Tangent-Normal Adversarial Regularization for Semi-Supervised Learning

A TensorFlow implementation for the paper:

Tangent-Normal Adversarial Regularization for Semi-Supervised Learning

Bing Yu^*, Jingfeng Wu^*, Jinwen Ma, Zhanxing Zhu

Requirements

1. python 3.6
2. tensorflow 1.9.0
3. numpy

Usage

Generate data

python cifar10_to_numpy.py

Obtain pre-trained VAE

The VAE checkpoint can be obtained via two ways:

• train by yourself: python train_vae.py
• download a pre-trained one: without augmentation, with augmentation

Play with TNAR

All of the parameters should be easy to understand by their naming conventions.

• Training: python train_tnar.py --resume vae-checkpoint
• Test: python test_tnar.py --resume tnar-checkpoint

Hyperparameters and Performance

See the paper.

Notes

• Bibtex, Slides, Poster
• Some codes are borrowed from vat_tf.