Variational Autoencoders for Semi-Supervised Learning

Train and analyze different types of models for semi-supervised learning. Contains implementations for logistic regression, linear VAE, convolutional VAE, M1 model and Prediction-constrained VAE. Currently supports MNIST and CIFAR-10 datasets.

Built by Harvey Mudd college students mentored by Prof. Gabriel Hope! Students: Rohan Subramanian, Sahil Rane, Saheli Patel

Setup for Lab Members

If you want to run this code, clone the repository into your own directory on the teapot server.

1. Installation

Next, create a virtual environment (most probably in the same directory as your cloned repository). You can also use conda or Docker.

python3.11 -m venv /path_to_your_cloned_repo/env

Then, activate the environment using source env/bin/activate.

Install the requirements using You can install requirements using

pip install -r requirements.txt --no-cache-dir

You will probably run into to issues installing PyTorch. Visit the Pytorch website to find the distribution matching your system. For the teapot server it should be:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 --no-cache-dir

Install any other packages that failed using pip. Do not forget the --no-cache-dir flag, as it prevents a copy of the packages from being backed up in your home directory. If you do forget and get I/O errors due to no memory, run rm -rf ~/.cache/pip.

2. Training

Now you can try training some models.

You should run the scripts from the root directory of the repository. You may have to run this command first:

export PYTHONPATH="/path_to_your_cloned_repo:$PYTHONPATH"

Example: python3 scripts/train_vae.py --config scripts/configs/config-vae-mnist.json

Make your own config file to edit training. Training is logged to Tensorboard. You can launch Tensorboard in your browser using one of the following commands:

tensorboard --logdir=logs python3 -m tensorboard.main --logdir=logs

You can specify to save a model then load the saved model later.

3. Version Control

Make your own branch to code up new models or features and submit pull requests as needed.

4. Weights and Biases

We have a Weights and Biases project to log important runs: https://wandb.ai/hopelab-hmc/torchVAE/. Use wandb disabled to turn off logging when testing.