Reproduction of "Generative Modeling by Estimating Gradients of the Data Distribution" by Yang Song and Stefano Ermon (NeurIPS 2019) in Tensorflow 2.0.

Github | OpenReview | Paper

Created for the Reproducibility Challenge @ NeurIPS 2019.

Instructions for running the code

The main file to run is main.py. Different options are available depending on the task to perform.

--experiment train for training the model.

--experiment generate to sample through Langevin dynamics with a trained model.

--experiment inpainting for performing inpainting (different patterns of occlusion are available).

--experiment toytrain to run the toy experiment.

--experiment evaluation for computing inception and FID score.

--experiment k_nearest to samples images and finding the k pixel-wise nearest images in the dataset.

--experiment intermediate to sample images and save them at each level of noise.

--experiment celeb_a_statistics for computing inception and FID score on CelebA.

--help for additional info, or check out utils.py.