Automated Music Composition Using Deep Reinforcement Learning

This repository is an implementation of automated music composition with Reinforcement Learning(RL) in PyTorch. We design a novel deep reinforcement learning-based framework with the self-attention-based neural networks for long-sequence music generation, which differs from the conventional Generative Adversarial Network (GAN). Additionally, we utilize Deep Q Network (DQN)-based and Proximal Policy Optimization (PPO)-based techniquein Reinforcement Learning (RL) with the transformer model to generate pop-style piano music sequentially.

Model architecture

DQN-Baed Architecture

PPO-Baed Architecture

Getting started

Install Dependencies

pip install -r requirements.txt

Dataset

We use the "AIlabs-Pop1K7" as training dataset, which can be download here.

Pre-trained Weights

Linear Transformer is our pre-trained weights of agent model, which can be downloaded here.

Run the Codes

Our expertiment is based on the unconditional generation with Reinforcement Learning using PyTorch framework in Python.

• Pretrain:

python agent_pretrain.py

• Training

python IRL_dqn_train.py

• Evaluation

## Generate midis songs in './gen_midis ##

python testing-no-type-cp.py

Demo

Our generated midi files can be downloaded here.