ssl-suite

A Semi-Supervised Learning suite using PyTorch.

The implementation of SSL methods are based on https://github.com/google-research/mixmatch

Currently, the following methods are implemented:

Interpolation Consistency Training
Mean Teacher
MixMatch
Pseudo Label
Virtual Adversarial Training

Updates

2019/12/25 Update WideResNet compatible with google's implementations

Requirements

Python>=3.7
PyTorch>=1.3
torchvision>=0.4.2
homura>=2019.11 (pip install -U git+https://github.com/moskomule/homura@v2019.11)
hydra>=0.11 (pip install -U hydra-core)

For data preparation, run backends/data.py.

How to run

python {ict,mean_teacher,mixmatch,pseudo_label,vat}.py

If you want to change configurations from the default values, do something like

python METHOD.py data.name=cifar100

For configurable values, see files in config.

Benchmarks

Following Berthelot+2019, the reported accuracy values are median of accuracy of last 20 epochs.

CIFAR-10

Number of Labeled images	ICT	Mean Teacher	MixMatch	Pseudo Label	VAT
4,000	0.89	0.89	0.93	-	-

Supervised learning on 50,000/4,000 images yields accuracy of 0.94/0.82.

SVHN

python {ict,mean_teacher,mixmatch,pseudo_label,vat}.py data.name=svhn data.labeled_size=1000 data.unlabeled_size=64931 data.val_size=7326 ${MODEL_SPECIFIC_SETTINGS}

Number of Labeled images	ICT	Mean Teacher	MixMatch	Pseudo Label	VAT
1,000	0.91	0.96	0.94	-	-

Supervised learning on 50,000/4,000 images yields accuracy of 0.97/0.88.

Citation

@misc{ssl-suite,
    author = {Ryuichiro Hataya},
    title = {ssl-suite: Semi-supervised Learning suite using PyTorch},
    year = {2019},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/moskomule/ssl-suite}},
}