pytorch-CartoonGAN
Pytorch implementation of CartoonGAN [1] (CVPR 2018)
- Parameters without information in the paper were set arbitrarily.
- I used face-cropped celebA (src) and anime (tgt) collected from the web data because I could not find the author's data.
Tensorflow version
Usage
1.Download VGG19
2.Train
python CartoonGAN.py --name your_project_name --src_data src_data_path --tgt_data tgt_data_path --vgg_model pre_trained_VGG19_model_path
Folder structure
The following shows basic folder structure.
βββ data
β βββ src_data # src data (not included in this repo)
β β βββ train
β β βββ test
β βββ tgt_data # tgt data (not included in this repo)
β βββ train
β βββ pair # edge-promoting results to be saved here
β
βββ CartoonGAN.py # training code
βββ edge_promoting.py
βββ utils.py
βββ networks.py
βββ name_results # results to be saved here
Resutls
paper results
celebA2anime face
Initialization phase (reconstruction)
Input - Result (this repo) |
Cartoonization
- I got the author's results from CaroonGAN-Test-Pytorch-Torch.
Input - Result (this repo) | Author's pre-trained model (Hayao) | Author's pre-trained model (Hosoda) |
Development Environment
- NVIDIA GTX 1080 ti
- cuda 8.0
- python 3.5.3
- pytorch 0.4.0
- torchvision 0.2.1
- opencv 3.2.0
Reference
[1] Chen, Yang, Yu-Kun Lai, and Yong-Jin Liu. "CartoonGAN: Generative Adversarial Networks for Photo Cartoonization." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
(Full paper: http://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_CartoonGAN_Generative_Adversarial_CVPR_2018_paper.pdf)