fashion-mnist-vae

TensorFlow.js: Training a variational autoencoder.

This example shows you how to train a variational autoenconder using TensorFlow.js on Node.

The model will be trained on the Fashion MNIST dataset.

This example is a port of the code for a multilayer perceptron based variational autoencoder from this link https://github.com/keras-team/keras-io/blob/master/examples/generative/vae.py See this tutorial for a description of how autoencoders work.

Prepare the node environment:

yarn
# Or
npm install

In the instructions below you can replace yarn with npm run if you do not have yarn installed.

Download the data

You can run yarn download-data or follow the instructions below t download it manually.

Download the train-images-idx3-ubyte.gz from this page.

Uncompress the file and put the resulting train-images-idx3-ubyte. Into a folder called dataset within this example folder.

Run the training script:

yarn train

or if you have CUDA installed you can use

yarn train --gpu

It will display a preview image after every epoch and will save the model at the end of training. At the end of each epoch the preview image should look more and more like an item of clothing. The way the loss function is written the loss at the end of a good training run will be in the 40-50 range (as opposed to more typical case of being close to zero).

Monitoring model training with TensorBoard

By using the --logDir flag of the yarn train command, you can log the batch-by-batch loss values to a log directory, so that it can be monitored using TensorBoard.

For example:

yarn train --logDir /tmp/vae_logs

Start TensorBoard in a separate terminal:

pip install tensorboard  # Unless you've already installed tensorboard.


tensorboard --logdir /tmp/vae_logs

Once started, the tensorboard process will print an http:// URL to the console. You can open it in the browser and see the loss curve, e.g, see the example below.

Serve the model and view the results

Once the training is complete run

yarn watch

to serve the model and the web page that goes with it.