Further Analysis of Outlier Detection with Deep Generative Models

Datasets

Most datasets will be downloaded automatically. For the GAN-generated dataset, see the instructions in gan_dset/.

For the other datasets, you will need to download them manually and change some hard-coded paths. The files below can be downloaded from http://ml.cs.tsinghua.edu.cn/~ziyu/static/ood/${file_name}:

• CelebA: celeba-32.npz
• SVHN: test_32x32.mat
• TinyImageNet: imgnet_32x32.npz
• For the experiment in Appendix B: {const,random,facescrub,omniglot,trafficsign}.npz

Using the Code

To run the experiments in paper, see the instructions in vae/ and pixelcnn/ (for PixelCNN++ and the linear model). All code is tested under Python 3.6 and TensorFlow 1.

The proposed test is implemented in pixelcnn/linear_tests.py which is mostly self-contained.

Acknowledgement

This repository contains code adapted from multiple sources. See the README in each directory.