CartoonGAN (PyTorch)

Unofficial PyTorch implementation of CartoonGAN. We followed the original Lua training implementation from the paper author (Yang Chen).

Repo Structure

├─checkpoints
├─data
│  ├─train 
│  │  ├─cartoon # You put cartoon images here
│  │  ├─cartoon_edge_pair 
│  │  └─photo   # You put photo images here
│  └─val
│      └─photo # You put photo images here
└─results
    ├─test  
    └─train

Dependenies

• Albumentations

pip install -U albumentations

• tqdm

pip install tqdm

To start training

0. Read https://vinesmsuic.github.io/2022/01/21/i2i-cartoongan/ to understand the implementation
1. Prepare the photo and cartoon data
2. Get the pre-trained VGG19 weight and put it in the CartoonGAN folder : https://download.pytorch.org/models/vgg19-dcbb9e9d.pth
3. Preprocess data through edge promoting

python edge_smooth.py

4. Edit config.py

• If you are using your custom data that are in random size, please enable RandomCrop in config.py.

transform_cartoon_pairs = A.Compose(
    #additional_targets: apply same augmentation on both images
    [   
        A.RandomCrop(width=IMAGE_SIZE*1.2, height=IMAGE_SIZE*1.2),
        A.Resize(width=IMAGE_SIZE, height=IMAGE_SIZE),
        A.HorizontalFlip(p=0.5),
        A.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], max_pixel_value=255.0),
        ToTensorV2(),
    ], 
    additional_targets={"image0": "image"},
)

5. Training

python train.py

• The training consist of initialization phase and training phase.

6. Wait for a long time and see the results at results folder

TODO

• [] Automatic Mixed Precision
• [] LR Scheduler
• [] Loss visualization
• [] WandB visualization
• [] Inference Code
• Explaining Code

Working Environments

• Windows with CUDA
• Ubuntu with CUDA