MIDI Embeddings

Overview

This project implements a Transformer-based model for generating embeddings from MIDI files, focusing on learning meaningful representations of musical pieces.

Main Script: main.py

Purpose

The main.py script serves as the primary entry point for training and evaluating a MIDI embedding model. When a user runs this script, the following will happen:

Workflow

1. Configuration Setup

   • A configuration dictionary is created with hyperparameters for the model and training process
   • Parameters include sequence length, embedding dimensions, attention heads, model layers, batch size, epochs, learning rate, and dropout rate

2. Dataset Preparation

   • Loads three datasets:
     • Training dataset
     • Validation dataset
     • Test dataset
   • Uses MIDIDatasetPresaved and MIDIDatasetDynamic for efficient data handling
   • Tokenizes MIDI files using a pre-trained tokenizer if available, otherwise trains a new one

3. Model Initialization

   • Creates a MIDITransformerEncoder with the specified configuration

4. Model Training

   • Trains the model using the training dataset
   • Validates performance on the validation dataset
   • Saves the best-performing model checkpoint

5. Model Evaluation

   • Evaluates the trained model on the test dataset
   • Prints out the test loss and perplexity metrics

6. Embedding Visualization

   • Generates and saves an interactive HTML visualization of embeddings for all the songs in the MAESTRO-sustain-v2 dataset using t-SNE

Example Usage

python main.py

Installation

pip install requirements.txt

Key Components

• transformer.py: Defines the MIDI Transformer Encoder architecture
• dataset.py: Handles MIDI dataset loading and preprocessing
• train.py: Contains training and evaluation functions
• visualize.py: Provides embedding visualization functions

Customization

Users can modify the configuration dictionary in main.py to experiment with different hyperparameters, such as:

• Embedding dimensions
• Number of attention heads
• Number of model layers
• Learning rate
• Batch size
• Dropout rate
• Sequence length
• Number of epochs

Visualization

The script generates an interactive HTML visualization of song embeddings, allowing users to explore how different musical pieces are represented in the embedding space.