danxan/iic-1

Invariant Information Clustering for Unsupervised Image Classification and Segmentation

Invariant Information Clustering for Unsupervised Image Classification and Segmentation

This repository contains PyTorch code for the IIC paper. Most of the code is ported to Python 3. I can confirm that the script cluster_greyscale_twohead.py now works.

IIC is an unsupervised clustering objective that trains neural networks into image classifiers and segmenters without labels, with state-of-the-art semantic accuracy.

We set 9 new state-of-the-art records on unsupervised STL10 (unsupervised variant of ImageNet), CIFAR10, CIFAR20, MNIST, COCO-Stuff-3, COCO-Stuff, Potsdam-3, Potsdam, and supervised/semisupervised STL. For example:

Commands used to train the models in the paper here. There you can also find the flag to turn on prediction drawing for MNIST:

How to download all our trained models here.

How to set up the segmentation datasets here.

To use a model trained on another system, the following paths in the config files need to be changed: root_dir, out_dir, root_dir, root_dir,

I started to change to code to be robust against this, but decided it is better to change the files.