From the course: AI Workshop: Hands-on with GANs with Deep Convolutional Networks

Generator and discriminator

From the course: AI Workshop: Hands-on with GANs with Deep Convolutional Networks

Start my 1-month free trial Buy for my team

Generator and discriminator

“

Now that we've understood convolutional layers and pooling layers, let's do a quick overview of the generator and the discriminator adversaries in a generative adversarial network. This is just to refresh our memory. Here is an architectural overview of the GAN model. Notice the components, the generator, and the discriminator. The objective or the role of the generator in this network is to generate samples of fake data, and in order to do so, the generator takes in some random noise at the input. In addition to train a GAN, you also need a database of real instances. Now we are going to be training our GAN using images, so let's just take images as our example. This data set of real images will make up our training data, and the generator needs to produce images that look like these real images. The real images from the real database, and the fake images generated by the generator are both fed in as input to the discriminator. Now the discriminator has to identify what images are real and what images are fake. It needs to output a probability score for each image, indicating whether it's real or fake. Essentially, the role of the discriminator is just that of a traditional classification model; identify real and fake images and categorize them as such. In order to have the generator generate images at its output, the generator needs some latent noise which it uses to create data, and this all put together gives us the architecture of a GAN.

Contents