audio-ml-template

Template code for audio model, using Pytorch Lightning and wandb.

Preparaton

First, excecute following line, in terminal. It ensures pyton to recognize its packages and modules.

export PYTHONPATH="${PYTHONPATH}:/path/of/the/project/src/directory"

Prepare Data

• All the sound file data ('*.wav') should be located in one directory, each for train and test.
• CSV file with two columns, 'file_path' and 'label' should be prepared, each for train and test. 'file_path' should be a path after the train(or test) directory. For example, a soundfile 'dataset/train/data1.wav's path on CSV is 'data1.wav'. 'label' should be an integer.
• In src/ml/script/main.py, type the path of data directory and csv file, each for train and test.

Train

1. To use wandb logger, Type your wandb api-key inside config/wandb-api-key.txt.
2. Modify config/config.yaml file as you want.
3. Modify DATA_DIR variable in src/ml/util/constants.py with the path of data for training. All the data should be inside that directory.
4. Run following line in a terminal.

python scripts/main.py --train

Wandb logger

python scripts/main.py --train -n 'wandb-run-name'

Resume training

python scripts/main.py --train -c 'model checkpoint file path' --run-id 'run-id of previous wandb run --last-epoch 'last epoch number'

Can omit '--run-id' if not using wandb.
Can omit '--last-epoch' if not using learning rate scheduler.

LR_finder

python scripts/main.py --train --tune

It will automatically find the best lr and start training with it. Graph image of lr_finder is saved in lr_finder.png.

Inference

python scripts/main.py --test -c 'model checkpoint file path'

It saves submission.csv file in a root directory.

EDA

python utils/EDA.py

Note

If you encounter an error while using a Hugging Face model, in most cases, reducing the batch size resolves the issue.