Closed-Form Flow Matching

This repository mirrors the full project folder for the Closed-Form Flow Matching course project.

The main implementation and report are in 13_Beaudouin_Collin/. The remaining top-level folders contain supporting course material, presentation files, and experiment notebooks/results kept with the original project workspace.

Main Project

13_Beaudouin_Collin/ contains:

• PyTorch code for closed-form, conditional, and empirical flow matching.
• Toy and image dataset loaders.
• MLP and compact UNet velocity models.
• Evaluation utilities and ODE samplers.
• Jupyter notebooks reproducing the main experiments.
• LaTeX report sources, generated figures, and PDFs.

Repository Layout

13_Beaudouin_Collin/   Main code, notebooks, report, figures, and final PDF
class/                 Flow matching course/reference material
diapo/                 Presentation source, exported PDF, and figures
result/                Additional experiment notebooks/results
*.zip                  Original export/submission archives

Quick Start

cd 13_Beaudouin_Collin
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python -m compileall -q src

The experiments are designed to run on a free Google Colab GPU. Image datasets are downloaded through torchvision when the notebooks or loaders are run.

Closed