• TODO
  • Randomize magnitude of randaug (someday in future...)
  • DDP & interleave for batch norm in multiple gpus? (someday in future...)

Unofficial Pytorch implementation of FixMatch (NIPS 2020)

• STL-10 dataset
  • Note that the original paper uses 1000 labels for training, here I used all possible labels
• WRN-28-2 (copied model code from xternalz/WideResNet-pytorch)

Links to official implementation

• Official GitHub (Tensorflow)
• Paper link

Environments

• Use docker image: pytorch/pytorch:1.12.1-cuda11.3-cudnn8-devel
• Install packages: pip install tensorboard termcolor yacs

Usage

• python train.py --input_size 96 --log_name 20230131 --randaug --num_epochs 300 --t_max 300

Trained weight

• RandomAugment
  • Best val acc model was not saved for some early experiments
  • Refer to papaerswithcode, 92.02 is the official score.

RA magnitude	RA number	epochs	Initial LR	val acc
1	2	200	0.005	91.05 (best was 91.275)
1	2	300	0.005	92.95 (best was 93.075)

• CTAugment
  • Note: Augmentation's behavior of the original paper and official implement seems different
    • Additional 'blur' operation
    • Blending after smoothing
    • ...
  • In this version, I followed the paper.
  • Refer to papaerswithcode, 94.83 is the official score.

epochs	Initial LR	val acc
300	0.05	91.50