mcmusic

PREAMBLE

This is a project by Pranav Dumpa, Siri Gunturi, Rithvik Harinath, Taran Singh, Nick Teeple, and Arianna Wilde. This is the implementation for a music classification project for the class CSC 480 at Cal Poly, instructed by Professor Rodrigo Canaan. This project is based on the following references:

1. Ferreira, L. N., Mou, L., Whitehead, J., & Lelis, L. H. S. (2022, September 1). Controlling perceived emotion in symbolic music generation with Monte Carlo Tree Search. arXiv.org. https://arxiv.org/abs/2208.05162
2. (PDF) automatic musical pattern feature extraction using convolutional neural network. (n.d.). https://www.researchgate.net/publication/44260643_Automatic_Musical_Pattern_Feature_Extraction_Using_Convolutional_Neural_Network
3. G. Tzanetakis, & P. Cook. (n.d.). Musical genre classification of Audio Signals. https://www.cs.cmu.edu/~gtzan/work/pubs/tsap02gtzan.pdf
4. Jvedarutvija. (2022, June 23). Music genre classification. Kaggle. https://www.kaggle.com/code/jvedarutvija/music-genre-classification
5. Zhang, Y., & Li, T. (2025, March 20). Music genre classification with parallel convolutional neural networks and Capuchin search algorithm. Nature News. https://www.nature.com/articles/s41598-025-90619-7
6. Colin Raffel. "Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching". PhD Thesis, 2016.

DEPENDENCIES

1. Install al dependencies found in requirements.txt by running the command "pip install -r requirements.txt" in your computer's terminal.
2. Ensure python-tk is installed. This can be done using the corresponding command: On macOS: brew install python-tk@3.10 Ubuntu/Debian: sudo apt-get install python3-tk Windows: Usually comes with Python installation. If not then use "pip install tk"

FOR THE MCTS RUN

1. CreateGenreMCTSModel.py
2. preprocess.py
3. main.py

To evaluate test data, run the script evaluate.py.

THOSE INSTRUCTIONS WILL ACTIVATE THE GUI

What This NEURAL NETWORK Does

Reads and parses genre labels from msd_tagtraum_cd1.cls Matches genre-labeled tracks with MIDI files from lmd_matched Extracts 20 musical features (e.g., pitch stats, rhythm, harmony, etc.) Trains a neural network classifier to recognize genres Predicts genre probabilities for any new MIDI input

Usage Instructions

1. Set Up Your Files

Download the Lakh MIDI Dataset (LMD-matched) from colinraffel.com/projects/lmd and extract it into a folder named lmd_matched/ in your project directory. Make sure the genre label file msd_tagtraum_cd1.cls is placed in the root directory of your project. 2. Train the Genre Classifier Open the Python notebook. Run the following blocks in order: Block 2: Loads and processes genre data Block 3: Matches MIDI files with genres and filters top genres Block 6: Initializes the classifier and trains the neural network This will extract features, train the model, and save it to disk.

3. Predict Genre for a New MIDI File Go to Block 7. Replace the example file ("Enya_-_Bard_Dance.mid") with the path to your own .mid file. Run the block to see the genre probabilities predicted by the trained model.

{ 'Pop_Rock': 0.74, 'Electronic': 0.18, 'Country': 0.08 }

Features Extracted

Each MIDI file is converted into a 20-dimensional feature vector including:

Song length Number of instruments Tempo statistics Pitch mean, std, range Rhythm patterns Chord size (harmony) Loudness stats (velocity) Melodic interval stats Percussion density Unique pitch diversity

Target Genres

By default, this model is trained on:

Pop_Rock Electronic Country You can easily modify this list in the filtered_genres variable.

Evaluation

After training, the model prints final test accuracy and shows training progress. You can further improve it by:

Increasing epochs Using more advanced architectures Expanding feature set Adding more genres

The model's test accuracy can also be evaluated by running neural_net/confusion_matrix_script.py.

Troubleshooting

No valid MIDI files found – check the midi_path is correct and matches the expected LMD format. Low accuracy? – consider balancing the dataset, checking feature quality, or increasing training time.

Saving/Loading

Saves the model as my_genre_classifier_model.h5 Saves scaler and label encoder in .pkl format for reuse Load with load_model() anytime

Credits

MIDI parsing via pretty_midi Genre labels via Tagtraum Industries Dataset: Lakh MIDI Dataset