GRN-hackathon

This repository contains implementations of two major diffusion model architectures for generative modeling on the Fashion-MNIST dataset:

1. DDPM (Denoising Diffusion Probabilistic Models)
2. Score-Based Diffusion (based on score matching)
3. Flow Matching (applied to Paul15 scRNA-seq data)

Both implementations include:

• Unified EMA (Exponential Moving Average) for stable training
• UNet based backbone
• FID score evaluation
• VAE baseline for benchmarking

Comments:

1. ddpm_diffusion.ipynb works well and generates high-quality images with fast sampling, we should use this.
2. score_based_diffusion.ipynb also works but the generated image quality is low and the EDM sampler is not working for now.
3. paul15_flow_matching.ipynb implements flow matching on the Paul15 dataset (included in data/), but on 2D embeddings.

Reference:

The ddpm_diffusion.ipynb implementation is based on https://github.com/dome272/Diffusion-Models-pytorch