Training CartoonGAN

This is the repo for training the CartoonGAN. The original paper of CartoonGAN can be found here.

Requirments

• Make sure you have set up a python environment for tensorflow. More details of instructions can be found here.
• The version of the modules we use is listed in requirements.txt.

Usage

1) Download this repository

Start by downloading this repo or clone this repository:

git clone https://github.com/Derek-Wds/training_CartoonGAN.git
cd training_CartoonGAN

2) Collect data

The dataset folder structure is as followed:

── dataset
      ├── photo_imgs
      |   ├── photo1.jpg
      |   ├── photo2.jpg
      |   └── ...
      |
      ├── cartoon_imgs
      |   ├── cartoon1.jpg
      |   ├── cartoon2.jpg
      |   └── ...
      |
      ├── smooth_cartoon_imgs
      |   ├── smooth1.jpg
      |   ├── smooth2.jpg
      |   └── ...
      |
      └── ...

You could use your own dataset, but pay ATTENTION: Make sure you move the download images to the dataset folder which contains three sub-folders: photo_imgs, cartoon_imgs and smooth_cartoon_imgs. After that you have to run the smooth_edge.py to get the Gaussian edge-smoothed images.

python smooth_edge.py

After making the dataset, now should use folllowing command to preprocess the images in order to feed them into the model. This step will create corresponding folders for the numpy files.

python preprocess.py

If you do not have available data on hand, you could use the download.py to get the photos as well as cartoon images to be used during training.

python download.py

3) Train

Run the training script with default arguments:

python main.py

Or you could specify your preferred hyperparameters settings:

python main.py --batch_size=32 \
--epochs=500 \
--gpu_num=2 \
--image_channels=3 \
--image_size=256 \
--init_epoch=30 \
--lr=0.0002

Or you could run bash script run.sh:

bash run.sh

If you want to resume the training, you could just use load_weights to load the saved weights of the modals (generator, discriminator, and train_generator) and keep training.

4) Visulization

We provide access to the visualization of loss and generated images. You could use following command to do this:

tensorboard --logdir logs --port 9090

Then you could have access to tensorboard at localhost:9090.

5) Use it

That's it and have fun with CartoonGAN in ml5js!