Using a GAN, the following code generates realistic faces that are completely AI generated. GAN stands for Generative Aversarial Network. A GAN uses two types of networks to train; a discriminator and a generator.
Determines if an input image is from the original dataset. Outputs a probability from 0 to 1. Numbers close to zero suggest the image was made by the genorator, where numbers close to 1 will be output if the discriminator suspects that the image is from the dataset.
Generates images from an input vector containing random numbers. Generates images similar to ones in the original dataset to fool the discriminator.
The model generates 128x128 images with three color channels. Using the trainNN function, the model can be trained and can generate images.
trainNN function:
epochs- Set to 0 to run the model without training. Every epoch represents one pass through the dataset of 70,000 images.batch_size- The number of images we train the model on at a time. We found that 128 works best.lr- How much the model parameters are adjusted after each iteration. Lr = 0.0002 works well for Adam Optimizer.save_time- The number of epochs before each save.save_dir- The file where the model is saved and loaded from.slide- Set true to use sliders and false to generate images.
There is a included slider option that allows the use of sliders to continuously transition between two different random vectors inputted in the generator. This creates a seamless transition between any 2 images the generator produces. This effect can be seen here.
The model was trained on the Flickr Dataset. This dataset contains high quality images with minimal artifacts. The dataset of 70000 images was stored in a numpy array with dimensions (70000, 3, 128, 128) saved to a .npy file.
Made by Kyler Nikolai and Emmett Jaakkola.
 |
 |
 |
 |
 |
 |