🌊 MiniFlow: Flow Matching on Moon Data

This project demonstrates the core concepts of Flow Matching by learning a vector field that transforms simple Gaussian noise into the structured make_moons distribution.

🚀 Setup & Installation

Follow these steps to set up the environment and install dependencies. Note that this configuration is optimized for CUDA 12.1.

# Create and activate the environment
conda create -n miniflow python=3.11 -y
conda activate miniflow

# Install PyTorch (Compatible with CUDA 12.1)
pip install torch torchvision torchaudio --index-url [https://download.pytorch.org/whl/cu121](https://download.pytorch.org/whl/cu121)

# Install core dependencies
pip install tqdm pyyaml scikit-learn imageio matplotlib wandb

🏋️ Training

python mini_flow/miniflow.py --config_path ./mini_flow/configs/miniflow.yaml

📊 Velocity Field Evolution

The following images visualize the transformation of the learned velocity field v(x,t). The model transitions from a state of total entropy to a structured flow that defines the moon manifold.

Before Training	After Training