This is a PyTorch implementation of the paper VIME: Extending the Success of Self- and Semi-supervised Learning to Tabular Domain on MNIST DATA

I use some codes from the original repo that uses tensorflow https://github.com/jsyoon0823/VIME ; I also used some functions from the repo https://github.com/aladdinpersson/Machine-Learning-Collection

The notebook that compares supervised learning, semi-supervised learning and self-supervised learning is VIME_training.ipynb

The semi-supervised learning is not working correctly and I'm still working on debugging It. So far Here are the accuracies that I get on the test set :

In the notebook VIME_income_data.ipynb I tried self supervised learning . However, I still get a better accuracy for the supervised xgboost model. This means that either I need to work on optimizing the hyperparameters or I need to change the loss definition on categorical variables.