Click "Run All" in plot.ipynb
- The performance provided in paper are the average of 5 pairs of the prime and odd models. You can get them by the train_mnist.py. We merged the i-th prime model with ((i+1)%5)-th odd model (because we found the i-th prime models i-th odd models are similar on weights due to the same random seed and training data).