MixMatch is a high-level algorithm for semi-supervised learning (SSL) that achieves state-of-the-art results on semi-supervised learning image classification tasks.
Its key steps include mixing in unlabeled samples with labeled (MixUp) and penalizing inconsistent prediction of unlabeled samples. In this work, we reproduce parts of the key findings of Google, and also discuss the limitations of the method.
In particular, we investigate the dependence of the results on the data augmentation used, the training length, and the unlabeled loss term of MixMatch. We are able to achieve 90.8% test accuracy on CIFAR10 when trained with 4000 labeled and 36000 unlabeled samples.
Tensorboard shows a view like this:
It also logs gradient values:
And plots the computation graph:
... and another view of the computation graph:
To run (currently set to train a WRN-28-2 on 1000 labeled images for 200 epochs), do
pip install -r requirements.txt
python src/main.pyWhen running you see text output showing the training progress. Data is also logged to tensorboard, which you can see by doing
(python -m tensorboard.main ./results) &
sensible-browser http://localhost:6006