Xnet

This is a implement example of Xnet. We take the transfer between Mnist and Mnist-M as an example.

The final result would be 0.892 after 10k epochs, 0.910 after 40k epochs(Domain accuracy: 0.569), 0.963 after around 0.6m epochs(Domain accuracy: 0.547). Different times of implementations might result in trivial different accuracies. If you want to stablize the result, add seed in your forked repo.

Run Training

python train.py

See logs

tensorboard --logdir=log

data benchmark

Here we put all the Mnist and Mnist-M data in this repo. Just download the whole repo and run.

All the six data benchmarks we collected and preprocessed will be released when the paper published.

Environment

All code was tested on Ubuntu 16.04 with Python 2.7. You also need the following package: tensorflow cv2 python-box ruamel.yaml

More further optimized code will be uploaded when the paper published.