Tensorflow implementation of GAN on Dataset CelebA.
| Generator | Discriminator |
|---|---|
| z, 256 | x, 128*128*3 |
| fc, 8*8*256, BN | 5*5 conv, 64, st.2 |
| 5*5 deconv, 256, st.2, BN | 5*5 conv, 128, st.2, BN |
| 5*5 deconv, 256, st.1, BN | 5*5 conv, 256, st.2, BN |
| 5*5 deconv, 256, st.2, BN | 5*5 conv, 512, st.2, BN |
| 5*5 deconv, 256, st.1, BN | fc, 1024 |
| 5*5 deconv, 128, st.2, BN | fc, 1 |
| 5*5 deconv, 64, st.2, BN | cross entropy loss |
| 5*5 deconv, 3, st.1 | |
| l2 loss |
- Python 3.3+
- Tensorflow 1.4.0
- SciPy
- pillow
To train a model with dataset CelebA:
$ python main.py
To test with an existing model:
$ python main.py --train=False
| After 1th epoch | After 20th epoch |
|---|---|
 |
 |
Details of the loss of Discriminator and Generator.
The following shows basic folder structure.
├── main.py # gateway
├── data
│ ├── celebA # celebA data (not included in this repo)
│ ├── xxxx.jpg
│ ├── xxxx.jpg
│ ├── xxxx.jpg
│ └── xxxx.jpg
├── GAN.py # build GAN
├── model.py # Generator and Discriminator
├── ops.py # some operations on layer
├── utils.py # utils
├── logs # log files for tensorboard to be saved here
└── checkpoint # model files to be saved here
This implementation has been based on this repository and tested with Tensorflow 1.4.0 on Ubuntu 16.04(GeForce GTX 1080 Ti).
Cong Yu / @congyucn