VAE Experiment

Working Demo

Steps

1. Train the model

Open and Run /model/variational-auto-encoder-mnist.ipbn

You'll see awesome data visualizations here like this Manifold:

And this latent space distribution:

2. Conver the model from .h5 to tensorflow.js

tensorflowjs_converter --input_format keras /Users/pedro/Documents/rants/basic-ae/vae/model/decoder_mlp_mnist.h5 /Users/pedro/Documents/rants/basic-ae/vae/model

note: replace /Users/pedro/Documents/rants/basic-ae/vae/ with the location on this file on your machine

3. Run the client

cd client # from the root folder
npm i
npm start

Resources that helped/motivated me

Building Autoencoders in Keras - Tutorial

Tutorial initially easy to follow

Intuitively Understanding Variational Autoencoders - Post/ Tutorial

Blog Post/Tutorial that complemented the previous Tutorial

Computer Generates Human Faces - Video + 'Raw' Code

The video that got me interested in Latent Spaces.

Video + Undocumented code

Variational Autoencoders - Video

Easy to understand explanation of Autoencoders, including bleeding edge ones