In this project, I defined and trained a DCGAN model with extra GDPP Loss on a dataset of faces. The goal was to get a generator network to generate new images of faces that look as realistic as possible!
The CelebA dataset contains over 200,000 celebrity images with annotations. Since you're going to be generating faces, you won't need the annotations, you'll only need the images. Note that these are color images with 3 color channels (RGB)#RGB_Images) each. You can download the CelebA dataset by clicking here.
A GAN is comprised of two adversarial networks, a discriminator and a generator.
- The inputs to the discriminator are 32x32x3 tensor images.
- The output should be a single value that will indicate whether a given image is real or fake.
(https://arxiv.org/abs/1812.00068).
DPP is an elegant probabilistic measure used to model negative correlations within a subset and hence quantify its diversity.
| Convolutional Layers | Input Dimension | Output Dimension | Activation Fucntion | Batch Normalization |
|---|---|---|---|---|
| layer 1 | 3 | 64 | leaky relu | False |
| layer 2 | 64 | 128 | leaky relu | True |
| layer 3 | 128 | 256 | leaky relu | True |
- The inputs to the generator are vectors of some length z_size.
- The output should be a image of shape 32x32x3.
| Deconvolutional Layers | Input Dimension | Output Dimension | Activation Function | Batch Normalization |
|---|---|---|---|---|
| Layer 1 | 512 | 256 | relu | True |
| Layer 2 | 256 | 128 | relu | True |
| Layer 3 | 128 | 64 | relu | True |
| Layer 4 | 64 | 3 | relu | False |
-
Epochs = 40
-
Optimizer = Adam
-
Z size = 100
-
Weight Inititalization
- Initialize the weights to a normal distribution, centered around 0, with a standard deviation of 0.02.
- The bias terms, if they exist, may be left alone or set to 0.
The dataset is biased; it is made of "celebrity" faces that are mostly white. After training the dataset over 40 epochs, I got the following result.
