Chainer IMSAT

Implementation of Learning Discrete Representations via Information Maximizing Self-Augmented Training (IMSAT) in Chainer.

Requirements

• chainer 3.0.0
• cupy 2.0.0
• munkres

How to train

• Train MNIST dataset with GPU

python train_cluster.py --epoch 50 --batchsize 250 --gpu 0 --dataset mnist

• Train 20news dataset with GPU

python train_cluster.py --epoch 50 --batchsize 250 --gpu 0 --dataset 20news

Reference

• Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto and Masashi Sugiyama. Learning Discrete Representations via Information Maximizing Self-Augmented Training. In ICML, 2017
• weihua916/imsat: Reproducing code for the paper "Learning Discrete Representations via Information Maximizing Self-Augmented Training"
• musyoku/IMSAT: Chainer implementation of Information Maximizing Self Augmented Training
• crcrpar/imsat at renewal