Face Generation GAN based on DCGAN architecture

Architecture

Examples

Mnist

Celeba

Details

codes for layers

• 1 layer = (size)F|C(ks,str,p)|DC(ks,str)|BN|D|LR
• (size): size of layer or filters
• F: Fully Connected
• C: Convulution
• DC(ks,str): Deconvulution with (Kernel Size, Stride, padding)
• BN: Batch normalization
• D: Dropout
• LR: Leaky Relu

Generator:

Full: (7x7x1024)F -> D -> LR

DCONV1: (512)DC(3,2,same) -> BN -> D -> LR

DCONV2: (256)DC(3,2,same) -> BN -> D -> LR

DCONV3: (128)DC(5,1,same) -> BN -> D -> LR

DCONV_OUT: (channles)DC(5,1)

Discriminator:

CONV1: (64)C(5,1,valid) -> LR

CONV2: (128)C(5,1,valid) -> BN -> D -> LR

CONV3: (256)C(5,1,valid) -> BN -> D -> LR

CONV4: (512)C(5,2,valid) -> BN -> D -> LR

OUT: (flat)F

Hyperparameters for celeba set

• Leaky Relu Slope: 0.2
• Adam Optimizaer beta: 0.5
• Learning rate: 0.0003
• Batch Size: 16
• latent vector dimension: 100
• dropout: 0.5
• 1 epoch