AMT-APC

AMT-APC is a method for training an automatic piano cover generation model by fine-tuning an AMT (Automatic Music Transcription) model.

• Project page: AMT-APC
• Paper: [2409.14086] AMT-APC: Automatic Piano Cover by Fine-Tuning an Automatic Music Transcription Model

Usage (Piano Cover Generation)

Python version: 3.10

1. Install dependencies

pip install torch torchaudio soundfile pretty-midi tqdm

2. Download the pre-trained model

wget -P models/params/ https://github.com/misya11p/amt-apc/releases/download/beta/apc.pth

3. Run the inference code

python infer input.wav

You can also input a YouTube URL (requires yt-dlp).

python infer 'https://www.youtube.com/watch?v=...'

You can also specify a style (level1, level2, level3).

python infer input.wav --style level3

Usage (Training & Evaluation)

Python version: 3.10

1. Install dependencies

pip install -r requirements.txt

2. Download the pre-trained AMT model

wget -P models/params/ https://github.com/misya11p/amt-apc/releases/download/beta/amt.pth

3. Download the dataset

python download.py

The dataset directory is set to dataset/ by default. You can change this directory by modifying path.dataset in config.json.

4. Create the dataset

python data/sync.py
python data/transcribe.py
python data/sv/extract.py
python data/create_labels.py
python data/create_dataset.py

5. Train the model

python train --n_gpus 1

5. Evaluate the model

Calculate $\mathcal Q_{\text{max}}$.

git clone https://github.com/albincorreya/ChromaCoverId.git eval/ChromaCoverId
python eval/cover.py
python eval/distance.py

Options

Detailed configuration can be done through config.json or by using command line options, which are explained with --help. The default values are those used in the experiments in the paper.

Citation

@article{komiya2024,
  title={AMT-APC: Automatic Piano Cover by Fine-Tuning an Automatic Music Transcription Model},
  author={Komiya, Kazuma and Fukuhara, Yoshihisa},
  journal={arXiv preprint arXiv:2409.14086},
  year={2024}
}