This is a reproducing code for IMSAT [1]. IMSAT is a method for discrete representation learning using deep neural networks. It can be applied to clustering and hash learning to achieve the state-of-the-art results. This is the work performed while Weihua Hu was interning at Preferred Networks.

You must have the following already installed on your system.

• Python 2.7
• Chainer 1.21.0, sklearn, munkres

For reproducing the experiments on MNIST datasets in [1], run the following codes.

• Clustering with MNIST: python imsat_cluster.py
• Hash learning with MNIST: python imsat_hash.py

calculate_distance.py can be used to calculate the perturbation range for Virtual Adversarial Training [2]. For MNIST dataset, we have already calculated the range.

All the datasets used in the paper can be downloaded here.

[1] Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto and Masashi Sugiyama. Learning Discrete Representations via Information Maximizing Self-Augmented Training. In ICML, 2017. Available at https://arxiv.org/abs/1702.08720

[2] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, Ken Nakae, and Shin Ishii. Distributional smoothing with virtual adversarial training. In ICLR, 2016.