MixMatch

This is an unofficial PyTorch implementation of MixMatch: A Holistic Approach to Semi-Supervised Learning. The official Tensorflow implementation is here.

Experiments on just CIFAR-10 are available.

Requirements

• python 3.10.11
• pytorch 2.0.0
• pytorch-cuda 11.7
• torchvision 0.15.0
• tensorboardX
• progress
• matplotlib
• numpy

Usage

Train

Train the model by 250 labeled data of CIFAR-10 dataset:

python train.py --gpu <gpu_id> --n-labeled 250 --out cifar10@250

Train the model by 4000 labeled data of CIFAR-10 dataset:

python train.py --gpu <gpu_id> --n-labeled 4000 --out cifar10@4000

Running using nohup [It may run overnight]

nohup python train.py --gpu <gpu_id> --n-labeled 1000 --out cifar10@1000 1>[email protected] 2>[email protected] &

To check if it is running.

ps -aux | grep python | grep <username>

Results (Accuracy)

#Labels	80	90	100	250	500	1000	2000	4000
Paper	-	-	-	88.92 ± 0.87	90.35 ± 0.94	92.25 ± 0.32	92.97 ± 0.15	93.76 ± 0.06
This code (mean)	76.166	80.245	81.9485	88.2325	89.299	91.113	92.545	93.562
This code (best)	80.6	82.12	83.3	89.28	90.3	91.96	93.24	94.2

References

@article{berthelot2019mixmatch,
  title={MixMatch: A Holistic Approach to Semi-Supervised Learning},
  author={Berthelot, David and Carlini, Nicholas and Goodfellow, Ian and Papernot, Nicolas and Oliver, Avital and Raffel, Colin},
  journal={arXiv preprint arXiv:1905.02249},
  year={2019}
}