FixMatch

Unofficial Pytorch Implementation "FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence"

Usage

Requirements

• pytorch (v2.0.1)
• torchvision (v0.15.2)
• PIL (v9.4.0)
• wandb (Optional)

Arguments

• Augmentation Policy (--augs)
  • 0: no augmentation
  • 1: weak augmentation
  • 2: strong augmentation (based on RandAug)
• Check config.py file for details. (Default parameters are set for cifar10)

Example Scripts

# Model Training
$ python main.py --mode 'train' --data 'cifar10' --num_X 4000 --augs 1 2  --nesterov --amp --include_x_in_u --save_path ./model-store/001
>>>...
>>>Tue Aug  8 00:17:33 2023: Iteration: [1046528/1048576], Ls: 0.1149, Lu: 0.1146, Mask: 0.9892, Accuracy(train/test): [1.0000/0.9554]
>>>Tue Aug  8 00:31:01 2023: Iteration: [1047552/1048576], Ls: 0.1145, Lu: 0.1142, Mask: 0.9897, Accuracy(train/test): [0.9999/0.9556]
>>>Tue Aug  8 00:44:49 2023: Iteration: [1048576/1048576], Ls: 0.1153, Lu: 0.1149, Mask: 0.9897, Accuracy(train/test): [0.9999/0.9556]

# Model Evaluation
$ python main.py --mode 'eval' --load_path ./model-store/001/ckpt.pth
>>>...
>>>Model Performance: 0.9556

Results

CIFAR10

Num Labaled Data	Top 1 Acc
4000	0.9556
250	0.9473
40	0.9352

Model weights (and training logs) of the above performance are on the release page.

References

• https://arxiv.org/abs/2001.07685
• https://www.zijianhu.com/post/pytorch/ema/
• https://github.com/kekmodel/FixMatch-pytorch