mixmatch-tensorflow2.0

This is an implementation of the research paper MixMatch: A Holistic Approach to Semi-Supervised Learning in Python and TensorFlow 2.0. All reported results from training use the WideResNet 28-2 architecture as described in the paper.

Results

CIFAR-10

Implementation/Labels	250	500	1000	2000	4000
MixMatch Paper	88.92±0.87	90.35±0.94	92.25±0.32	92.97±0.15	93.76±0.06
mixmatch-tensorflow2.0	88.60

Prerequisites

pip installs:

numpy>=1.17.2
pyyaml>=5.1.2
tensorflow>=2.0
tensorflow-datasets>=1.2.0
tqdm>=4.36.1

Training

To run a training session simply run the main.py file from the project directory

python3 main.py

To run a training session with non-default hyperparameters you have two options

Option #1, run the training session with command line arguments for the hyperparameters you wish to change:

python3 main.py --dataset "cifar10" --labelled-examples 250 --learning-rate 0.02

Option #2, run the training session with a .yaml config file:

python3 main.py --config-path "configs/cifar10@250.yaml"

Please note if you use command line arguments and a .yaml config file any overlapping arguments will use the value from the config file instead of the value provided by the command line argument

Tensorboard

To write tfevent files for tracking training progress in tensorboard simply run a training session using the tensorboard flag:

python3 main.py --tensorboard

All tfevent files are written to a .logs directory under the project directory, so to run tensorboard on the logs written during a training session run the following command from inside the project directory:

tensorboard --logdir .logs/*

Citations

@misc{berthelot2019mixmatch,
    title={MixMatch: A Holistic Approach to Semi-Supervised Learning},
    author={David Berthelot and Nicholas Carlini and Ian Goodfellow and Nicolas Papernot and Avital Oliver and Colin Raffel},
    year={2019},
    eprint={1905.02249},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}