This repository shows how to make pictures like the one above.

This simple repo contains code for training a VAE on fashion MNIST data and then using the encoder from the VAE to generate features for each fashion MNIST image. These are then used as the inputs for the UMAP projection. The latter relies heavily on the excellent umap-learn python library.

Set up the environment

Clone this repository somewhere:

git clone https://github.com/dhudsmith/nnet_umap.git

With anaconda installed, navigate to the project root and execute

conda env create -f environment.yml

To start the environment, execute

conda activate nnet_umap

Use conda deactivate to exit.

Starting jupyter

Within the conda environment, execute

jupyter lab

Jupyter should open in your default browser. If not, follow the link in your terminal.

Training the VAE

You can optionally run through train_vae.ipynb jupyter notebook to train the VAE yourself using the fashion-MNIST dataset. A pretrained encoder is available in the models/ folder, so you can skip this unless you want to customize the model or use a different dataset.

Generating the graphic

Run through umap_projection.ipynb to generate the visual. Note the settings at the top these control the number of data points used, the color scheme, the resolution, and the size of data points. The code can take an hour or more run if you use lots of data points. The default parameters should run in a few minutes on most machines. If all goes well, you should get something like this with default settings: