Langevin Dynamics and Score Matching

How to?

Setup

pip install -r requirements.txt

Train

python3 main.py config/ring.yml train

This command produces the model (learnt scaled score function): ring.pth

Test

python3 main.py config/ring.yml test

This command produces a list of frames at each stage of sampling under ring/ and an animation ring.gif made from them.

Demo (with carefully tuned hyperparameters)

• Left: sample from the real pixel distribution
• Right: sample using Langevin dynamics based on the score function learnt via denoising score matching.

Ring

Waddles

Stewie

Reference

• A good lecture on Langevin dynamics for sampling.
• A recent paper combining Langevin dynamics with denosing score matching with a more complicated multi-level pertubation scheme.
• A paper introducing the denoising score matching.