TensorFlow.js: Generative Adversarial Network

Demonstrates the generator part of the ACGAN in the browser.

Description

This is the web demo part of the dual-environment TensorFlow.js example of Auxiliary Classifier Generative Adversarial Network (ACGAN).

The training code is in gan.js, which runs in Node.js using tfjs-node or tfjs-node-gpu. In this web page, we load the generator part of a pre-trained GAN to generate MNIST images. How real those generated MNIST images look depends on how well the model has been trained. After 5 epochs of training, you should start seeing reasonable images. After 15 epochs, the images should start to look good. Close-to-perfect images (ones that are hard to distinguish from the real ones) should appear toward the end of the default 100 epochs of training.

The ACGAN is a type of GAN. It consists of

• A disriminator, which is trained to distinguish a "counterfeit" MNIST image from a real MNIST image. In addition, the disciminator part of ACGAN is also trained to classify an MNIST image into the ten digit categories.
• A generator, which is trained to generate realistic-looking MNIST images and "fool" the discriminator. The generator takes two inputs:
  1. A fixed-length latent vector, which can be viewed as a "random primer" for the fake generation process.
  2. An integer in the interval [0, 10), which indicates the desired digit category.
  Using these two inputs, it outputs a fake MNIST image.

Load Hosted model Generate