LAMIR: Beat Tracking Tutorial

source code for the 2024 LAMIR beat tracking tutorial.

in this repo, we implement a PyTorch version of the TCN proposed by Sebastian Böck and Matthew E. P. Davies in the paper "DECONSTRUCT, ANALYSE, RECONSTRUCT: HOW TO IMPROVE TEMPO, BEAT, AND DOWNBEAT ESTIMATION".

code authored by Giovana Morais

---

file overview

• config.py: hyperparameters for training and finetuning.
• dataloader.py: dataloader for audio and beat data
• model.py: torch beat tracking model
• pl_model.py: pytorch lightning model
• train.py: training script
• finetune.py: finetuning script

installing dependencies

pip install -r requirements.txt

training

alter whatever hyperparameters you with on the PARAMS_TRAIN inside config.py

python train.py

finetuning

we provide one pre-trained model for the finetuning. to use it, keep the PARAMS_FINETUNE dictionary as is.

if you trained a new model and want to use it, please make sure to update the CHECKPOINT_NAME variable in the config.py file accordingly.

PARAMS_FINETUNE = {
    "MAX_NUM_FILES": 5,
    "LEARNING_RATE": 0.005,
    "N_FILTERS": 20,
    "KERNEL_SIZE": 5,
    "DROPOUT": 0.15,
    "N_DILATIONS": 11,
    "N_EPOCHS": 5,
    "LOSS": "BCE",
    "NUM_WORKERS": 7,
    # Change here with your checkpoint filename
    "CHECKPOINT_FILE": <new_filename>
}

run the script by specifying which dataset you want to finetune it:

python finetune.py --dataset=brid
# or
python finetune.py --dataset=candombe

if you want to download the datasets, just provide the --download flag

python finetune.py --dataset=brid --download
# or
python finetune.py --dataset=candombe --download

NOTE: this will run the download with force_overwrite.