🎹 High-Resolution Sustain Pedal Depth Estimation

This repository contains code for training a Transformer-based model to detect piano sustain pedal usage from audio.

📂 Project Structure

.
├── train_basic.py                # Main training script
├── inference_basic.py            # Main inference & metrics calculation script
├── calculate_metric.py           # Helper functions related to evaluation metrics
├── requirements.txt              # Python dependencies for the project
├── sample_data/                  # JSON file lists for training, validation, and test data
├── src/
│   ├── model_basic.py            # Transformer-based model for pedal detection
│   ├── dataset_basic.py          # PyTorch dataset class for pedal data
│   ├── trainer_basic.py          # Trainer with MSE loss for pedal depth estimation
│   ├── trainer_bce.py            # Trainer with BCE loss for pedal depth estimation
│   ├── utils.py                  # Utility functions
│   ├── cnn_block.py              # Convolutional blocks
│   ├── transformer.py            # Transformer
│   ├── dirs.py                   # Directory and path utilities

🚀 Getting Started

1. Install dependencies

pip install -r requirements.txt

2. Prepare data

Place your .h5 data files in a directory (e.g., /path/to/data/) and update paths in sample_data/train.json, sample_data/val.json and sample_data/test.json.

3. Run training

python train_basic.py \
  --data_dir /path/to/data \
  --datasets r0-pf1 \
  --save_dir results \
  --loss_function mse

Some useful training options:

Argument	Description	Default
--checkpoint_path	Resume training from checkpoint	None
--data_dir	Path to H5 feature data	/path/to/data/
--datasets	Dataset names to include	["r0-pf1"]
--save_dir	Output directory for logs and checkpoints	results
--loss_function	Loss type (mse or bce)	mse
--batch_size	Batch size	24
--train_rand_sample	Use random sampling	False

📚 Paper

@inproceedings{FZ25pedal,
  title={High-Resolution Sustain Pedal Depth Estimation from Piano Audio across Room Acoustics},
  author={Kun Fang and Hanwen Zhang and Ziyu Wang and Ichiro Fujinaga},
  booktitle={Proceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR)},
  year={2025},
  address={Daejeon, Korea},
  month={September},
  day={21--25}
}

📦 Dataset

The dataset associated with this paper is available on Zenodo.