MDSM

Code for reproducing results in Multiscale Denoising Score Matching

About MDSM

MDSM train a neural network Energy-Based Model Efficiently without sampling.

The resulting Energy function can be sampled with Annealed Langevin Dynamics.

CIFAR and CelebA samples:

Requirements

• PyTorch

• torchvision

Usage

Train EBM on Fashion MNIST:

sh exps/fmnist_train.sh

Generate samples on Fashion MNIST (modify the time string in .sh file to that of your saved experiment before running):

sh exps/fmnist_sample_single.sh

Train EBM on CIFAR (takes about 24h on 2 GPUs):

sh exps/cifar_train.sh

Generate samples for range of saved networks in a folder: (modify time argument in sh file to that of you logging folder)

sh exps/cifar_sample_all.sh 

Generate more samples from one network (modify --log argument to folder name and --time argument to time string):

sh exps/cifar_sample_single.sh

Pretrained models

Download pretrained cifar model at here, and unpack to logs folder

Then, visualize samples from pretrain model with:

sh exps/cifar_visualize_pretrain.sh

Citation

@article{li2019learning,
  title={Learning energy-based models in high-dimensional spaces with multi-scale denoising score matching},
  author={Li, Zengyi and Chen, Yubei and Sommer, Friedrich T},
  journal={arXiv preprint arXiv:1910.07762},
  year={2019}
}