LC-beating

This repository contains the research code for LC-beating, a low-latency beat and downbeat activation estimation framework. It includes the training pipeline, evaluation utilities, neural network definitions, and reference inference code used for the corresponding paper.

The repository is released as source code only. Training datasets, generated checkpoints, logs, and audio examples are not included. Please prepare licensed beat/downbeat datasets separately and point the code to your local data directory.

Repository Structure

.
├── networks/          # CRNN, TCN, and Transformer variants
├── param_file/        # Dataset and training configuration
├── train.py           # Training / testing entry point
├── trainer.py         # Training loop
├── simple_test.py     # Evaluation wrapper
├── eval.py            # Beat/downbeat metrics
├── inference.py       # Reference realtime inference interface
└── network_infer.py   # Frame-wise inference wrapper

Environment

The code was developed with Python 3.7 and PyTorch. A minimal environment can be installed with:

pip install -r requirements.txt

Some legacy CRNN utilities depend on BeatNet-style log-spectrogram code. If you use those modules, set:

export BEATNET_PATH=/path/to/BeatNet

Data

The training code expects HDF5 beat/downbeat datasets. Set the dataset root with:

export LC_BEATING_DATA_ROOT=/path/to/beat_chord_hdf5

Dataset names are configured in param_file/params.py.

Training

Example command:

python train.py \
  --mode train \
  --network lc42_dt_1 \
  --mark round4_device1 \
  --eval_mode online \
  --peak_type simple

Checkpoints and TensorBoard logs are written to local output directories such as model/, model_pp/, vis_scalar/, and vis_scalar_pp/.

Evaluation

Use test mode with a checkpoint number:

python train.py \
  --mode test \
  --network lc42_dt_1 \
  --mark round4_device1 \
  --check_num 135 \
  --eval_mode online

Inference

inference.py provides a reference realtime interface. The public repository does not include pretrained checkpoints or demo audio; load your own trained checkpoint and licensed audio files.

Citation

If you use this code, please cite the corresponding LC-beating paper. Citation metadata will be added after publication.

License

This source code is released under the MIT License. See LICENSE.