PyTorch Implementation of "Towards K-Means-Friendly Spaces: Simultaneous Deep Learning and Clustering," Bo Yang et al., ICML'2017.
I found the official implementation of deep clustering network (DCN) is outdated (https://github.com/boyangumn/DCN-New). This repo is a re-implementation of DCN using PyTorch.
An interesting work that jointly performs unsupervised dimension reduction and clustering using a neural network autoencoder.
Here I offer a demo on training DCN on the MNIST dataset (corresponding to Section 5.2.5 in the raw paper). To run this demo, simply type the following command:
python mnist.py
For anyone with interests, you can also refer to the implementation of Günther Eder: https://github.com/guenthereder/Deep-Clustering-Network, which has more details on the reproducibility.
I trained the DCN model on MNIST dataset, hyper-parameters like network structure were set as values reported in the paper. The left figure presents the reconstruction error of the autoencoder during the pre-training stage, and the right figure presents changes on NMI and ARI (two metrics employed in the paper) during the training stage. The best NMI result I have got is around 0.65.
In my practice, this implementation also works fine on PyTorch 0.4.1. Feel free to open an issue if there were incompatibility problems.