This was my selected paper from NIPS 2020 for my Statistical Learning project I found a missing function in the official code presented by the author of the paper (The planar flow function was missing, which I fixed in this forked repo)
The folder Report is consisted of a report of my understanding of the paper with some clarifications on the terminologies used in the paper The folder Code is my edited version of the official notebooks (I had to tweak some hyperparameters in order to run it in a free gpu server by google colab, because the hyperparameters used in the official notebooks required expensive, powerful GPU's , in which the original model was trained on, which I lacked.)
This repository presents the experiments of the paper:
Posterior Network: Uncertainty Estimation without OOD Samples via Density-Based Pseudo-Counts
Bertrand Charpentier, Daniel Zügner, Stephan Günnemann
Conference on Neural Information Processing Systems (NeurIPS), 2020.
To install requirements:
conda env create -f environment.yaml
conda activate posterior-network
conda env list
To train the model(s) in the paper, run one jupyter notebook in the folder notebooks. All parameter are described.
To dowload the datasets, follow the following links:
- 2DGaussians vs anomalous2D
- Segment (No sky) vs Segment (Sky only)
- SensorlessDrive (No 10, 11) vs SensorlessDrive (10, 11 only)
- MNIST vs FashionMNIST / KMNIST
- CIFAR10 vs SVHN
You can find pre-trained models in the folder saved_models. Models in saved_models/MNIST-postnet and saved_models/CIFAR10-postnet are trained on classic MNIST and CIFAR10 splits.
Please cite our paper if you use the model or this code in your own work:
@incollection{postnet,
title = {Posterior Network: Uncertainty Estimation without OOD Samples via Density-Based Pseudo-Counts},
author = {Charpentier, Bertrand, Daniel Z\"{u}gner and G\"{u}nnemann, Stephan},
booktitle = {Advances in Neural Information Processing Systems 33},
year = {2020},
publisher = {Curran Associates, Inc.},
}