Flow Matching

Inspired by Ezra Erives and Jiarui Hai

This repository provides the codebase and environment setup for the Flow Matching experiment, designed to deliver a clean, minimal, and extensible framework.

Quick Start

1. Clone the repository

# clone the repository
git clone https://github.com/hanbinzheng/flow-matching.git
cd flow-matching

---

2. Set up the environment

If you are using WSL, install the appropriate python3-venv package before proceeding. To detect and install the correct package version, run:

# install python3-venv package for wsl users
ver=$(python3 --version 2>&1 | awk '{print $2}' | cut -d. -f1,2)
sudo apt install -y python${ver}-venv

Then, execute the environment setup script:

# set up the environment
chmod +x setup_env.sh
./setup_env.sh

This process may take several minutes depending on your internet connection. Please be patient.

---

3. Activate the virtual environment and run training

# run the training code
source .venv/bin/activate
python train.py

Training duration depends on your hardware configuration, but typically completes within an hour.

Training outputs will be saved under the flow-matching/results/ directory.

---

Notes

• The setup_env.sh script automates the creation of a Python virtual environment, upgrades pip, installs required packages (excluding PyTorch), detects your system’s CUDA version, and installs the appropriate PyTorch build accordingly.

• This setup supports WSL, macOS, Ubuntu, and the vast majority of Unix-like operating systems.

• CPU-only training is fully supported, with automatic adaptation for environments lacking compatible GPUs.

• The .venv directory is excluded from version control to maintain a clean repository.

• (Only for GPU Training) Multi-GPU configurations are not supported by the automated setup and require manual intervention.

• (Only for GPU Training) Ensure NVIDIA drivers and CUDA are properly installed, and that the nvidia-smi command is available for CUDA-based GPU detection.

• For troubleshooting installation issues, consult the packages_install_log.txt file generated during setup (only for those who failed the 2: Set up the environment).

---

This setup guarantees a reproducible and efficient environment for your Flow Matching experiments.