You can deploy this model and start earning money today!

jnmusic/
sqgeneration

Views

No views yet

SQGen: Mamba2-Based String Quartet Generation Model

Python 3.10+

PyTorch 2.0+

License: MIT

SQGen is a symbolic music generation model based on the Mamba2 architecture, designed to generate high-quality string quartet scores (ABC notation).

🎵 Core Features

Hierarchical Mamba2 architecture: 26-layer model (20 patch-level + 6 character-level), 307 million parameters
Three-stage training pipeline: pretraining, finetuning, and reinforcement learning optimization

🏗️ Model Architecture

SQGen adopts a hierarchical architecture with two main components:

Patch-level Decoder (20 layers)

Processes music at the patch level (16 characters per patch)
Efficiently handles long-sequence dependencies
Uses Mamba2 blocks and selective state space modeling

Character-level Decoder (6 layers)

Generates ABC notation characters within each patch
Provides fine-grained control over musical structure
Uses teacher-forced autoregressive generation during training
Uses Mamba2 blocks and selective state space modeling

Model specs:

Hidden size: 1280
Patch size: 16 characters
Patch length: 2048 (pretraining) / 1024 (finetuning)
Total parameters: about 307 million
Context length: up to 32,768 characters

📦 Environment Setup

System Requirements

Python 3.10 or higher
CUDA 11.8 or higher (for GPU training)
Training requires 40GB+ GPU memory, inference requires 24GB+

Step 1: Create the environment

bash

conda create -n sqgen python=3.10
conda activate sqgen

Step 2: Install PyTorch

bash

# CUDA 11.8
conda install pytorch==2.3.0 pytorch-cuda=11.8 -c pytorch -c nvidia

# CUDA 12.1
conda install pytorch==2.3.0 pytorch-cuda=12.1 -c pytorch -c nvidia

Step 3: Install dependencies

pip install -r requirements.txt

Step 4: Install Mamba2 components

bash

# Install Mamba SSM
cd mamba/
pip install -e .

# Install Causal Conv1D
cd ../causal-conv1d/
pip install -e .

cd ..

Step 5: Verify installation

bash

cd generation/
python test_env.py

📂 Project Structure

sqgen/
├── core/                          # Core model implementation
│   ├── model.py                   # SQGen model based on Mamba2
│   ├── config.py                  # Model configuration class
│   └── utils.py                   # Patchilizer and utility functions
│
├── training/                      # Training scripts
│   ├── pretrain/                  # Pretraining stage
│   │   ├── train.py
│   │   └── config.py
│   ├── finetune/                  # Finetuning stage
│   │   ├── train.py
│   │   └── config.py
│   └── optimization/              # Reinforcement learning optimization stage
│       ├── train.py
│       ├── config.py
│       ├── data.py
│       └── utils.py
│
├── generation/                    # Music generation
│   ├── generate.py                # Main generation script
│   ├── config.py                  # Generation configuration
│   ├── utils.py                   # Generation utilities
│   ├── test_env.py                # Environment test
│   └── test_gpu.py                # GPU inference test
│
├── evaluation/                    # Evaluation tools
│   └── clamp2/                    # CLaMP-2 reward model
│       ├── extract_clamp2.py
│       ├── statistics.py
│       ├── config.py
│       └── utils.py
│
├── data/                          # Data processing
│   ├── 1_batch_xml2abc.py         # MusicXML to ABC
│   ├── 2_data_preprocess.py       # Data preprocessing
│   ├── 3_batch_abc2xml.py         # ABC to MusicXML
│   ├── abc2xml.py                 # ABC parser
│   └── xml2abc.py                 # XML parser
│
├── scripts/                       # Utility scripts
│   ├── create_data_index.py       # Create data index
│   ├── create_string_quartet_index.py
│   └── organize_by_key.py
│
├── mamba/                         # Mamba SSM library
├── causal-conv1d/                 # Causal convolution library
├── requirements.txt               # Python dependencies
└── README.md

🎼 Data Format

SQGen uses ABC notation, a text-based music notation system. Example:

abc

X:1
T:String Quartet - Classical
C:Mozart
M:4/4
L:1/8
Q:1/4=120
K:C
V:1 violin1
V:2 violin2
V:3 viola
V:4 cello
[V:1]|: G2AB c2d2 | e2dc B2A2 | G2FE D2E2 | F2ED C4 :|
[V:2]|: E2FG A2B2 | c2BA G2F2 | E2DC B,2C2 | D2CB, A,4 :|
[V:3]|: C2DE F2G2 | A2GF E2D2 | C2BA, G,2A,2 | B,2A,G, F,4 :|
[V:4]|: C,2D,E, F,2G,2 | A,2G,F, E,2D,2 | C,2B,,A,, G,,2A,,2 | B,,2A,,G,, F,,4 :|
......

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

Mamba - state space model
abc2xml/xml2abc - ABC notation tools
ABCToolkit - ABC notation processing
CLaMP 2 - contrastive language-music pretraining

jnmusic/sqgeneration

SQGen: Mamba2-Based String Quartet Generation Model

🎵 Core Features

🏗️ Model Architecture

Patch-level Decoder (20 layers)

Character-level Decoder (6 layers)

📦 Environment Setup

System Requirements

Step 1: Create the environment

Step 2: Install PyTorch

Step 3: Install dependencies

Step 4: Install Mamba2 components

Step 5: Verify installation

📂 Project Structure

🎼 Data Format

License

Acknowledgments

jnmusic/
sqgeneration