This repository is built for the paper StyleGAN: Large Scale Style Transfer Using Generative Adversarial Networks. Please refer to it for more details.
- Python3 (only tested on 3.6.4)
- Pytorch 0.4.0 (does not support other versions)
- NumPy (only tested on 1.14.2)
- Matplotlib (only tested on 2.2.0)
- Pillow (only tested on 5.1.0)
- Jupyter Notebook
- GPU support (a Nvidia GTX 1080 Ti is recommended)
- datasets ------ where you should set your datasets
- dev ------ some in-developing codes
- images ------ images for this README
- models ------ codes for building our models
- paper ------ a copy of our paper
- savefigs ------ HD images generated from demo notebooks will be saved here
- savemodels ------ two pretrained generators are provided here
- styles ------ images for style transfer
- utils ------ codes for data processing, visualization, and small modules
- Generator_layer_understanding.ipynb ------ notebook for reproducing our layer analysis
- StyleGAN_training.ipynb ------ notebook for reproducing our fine-tuning process
- train_fixed_DCGAN.py ------ script for training a 64x64 DCGAN
- train_fixed_LSGAN.py ------ script for training a 112x112 LSGAN
- Clone this repo:
git clone -b master --single-branch https://github.com/dashidhy/styleGAN.git
cd ./styleGAN- Make sure you have all prerequisites ready
Pretrained models and Jupyter Notebooks are provided for our fine-tuning process. You can play with them without downloading datasets. If you would like to reproduce the whole work, please read the sections below.
In our work, we use church-outdoor class of LSUN dataset for GAN training. Please refer to their README file for downloading. Finally, you should have a folder named church_outdoor_train_lmdb under ./datasets/LSUN/. You can also implement your own datasets, but this may lead to some works on the source codes.
To train a 64x64 DCGAN, run:
python3 train_fixed_DCGAN.py -dTo train a 112x112 LSGAN, run:
python3 train_fixed_LSGAN.py -dBy using -d argument, you will run a training under our default settings. If you would like to set you own hyperparameters, use argument -h for help, or read and modify source codes directly.