This is an implementation of "FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence," NeurIPS'20. in pytorch.
We reproduced the experiments on CIFAR-10 with 40, 250, 4000 labeled data and 5000 validation samples as the official implementation of FixMatch. But due to the limitation of computational resources, we didn’t reproduce 5 "folds". Thus, our result based on 1 fold doesn’t have the standard deviation. Our model uses the Wide ResNet-28-2 with leaky ReLU activation function. Our results are comparable to the performance in the original paper.
| CIFAR-10 | |||
|---|---|---|---|
| Method | 40 labels | 250labels | 4000 labels |
| Official FixMatch(RA) | 13.81±3.37 | 5.07±0.65 | 4.26±0.05 |
| Ours(RA) | 10.04 | 5.29 | 4.36 |
pip install --upgrade git+https://github.com/pytorch/ignite
pip install -r requirements.txt
python run.py DATASET.label_num=250 DATASET.strongaugment='RA'
python run_load_temp.py EXPERIMENT.resume_checkpoints='./checkpoints/'