Face Generation using Deep Convolutional GANs

Overview

In this project, I applied Deep Convolutional GANs (DCGAN) to generate new human face images. CelebA dataset was used as our training input, and below shows some example pictures in CelebA.

CelebA examples

Database

• Large-scale CelebFaces Attributes (CelebA) Dataset

Model Architecture:

• Discriminator: 5-layer CNN with batch normalization layer in between.
• Generator: 4-layer CNN with batch normalization layer in between and tanh as active function.

Hyperparameters:

• Batch Size: 64
• Epoch: 20
• Learning Rate (Discriminator): using Adam with lr = 0.0005 and [beta1, beta2] = [0.5, 0.999]
• Learning Rate (Generator): using Adam with lr = 0.0004 and [beta1, beta2] = [0.5, 0.999]
• Generated Image Size: 32 x 32
• Length of latent vector z: 100
• Number of Filters in Discriminator's first hidden layer: 32
• Number of Filters in Generator's first hidden layer: 32

Result

New human images generated by the trained model