/Tensorflow_FixMatch

[Re-implementation] FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence

Primary LanguagePython

FixMatch

# Motivation

# Summary of paper

# Commands

  • for training,
# You want to train FixMatch using single gpu.
python train.py --use_gpu 0 --number_of_labels 40 
python train.py --use_gpu 1 --number_of_labels 250 
python train.py --use_gpu 2 --number_of_labels 4000 

# You want to train FixMatch using multiple gpus.
python train.py --use_gpu 0,1,2,3 --number_of_labels 40 
python train.py --use_gpu 0,1,2,3 --number_of_labels 250 
python train.py --use_gpu 0,1,2,3 --number_of_labels 4000
  • random seed is 0.
The number of labels 40 250 4000
Official implementation (with RA) 86.19 ± 3.37 94.93 ± 0.65 95.74 ± 0.05
My implementation (with RA) 92.39 95.14 95.62

# References

  • Official Tensorflow implementation of "FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence" (google-research/fixmatch) [Code]
  • Unofficial PyTorch implementation of "FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence" [Code]
  • FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence [Paper]