Domain-Adversarial Training of Neural Networks in Tensorflow

Requires TensorFlow>=1.0 and tested with Python 3.5

MNIST Experiments

This experiment is tested under the mnist dataset

Build MNIST-M dataset

First you should download data from http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/BSR/BSR_bsds500.tgz

and run

python create_mnistm.py

This may take a couple minutes and should result in a mnistm.pkl file containing generated images.

Training

python main.py

Statement

This experimental idea from "Domain-Adversarial Training of Neural Networks" and this is test version.