jzsherlock4869 / cyclegan-pytorch

An easy-to-modify and easy-to-follow re-implementation of CycleGAN (cycle-consistent generative adversarial network) in PyTorch

Implementation of CycleGAN (pytorch)

A simple & easy to modify implementation of CycleGAN (Cycle-consistent generative adversarial network) in pytorch

Basic illustration of CycleGAN (from original paper):

Results from re-implementation of this repo using "horse2zebra" dataset (without parameter tuning, which still has some obvious artifacts, you can tune the hyper-parameters to make it better~ )

Installation and preparation

Download Commonly used datasets for CycleGAN and use them to train and validate the code pipeline

1. Monet-Photo transfer: Kaggle Monet-Photo transfer
2. Horse-Zebra transfer: Kaggle Horse-Zebra transfer

then prepare the dataset folder in following structure:

├── monet_dataset
│   ├── monet_jpg
│   └── photo_jpg
└── zebra_dataset
    ├── testA
    ├── testB
    ├── trainA
    └── trainB

Git clone this repo and cd to the root folder

git clone https://github.com/jzsherlock4869/cyclegan-pytorch
cd cyclegan-python

install necessary python packages according to requirements.txt in the folder

Start Training

modify the dataroot of config/000_base_cyclegan_horse2zebra.yml to your own path for dataset, and run the train process:

python train_cyclegan.py --opt configs/000_base_cyclegan_horse2zebra.yml

Dataset class for "horse2zebra" and "photo2monet" are already implemented in this repo.

For training on your own datasets (domain A and domain B), write your own dataloader as get_photo2monet_train_dataloader and get_horse2zebra_train_dataloader in data/sample_dataloader.py

def get_your_custom_train_dataloader(root_dir="your_path", 
                                    batch_size=8, 
                                    img_size=(256, 256)):
    imgA_sub, imgB_sub = "subdirnameA", "subdirnameB" # sub directory name to your root_dir
    postfix_set=["jpg"]  # which postfix is your images
    train_dataset = CycleGANDataset(root_dir, imgA_sub, imgB_sub, postfix_set, img_size)
    train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=False)
    return train_dataloader

and then modify the train_cyclegan.py line 54-60 to add your dataset (remember to import them first! )

if which_dataset == 'horse2zebra':
    train_dataloader = get_horse2zebra_train_dataloader(dataroot, 
                                                        batch_size=batch_size, 
                                                        img_size=img_size)
elif which_dataset == 'photo2monet':
    train_dataloader = get_photo2monet_train_dataloader(dataroot, 
                                                        batch_size=batch_size, 
                                                        img_size=img_size)
 # add lines here
 elif which_dataset == 'your_custom_dataset':
    train_dataloader = get_your_custom_train_dataloader(dataroot, 
                                                        batch_size=batch_size, 
                                                        img_size=img_size)
# add lines here
else:
    raise NotImplementedError(f"Unrecognized dataset type : {which_dataset}")

References

This code is an re-implementation for CycleGAN, which is more easy-to-follow and easy-to-modify, especially for beginners. Original paper is:

@inproceedings{CycleGAN2017, title={Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networkss}, author={Zhu, Jun-Yan and Park, Taesung and Isola, Phillip and Efros, Alexei A}, booktitle={Computer Vision (ICCV), 2017 IEEE International Conference on}, year={2017} }

A keras version and tutorial with detailed explanations about the theory and process of CycleGAN:

https://machinelearningmastery.com/cyclegan-tutorial-with-keras/

Moreover, this codebase also refered BasicSR and UnpairedSR for the code architecture and style, and some functions are directly borrowed from them. Appreciations for their great works~

Welcome to star⭐ if this repo helps you :)