be also provided with your absolute path of your checkpoint.

If you are using multiple(4) gpus to finetune, you can use the following command:

dora run -d solver=musicgen/multigpu_finetune \
    conditioner=chord2music_inattn.yaml \
    continue_from=//sig/musicongen \ 
    compression_model_checkpoint=//pretrained/facebook/encodec_32khz \
    model/lm/model_scale=medium dset=audio/example \
    transformer_lm.n_q=4 transformer_lm.card=2048

export weight

use export_weight.py with your training signature sig to export your weight to output_dir.

License

The license of code and model weights follows the LICENSE file, LICENSE of MusicGen in LICENSE file and LICENSE_weights file.

Citation

@misc{lan2024musicongenrhythmchordcontrol,
      title={MusiConGen: Rhythm and Chord Control for Transformer-Based Text-to-Music Generation}, 
      author={Yun-Han Lan and Wen-Yi Hsiao and Hao-Chung Cheng and Yi-Hsuan Yang},
      year={2024},
      eprint={2407.15060},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
}