mean-teacher

Why we included this project

If you work with partially labeled image data, this reference implementation is worth a look. The method trains a student network alongside a teacher whose weights are an exponential moving average of the student's, and adds a consistency cost between their outputs so unlabeled examples help drive learning. Both TensorFlow and PyTorch versions accompany the NIPS 2017 paper; the PyTorch code follows common idioms, so swapping in your own model and data loader is straightforward. The repo also includes the results from ImageNet and CIFAR-10 experiments with only a fraction of the labels, which is handy for sanity-checking other semi-supervised baselines. Because the approach is small and not tied to a specific architecture, it works both as a way to learn consistency-based regularization and as a base for larger experiments.