FixMatch is an alternate method to perform semi-supervised learning with limited data labelling, also using Contrastive Learning.
We evaluated the performance of FixMatch versus SimCLR (with Fine-Tuning and Knowledge-Distillation) for 1%, 2% and 5% labels and found SimCLR to be faster and more accurate. Refer to the Final Report for the comparison of results.
The code and instructions to run FixMatch on our dataset can be found in FM.ipynb. It uses fm.zip, which contains the FixMatch code (obtained from Google Research) and data preparation and execution script, modified to work on our dataset.