Keras implementation of Progressive Growing of GANs for Improved Quality, Stability, and Variation.
Developed by BUAA Microsoft Student Club.
Leader Developers: Kun Yan, Yihang Yin, Xutong Li
Developers: Jiaqi Wang, Junjie Wu
- Python3
- keras 2.1.2 (TensorFlow backend)
- CelebA Dataset
- Set main.py as the Startup File.
- Right click main.py, click "Start without Debugging" or "Start with Debugging" context menus.
- Please read the "Prepare the dataset" carefully if you had dataset issues.
- If you're able to access Google Drive, when you run main.py, it will automatically download the CelebA Dataset (1.5GB) to /datasets directory and unzip it to /datasets/CelebA directory.
- If you can not access Google Drive, please download the CelebA Dataset (It has Baidu Cloud Disk link), notice that we only need the "img_align_celeba.zip" file. Unzip it to /datasets/CelebA directory. Then when you run main.py it will skip the download process and begin training.
Notice: /datasets/CelebA should just contains 202599 pictures.
Parameters in train.py will determine the frequency of saving the training result snapshot. And if you want to resume a previous result, just modify train.py:
# In train.py:
image_grid_type = 'default',
# modify this line bellow
# resume_network = None,
# to:
resume_network = <weights snapshot directory>,
resume_kimg = <previous trained images in thousands>,
First download CelebA Dataset.
Run h5tool.py to create HDF5 format datatset. Default settings of h5tool.py will crop the picture to 128*128, and create a channel-last h5 file.
$ python3 h5tool.py create_celeba_channel_last <h5 file name> <CelebA directory>
Modify config.py to your own settings.
# In config.py:
data_dir = 'datasets'
result_dir = 'results'
dataset = dict(h5_path=<h5 file name>, resolution=128, max_labels=0, mirror_augment=True)
# Note: "data_dir" should be set to the direcory of your h5 file.
We only support CelebA dataset for now, you may need to modify the code in dataset.py and h5tools.py if you want to switch to another dataset.
$ python3 train.py
In train.py:
# In train.py:
speed_factor = 20
# set it to 1 if you don't need it.
"speed_factor" parameter will speed up the transition procedure of progressive growing of gans(switch resolution), at the price of reducing images' vividness, this parameter is aimed for speed up the validation progress in our development, however it is useful to see the progressive growing procedure more quickly, set it to "1" if you don't need it.
The operations are the same as VStdio/VSCode users.
So far, if your settings have no problem, you should see running information like our running_log_example
These two pictures are the training result we get so far, trained for 5 days on a NVIDIA GeForce 1080-ti GPU. You should be able to see the changes of resolution during the progressively growing procedure of our model.
We provide a set of pre-trained weights. Available at Google Drive and Baidu Cloud Disk.
There are two ways to use the pretrained weight:
- Use it as trained network to resume training.
- Use it to generate fake CelebA faces directly.
For the second purpose, we provide predict.py, download the weights and put it /pre-trained_weight directory (Under the same directory as /datasets )
Any bug report or advice, please contact us:
Kun Yan (naykun) : [email protected]
Yihang Yin (Somedaywilldo) : [email protected]
-
Progressive Growing of GANs for Improved Quality, Stability, and Variation, Tero Karras (NVIDIA), Timo Aila (NVIDIA), Samuli Laine (NVIDIA), Jaakko Lehtinen (NVIDIA and Aalto University) Paper (NVIDIA research)
-
tkarras/progressive_growing_of_gans (https://github.com/tkarras/progressive_growing_of_gans
-
github-pengge/PyTorch-progressive_growing_of_gans (https://github.com/github-pengge/PyTorch-progressive_growing_of_gans)
Our code is under MIT license. See LICENSE
We have translated the original paper to Chinese briefly, a markdown version is available now, hope this will benefits chinese users.