Simple Flow Matching

A minimal demo of Conditional Flow Matching with Optimal Transport adapted from the 100 lines of code implementation here.

The example notebook walks through the training and inference workflow

The Torch-CFM library provides scalable and flexible support for training CNFs.

Reference Paper

The algorithms implemented and notation used in the notes and code follows:

Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, Matt Le. Flow Matching for Generative Modeling, 2023.

@article{lipman2023flowmatchinggenerativemodeling,
      title={Flow Matching for Generative Modeling}, 
      author={Yaron Lipman and Ricky T. Q. Chen and Heli Ben-Hamu and Maximilian Nickel and Matt Le},
      year={2023},
      eprint={2210.02747},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2210.02747}, 
}

Installation

Follow the steps given below to install simple-flow-matching in a conda virtual environment.

# Clone this repository to your machine
git clone https://github.com/anoushka2000/simple-flow-matching.git

# On artemis
module purge
module --ignore_cache load python/3.11.5 cuda/12.1.1

# Create a virtual environment
conda create --name flow_matching python=3.11

# Activate the environment
conda activate flow_matching

# Install package in the environment
cd simple-flow-matching
pip install .