A Bestiary of Autoencoders

This repository illustrates some simple autoencoder architectures. You might want to check out the slides. To run the example, follow these steps:

1. Install poetry, a framework for simplifying the management of virtual environments.
2. Issue poetry install to set up and install all dependencies.
3. To run the model with default parameters, issue poetry run python -m bestiary.mnist.

This should result in the following output:

  | Name    | Type   | Params
-----------------------------------
0 | encoder | Linear | 1.6 K
1 | decoder | Linear | 2.4 K
-----------------------------------
3.9 K     Trainable params
0         Non-trainable params
3.9 K     Total params
0.016     Total estimated model params size (MB)
Epoch 1:  62%|█████████████████████████████████████████████████████████████████▊                                        | 534/860 [00:06<00:04, 80.91it/s, loss=0.0557, v_num=127]

Afterwards, see the folder output for some latent space reconstructions. To see available options, run poetry run python -m bestiary.mnist --help. The current options of the code are as follows:

usage: mnist.py [-h] [-m {linear,vae}] [--max-epochs MAX_EPOCHS]

optional arguments:
  -h, --help            show this help message and exit
  -m {linear,vae}, --model {linear,vae}
                        Model to use for training
  --max-epochs MAX_EPOCHS
                        Maximum number of epochs to train

If you are interested in autoencoders, check out these awesome other projects:

• latent-lego: a Python library for building autoencoders for single-cell genomics, developed by the Quantitative Developmental Biology Lab of D-BSSE.
• topological-autoencoders: a new type of architecture for preserving topological properties in your data, developed by the Machine Learning and Computational Bioloy Lab of D-BSSE.