You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
a174e3b572 | 7 years ago | |
---|---|---|
samples | 7 years ago | |
README.md | 7 years ago | |
dcgan.png | 7 years ago | |
dlnd_face_generation.html | 7 years ago | |
dlnd_face_generation.ipynb | 7 years ago | |
helper.py | 7 years ago | |
problem_unittests.py | 7 years ago |
README.md
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