Experimenting with AutoEncoders

This repo (roughly) experiments with various autoencoder architectures to tease out their ability to learn meaningful representations of data in an unsupervised setting.

• Each model is trained (unsupervised) on the 60k MNIST training set.
• Models are evaluated via transfer learning on the 10k MNIST test set.
• Each model's respective encodings of varying subsets of the 10k MNIST test set are input to train a linear classifier:
  • Each linear classifier is trained on increasing split sizes ([10, 100, 1000, 2000, 4000, 8000]) and evaluated on remaining subset of the 10K MNIST test set.
  • See autoencoders/compare.py for details.
• Classifier performance is measured by mulit-class accuracy and (ROC)AUC.

Notes

• Most models follow the architectures from their respective papers (see autoencoders/models/).
• Hyperparameter optimization is not performed on the encoder models nor the linear classifiers.
• Model specification/configuration located in the outputs/ directory.

Results

Install

Create a virtual environment with Python >= 3.10 and install from git:

pip install git+https://github.com/chris-santiago/autoencoders.git
conda env create -f environment.yml
cd autoencoders
pip install -e .

Use

Prerequisites

Hydra

This project uses Hydra for managing configuration CLI arguments. See autoencoders/conf for full configuration details.

Task

This project uses Task as a task runner. Though the underlying Python commands can be executed without it, we recommend installing Task for ease of use. Details located in Taskfile.yml.

Current commands

> task -l
task: Available tasks for this project:
* check-config:          Check Hydra configuration
* compare:               Compare encoders using linear baselines
* eval-downstream:       Evaluate encoders using linear baselines
* plot-downstream:       Plot encoder performance on downstream tasks
* train:                 Train a model
* wandb:                 Login to Weights & Biases

Example: Train a SiDAE model

The -- forwards CLI arguments to Hydra.

task train -- model=sidae2 data=simsiam callbacks=siam

Weights and Biases

This project is set up to log experiment results with Weights and Biases. It expects an API key within a .env file in the root directory:

WANDB_KEY=<my-super-secret-key>

Users can configure different logger(s) within the conf/trainer/default.yaml file.

Documentation

Documentation hosted on Github Pages: https://chris-santiago.github.io/autoencoders/