FixMatch

This is an unofficial PyTorch implementation of FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. The official Tensorflow implementation is here.

This code is only available in FixMatch (RandAugment).

Results

CIFAR10

* November 2020. Retested after fixing EMA issues.

CIFAR100

* Training using the following options --amp --opt_level O2 --wdecay 0.001

Usage

Train

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

python train.py --dataset cifar10 --num-labeled 4000 --arch wideresnet --batch-size 64 --lr 0.03 --expand-labels --seed 5 --out results/[email protected]

Train the model by 10000 labeled data of CIFAR-100 dataset by using DistributedDataParallel:

python -m torch.distributed.launch --nproc_per_node 4 ./train.py --dataset cifar100 --num-labeled 10000 --arch wideresnet --batch-size 16 --lr 0.03 --wdecay 0.001 --expand-labels --seed 5 --out results/[email protected]

Monitoring training progress

tensorboard --logdir=<your out_dir>

Requirements

• python 3.6+
• torch 1.4
• torchvision 0.5
• tensorboard
• numpy
• tqdm
• apex (optional)

My other implementations

• Meta Pseudo Labels
• UDA for images

References

• Official TensorFlow implementation of FixMatch
• Unofficial PyTorch implementation of MixMatch
• Unofficial PyTorch Reimplementation of RandAugment
• PyTorch image models

Citations

@article{sohn2020fixmatch,
    title={FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence},
    author={Kihyuk Sohn and David Berthelot and Chun-Liang Li and Zizhao Zhang and Nicholas Carlini and Ekin D. Cubuk and Alex Kurakin and Han Zhang and Colin Raffel},
    journal={arXiv preprint arXiv:2001.07685},
    year={2020},
}