This is implementation of Noisy Student [paper][tensorflow] in PyTorch using smaller dataset(CIFAR10/CIFAR100) and smaller model architecture(ResNet).
- RandAugment : Substituted to AutoAugment
- Dropout : O
- Stochastic Depth : X
- gamma 0.97 (every 4.8 epochs if small model 700 epochs / every 2.4 epochs if large model 350 epochs) : O? X?
- increase test time crop size & fine tune : X
- filter images that the teacher has low confidences on : O
- balance the number of unlabeled images for each class : X
To run the code,
python main.py \
--lr=0.001 \
--dataset='CIFAR10'