music-genre-classification-and-recommendation

Slides: https://www.canva.com/design/DAGlAnd3rKA/ex9TRD9362weA46tJfHbog/edit

This project builds an end-to-end system for automatically classifying music by genre and recommending similar songs. Our approach uses both traditional machine learning models and deep learning architectures to tackle the problem.

Project Overview

• Problem Statement:
  Create a system that automatically classifies music by genre and then recommends similar tracks based on learned features.

• Dataset:
  We use the GTZAN Genre Collection because it is balanced (100 tracks per genre across 10 genres) and small enough for a class project. (The alternative, Free Music Archive, is too large for our purposes.)

• Components:

  1. Data Preprocessing & Exploration:
     • Loading audio files from the GTZAN dataset.
     • Extracting MFCC features and visualizing genre distribution.
  2. Baseline Models:
     • Implementing two baseline models (Random Forest and SVM) using scikit-learn.
  3. Deep Learning Architectures:
     • One deep learning model (a simple CNN built with PyTorch and PyTorch Lightning) is implemented here.
     • One deep learning model (a modified EfficientNet-B7 that accepts single-channel mel-spectrogram inputs) is implemented here.
  4. Validation & Evaluation:
     • Splitting the dataset into training and validation sets.
     • Reporting accuracy, loss, and confusion matrices.
  5. Hyperparameter Tuning & Final CNN
     • Optuna‑driven search over learning rate, dropout, filter count, and batch size.
     • Full‑dataset retrain for 20 epochs with EarlyStopping and a TQDMProgressBar for live progress.
     • Final test‑set performance: 62 % accuracy | Macro‑F1 ≈ 0.59.
     • Saves best checkpoint at notebooks/checkpoints/best_cnn.ckpt.
  6. End‑to‑End Inference & Recommendation Demo
     • Preprocessing: preprocess_audio() loads a raw .au file and converts it to a fixed‑size (1×128×128) Mel‑spectrogram tensor.
     • Embedding extraction: monkey‑patches model.get_embedding() to run the CNN’s conv‑blocks + global‑avg‑pool, returning a 1D feature vector.
     • Recommendation: predict_and_recommend(sample_clip, k=5) classifies the clip (via the loaded checkpoint) and computes cosine similarity against an embedding bank to pick the top‑5 closest tracks.
     • Demo output: prints the input filename, predicted genre, and displays a pandas DataFrame with columns file, genre, and cosine_dist for the recommended songs.

How to run

1. Install dependencies

   pip install -r requirements.txt