dihjiang/mtq

Multivariate Tirangular Quantiles for Novelty Detection

Multivariate Triangular Quantiles for Novelty Detection

Pytorch implementation to replicate the experiments in our NIPS2019 paper "Multivariate Triangular Quantiles for Novelty Detection" (published soon).

Datasets

• Thyroid: http://odds.cs.stonybrook.edu/thyroid-disease-dataset/
• KDDCUP: http://kdd.ics.uci.edu/databases/kddcup99/kddcup.testdata.unlabeled_10_percent.gz
• MNIST/Fashion MNIST: automatically download by torchvision

Models

This implementation includes the following models for novelty detection.

• LSA: Autoencoder part in [2]
• LSA_EN[2]: Autoencoder + Density Estimator in [2]
• LSA_SOS: Autoencoder + TQM based on SOS-flow
• LSA_MAF: Autoencoder + TQM based on MAF-flow
• EN: density estimator in [2]
• SOS: TQM based on SOS-Flow [3]
• MAF: TQM based on MAF-Flow [1]

The LSA_SOS and LSA_MAF are two instances for our TQM. In our paper, we mainly apply LSA_SOS as the novelty detection model.

Environments

This code runs on Python >=3.6. Set up environment by:

pip install -r requirements.txt

Demo

In demo.sh, we give some examples to show the how to train/test a model on one dataset

Reference

Our implementation based on the following reference/code:

• Autoencoder(https://github.com/aimagelab/novelty-detection): Pytorch implementation to replicate experiments in the CVPR19 paper "Latent Space Autoregression for Novelty Detection"

• MAF-flow(https://github.com/ikostrikov/pytorch-flows): PyTorch implementations of Masked Autoregressive Flow

• SOS-flow: PyTorch implementations of Sum-of-Square Polynomial Flow

[1] Papamakarios G, Pavlakou T, Murray I. Masked autoregressive flow for density estimation[C]//Advances in Neural Information Processing Systems. 2017: 2338-2347.

[2] Abati D, Porrello A, Calderara S, et al. Latent Space Autoregression for Novelty Detection[C]//International Conference on Computer Vision and Pattern Recognition. 2019.

[3] Jaini, P., Selby, K. A., & Yu, Y. (2019). Sum-of-Squares Polynomial Flow. arXiv preprint arXiv:1905.02325.