Clone this repository:

git clone https://github.com/jaepoong/Colorization-GAN.git
cd Colorization-GAN

Install dependency:

conda create -n Colorization-GAN python=3.6.7
conda activate Colorization-GAN
pip install requirments.txt
conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

I provide a script to download afhq datasets. The datasets will be downloaded to the data directories.
In this code, I used AFHQ for Colorizing

AFHQ. To download the AFHQ dataset run the following commands:

bash download.sh afhq-dataset

In this code, I used Landscape for the original image of cartoonizing
Landscape. To download the Landscape Dataset please click the below link

Kaggle Landscape Dataset

python video_save.py

By Copyright, I cant publish the Cartoon dataset. If you don`t have one, I recommend useing the upright code. In the implementation, I used the 4500 Shinkai Makoto movie frame dataset.

example

After downloading dataset, you can train Networks by below implementation.
This code have two kinds Network Base and *Modified
Base model consists of U-Net architecture that is inferior than Modified model
*Modified" model changed a lot, including instance norm, Wgan loss etc.

*example for train Modified "

 python Gan_main.py --Mod True --model_save_path checkpoints/CycleGAN/ \  ## --model_save_path : save checkpoint at this directory
                    --cartoonizing True \ ## if true, dataloader receive img and target img dir else img become gray scaled of target img
                    --generated_image_save_path \ ## --generated_image save_path is the directory that save generated image each epoch
                    --photo_img_dir \ ##img directory path
                    --photo_image_target_dir \ ## target image dir path
                    --initialization_epochs \ ## for stable training, initialize small epochs for train Generator
                    --num_epochs \ ## i trained colorizing 50 epochs for AFHQ datset
                    --batch_size \ ## in 12GB vram, i used 6 epochs

After training, for evaluating the performance of the generating model, You can evaluate the Frechet Inception Distance(FID) the below code.
FID evaluates fidelity and diversity of output of the model. (lower = good)
In this implementation, Because of many parameters, The FID of base model is a little good than Modified model.

  python metrics/fid.py --paths path1 path2 ## --paths argument need two path for evaluate  

In this implementation, after training, because Base FID perfomance was better than the Modified model, I surveyed the performance of generated images.
Despite many artifacts in the base, the Low FID of the Modified model was not reliable.
So I surveyed output-image by approximately 100 people.
The result shows Modified model has high performance for human.

To test the model invariability, I implemented affine transform for cartoonizing.
It show that the Modified model is invariant for affine transform(rotation, cropping, etc)

At first, I made the same parameter for each base and Modified the model.
The performance of the base model was too inferior. The result is below.

This implementation is colorizing & cartoonizing by two models.
By this code, the User can easily train the nice model for the 256 resolution style transfer model.
If any questions, please contact [email protected] or issue