On the Mode-Seeking Properties of Langevin Dynamics

This repo contains the official implementation for the paper On the Mode-Seeking Properties of Langevin Dynamics by Xiwei Cheng, Kexin Fu, and Farzan Farnia.

Getting started

To train a neural network for chained Langevin dynamics on the original and flipped images from the MNIST dataset, run

python3 main.py  --runner ChainedRunner  --doc chained_mnist_flip  --config chained_mnist_flip.yml

Then the model will be trained according to the configuration files in configs/chained_mnist_flip.yml, and the log files will be stored in run/logs/chained_mnist_flip.

To generate images using chained Langevin dynamics with 30000 iterations, run

python3 main.py  --test  --test_iter 30000  --runner ChainedRunner  --doc chained_mnist_flip  --config chained_mnist_flip.yml

Then the generated samples will be saved in run/logs/chained_mnist_flip/images_iter30000.

References

The implementation is based on the paper Generative Modeling by Estimating Gradients of the Data Distribution (code).