This is an interactive variational autoencoder sampler that I've created to support the discussion of Auto-Encoding Variational Bayes at our reading group. It may also help others reading the paper. The autoencoder implementation is based on Keras.
You can play with the live version or download, build and experiment with the source code, which I recommend. The live app uses a low-power compute server for autoencoder computations. So, you may experience some lag while interacting with the grid visualizations (third column). Otherwise, slider interactions should be fluid.
The viewer enables interactive sampling from the 2- and 4-dimensional latent encoding spaces of the MNIST(first row) and Fashion MNIST(second row) datasets, obtained using the probabilistic encoding proposed in the paper. The encoding space learned by the variational autoencoder here is the parameter space (z) of a probabilistic representation of the dataset, which can be sampled from (i.e., generative). You can change the z values using the sliders and see the corresponding image samples. Sliders range between 0.05 and 0.95, and z values for sampling are obtained by applying the inverse cumulative distribution function (CDF) of the Gaussian to the slider values. You can sample a single image from the encoding space (first two columns) or a grid of images (third column), where the first two latent dimensions are used (fixed) to create image samples using z values sampled on a 4-by-4 grid and the remaining two can be dynamically changed using the sliders. For example, the image sample on the grid location (row#3, col#4) represents the sample obtained using z = GaussianInverseCDF([0.95, 0.65, Slider#0Value, Slider#1Value]). You can try to increase the grid resolution if you have the compute power.
To install, download or clone the repo and then run npm install in the
va-sampler folder. As for the backend, you'll need to run pip install -r requirements.txt in the analytics_server folder. If you don't have pip
installed, install it by following the instructions
here.
To build & run, first enter npm run analytics under va-sampler/
(alternatively, change dir to analytics_server/ and python ./server.py) to
start the Python server and then open another terminal and enter npm run start to start the web client. The last step will open a tab in your default
browser & show the running application.