Pytorch-CartoonGAN

An unofficial Pytorch demo of Paper: CartoonGAN: Generative Adversarial Networks for Photo Cartoonization (http://openaccess.thecvf.com/content_cvpr_2018/papers/Chen_CartoonGAN_Generative_Adversarial_CVPR_2018_paper.pdf)

Prerequisites

Code is intended to work with Python 3.6.x, it hasn't been tested with previous versions
For training and testing, the code needs pytorch
For data preparation, the code needs opencv

Training

1. Setup the dataset

First, you will need to download and setup a dataset.
You can download the dataset from the Baidu net disk: https://pan.baidu.com/s/19FIv4Fjby1knOTXJDWcXBw

Alternatively you can build your own dataset by setting up the following directory structure:

.
├── data                   
|   ├── train              # Training
|   |   ├── Cartoon        # Contains Cartoon images, with screenshots of cartoons, you can make by FameCrop.py
|   |   ├── Cartoon_blur   # Contains Cartoon_blur images, with images in data/train/Cartoon, you can make by edgeDilate.py
|   |   └── Photo          # Contains Photo images
|   └── test               # Testing
|   |   └── B              # Contains test photo images

python FameCrop.py

This command will randomly crop screenshots(under the data/train/screenShots directory) to 256 * 256

python edgeDilate.py

This command will dilate edges of cartoon images(under the data/train/Cartoon directory)

2. Train

python ./train --dataroot ./data/ --cuda --initialization

This command will start a training session of Initialization phase.
After training 10 epochs, it will save model ./output/initial_checkpoint.pth

python ./train --dataroot ./data/ --cuda --load_model ./output/initial_checkpoint.pth

This command will start a training session of Cartoon GAN.
Both generators and discriminators weights and the will be saved under the output directory, the generated test images will save in the directory output/cartoon_Gen.

If you don't own a GPU remove the --cuda option, although I advise you to get one!

Testing

python ./test --dataroot ./data/ --cuda --load_model ./output/checkpoint100.pth

This command will take the images under the dataroot/test directory, run them through the generators and save the output under the output/cartoon_Gen directories.
As with train, some parameters like the weights to load, this can be set by --load_model.

Examples of the generated outputs:

Notes

Due to the lack of coumputing resources, this Code has some problems that have not been solved. The results that have been shown is the best results in epoch 93. The problems are as follows:

1. With the increases of epochs, the results are not always good.
2. The results have not been as good as the official predict implementation, in the extent of simplifying and the lack of clear black edge of generated images.
3. Althought balanced by content loss, it has the problem of model collapse in some epochs.
   

Acknowledgments

Code is base on https://github.com/aitorzip/PyTorch-CycleGAN