🎵 Music Generation with AI — Task 3

An end-to-end deep learning pipeline that composes original piano music using LSTM neural networks trained on classical MIDI data.

Author: Roshni Sukhnani | @weiwei-gitch
Internship: DecodeLabs — Task 3

---

🧠 What It Does

This project builds an AI model that learns musical patterns from classical MIDI files and generates brand new compositions — no human input required after training.

Classical MIDI Files → Parse Notes → Train LSTM → Generate Music 🎵

---

🛠️ Tech Stack

Tool	Purpose
Python 3.14	Core language
music21	MIDI parsing and generation
PyTorch	LSTM neural network
NumPy	Data processing
tqdm	Training progress bars

---

📁 Project Structure

music_ai/
│
├── data/
│   ├── midi_files/        ← Input MIDI training data (14 files)
│   ├── notes.pkl          ← Preprocessed note sequences
│   └── model.pth          ← Trained model weights
│
├── output/
│   └── generated_music.mid ← AI-composed music output
│
├── preprocess.py          ← Step 2: Parse MIDI → note sequences
├── model.py               ← Step 3: LSTM architecture
├── train.py               ← Step 4: Model training
└── generate.py            ← Step 5: Music generation

---

🚀 How to Run

1. Install Dependencies

pip install music21 numpy tqdm
pip install torch --index-url https://download.pytorch.org/whl/cpu

2. Add MIDI Files

Place .mid files inside data/midi_files/
(Classical piano recommended — bitmidi.com)

3. Preprocess

python preprocess.py

4. Train the Model

python train.py

5. Generate Music

python generate.py

Open output/generated_music.mid with VLC or Windows Media Player 🎵

---

📊 Results

Metric	Value
MIDI files used	14
Total notes extracted	11,254
Unique notes/chords	480
Model parameters	4,391,648
Training epochs	20
Final training loss	4.08
Notes generated	200

---

🏗️ Model Architecture

Input (480) → LSTM(512) → LSTM(512) → Dense(256) → ReLU → Output(480)

• 2-layer LSTM with hidden size 512
• Dropout (0.3) to prevent overfitting
• CrossEntropyLoss + Adam optimizer (lr=0.001)
• Temperature sampling (0.8) for creative note selection

---

🎼 Pipeline Explained

Step 1 — Data Collection

Downloaded 14 classical piano MIDI files from bitmidi.com covering composers like Beethoven, Mozart, and Chopin.

Step 2 — Preprocessing

Used music21 to parse each MIDI file into individual note and chord names. Extracted 11,254 total notes with a vocabulary of 480 unique notes/chords.

Step 3 — Model Design

Built a 2-layer LSTM network with 4.3M parameters using PyTorch. The model takes sequences of 50 notes and predicts the next note.

Step 4 — Training

Trained for 20 epochs with batch size 64. Loss decreased from 4.79 → 4.08, showing the model successfully learned musical patterns.

Step 5 — Generation

Fed a random 50-note seed sequence to the trained model and generated 200 new notes using temperature-based sampling for musical variety.

---

📌 Notes

• Training runs on CPU — no GPU required
• Generation takes under 5 seconds
• Output is a standard .mid file playable on any media player

---

Built as part of DecodeLabs Internship — Task 3