Skip to content

Repository with code for paper "Inhibited Softmax for Uncertainty Estimation in Neural Networks"

Notifications You must be signed in to change notification settings

MSusik/Inhibited-softmax

Repository files navigation

Update (12.03.2019):
* We updated the repo so that it supports Pytorch 1.0.
* We updated the methodology - please see the updated paper on Arxiv.
* We removed the implementation of other methods - they are still available in the older commits.
* The repository contains now the matrices for running the sentiment experiments. They are on LFS, so you need to fetch them separately.

This repository contains source code that can reproduce results from the paper "Inhibited Softmax for Uncertainty Estimation in Neural Networks".

Please note:
* requirements.txt install CPU version of Pytorch that might be too slow for you
* In order to run experiments you have to download datasets and put them in correct directories

I provide a script download.sh that downloads all the required files.

The datasets that have to be downloaded:
* NOTMNIST - place in ./notMNIST_small - in case some of the images are corrupted, remove them
* Omniglot - ./omniglot
* LFW-a - ./lfw-a
* Reuters-21578 - ./nonmovies/reuters21578
* Customer Review data - ./nonmovies/customer\ review\ data
* Review Polarity - ./nonmovies/review_polarity
* Movie Reviews (IMDB) - ./movies/aclImdb

About

Repository with code for paper "Inhibited Softmax for Uncertainty Estimation in Neural Networks"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published