In progress pytorch implementation of FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence
- SGD optimizer + EMA
- Cosine Annealing with warm restarts, with the schedule stepping suggested by the paper.
- Batchsize of 64
python train.py --batch-size=64 --mu=7 --dataset-name=CIFAR10C --data-dir=path/to/your/data
CIFAR10 classification accuracy for now, for mu=7
Epochs | 100 | 500 |
---|---|---|
Paper | xx | xx |
This repo | 85.0 | 93.7500 |