This is the source code for the ICLR 2017 workshop paper "Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results".
The code runs on Python 3. Install the dependencies and prepare the datasets with the following commands:
pip install tensorflow numpy scipy
./download_svhn.py
To train the model, run:
python train_svhn.py
to train on SVHN using 500 labelspython train_cifar10.py
to train on CIFAR-10 using 4000 labels
To reproduce the results of the poster above, run:
python train_svhn_final_eval.py
to reproduce the SVHN resultspython train_cifar10_final_eval.py
to reproduce the CIFAR-10 results