This is the TensorFlow source code for the Mean Teacher paper. The code runs on Python 3. Install the dependencies and prepare the datasets with the following commands:

pip install tensorflow==1.2.1 numpy scipy pandas
./prepare_data.sh

Note that TensorFlow versions >= 1.3 have a bug that causes the code to hang up in the beginning. If you want to provide a workaround, pull requests are welcome.

To train the model, run:

• python train_svhn.py to train on SVHN using 500 labels
• python train_cifar10.py to train on CIFAR-10 using 4000 labels

These runners converge fairly quickly and produce a fair accuracy.

To reproduce the experiments in the paper run: python -m experiments.cifar10_final_eval or similar. They use different hyperparameters, and each of the runs takes roughly four times as long to converge as the example runners above. See the experiments directory for the complete set of experiments.