khiner / drumclassification

Drum classification model

Drum classification

A PyTorch drum classification model trained on Magenta's Expanded Groove MIDI Dataset.

Final project for Computational Data Analysis (CSE-6740).

• Paper
• Slide deck
• Blog post

Training

Download data

Download and extract the full E-GMD dataset zip into the dataset/ directory at the repository root.

This zip includes all MIDI, audio, and metadata. Your directory structure should look like this:

DrumClassification/
    dataset/
        e-gmd-v1.0.0/
            drummer1/
            drummer2/
            ...
            e-gmd-v1.0.0.csv
    ...

Prepare data for training

This step is already done, resulting in the CSV files at the root of the dataset/ directory. However, if you would like to change any aspect of the data preprocessing, such as including a different set of drum kit types, the full data preprocessing pipeline can be run as follows:

See the scripts for explanations of their roles in data preparation.

$ pip install pandas mido torchaudio  # Install packages needed for preprocessing.
$ python create_slim_metadata.py  # Outputs `dataset/e-gmd-v1.0.0-slim.csv`
$ python create_label_mapping  # Outputs `dataset/note_occurrences_slim.csv` and `label_mapping.csv`
$ python chop_dataset.py  # Outputs `dataset/chopped_raw.csv`
$ python clean_chopped_raw.py  # Outputs `dataset/chopped.csv`

Explore the dataset

The explore_dataset.ipynb notebook provides a variety of data exploration tools, such as:

• visualizing relevant data distrubutions
• previewing the "chopped" drum hit clips
• listening to random supercuts of clips for all training labels

Train

$ pip install torch torchaudio pandas numpy matplotlib tqdm
$ pip install SoundFile  # torchaudio needs a backend to load wav files.
$ python train.py

Run pretrained model

A pretrained classification model is provided at pretrained/final.zip. To evaluate this model over the test set, unzip it to pretrained/final.pth and run:

$ pip install pandas seaborn torch torchaudio SoundFile
$ python inference.py

This assumes you have already downloaded the dataset as explained above.