Generative modeling is an unsupervised learning task in machine learning that involves automatically discovering and learning the patterns in input data in such a way that the model can be used to generate new examples which are not in the original dataset, If you are new to GAN you can read my blog on GAN. DCGAN, or Deep Convolutional GAN, is such a network architechture that is capable of generating RGB images of size 64x64. The idea of Deep Convolutional generative adversarial network was proposed in this paper.
The network architecture is shown in the below figure.
The Network hyperparameters are,
Face_images_for_training = 60000
IMAGE_HEIGHT = 64
IMAGE_WIDTH = 64
IMAGE_CHANNEL = 3
NOISE_SIZE = 100
DISCRIMINATOR_LEARNING_RATE = 0.0002
GENERATOR_LEARNING_RATE = 0.0002
BATCH_SIZE = 128
EPOCHS = 150
BETA1 = 0.5
WEIGHT_INIT_STDDEV = 0.02
EPSILON = 0.00001
This project uses DCGAN to generate fake face images of celebrities, the model was trained on the CelebFaces Attributes dataset. The below GIF shows how the generator learns to generate a fake face over time.
The model was trained for 150 epochs and then it was able to generate decent fake face images, some of them are given below.
