CycleGAN-Finphot2Art

First words

In this repository, I build and train a CycleGAN model from scratch, applying it to photos of landscapes from where I live in Joensuu, Finland. The goal is to create some amazing artwork.

Requirements

• Python3
• tqdm
• numpy
• matplotlib
• seaborn
• opencv-python
• torch
• torchvision

Install

First, clone the repository.

git clone https://github.com/HoangPham3003/CycleGAN-Finphot2Art.git
cd CycleGAN-Finphot2Art

Second, create and activate the python environment.

python3 -m venv .venv
source .venv/bin/activate

Finally, install requirements.

pip install --upgrade pip
pip install -r requirements.txt

Train

Default parameters
To start training with the default parameters, run:

python train.py

Custom parameters
To customize training, you can adjust the parameters as follows:

python train.py -pt CycleGAN.pt -lr 0.0002 -ep 20 -bs 1 -ds 200 -ts 256 -d cuda -s True 

For a detailed explanation of the training parameters, refer to the train.py

Infer

To generate a new artwork from a photo:

python infer.py -pt CycleGAN.pt -ip img1.jpg -sd inference -d cuda

For a detailed explanation of the inference parameters, refer to the infer.py

Results

In this experiment, I use the vangogh2photo dataset for training. More dataset (cityscapes, horse2zebra, maps, monet2photo,...) can be found at here.The model is trained for 20 epochs with a batch size of 1, resulting in a total of 130,000 steps.

• First, I examine the generator and discriminator losses:



• Second, I review the fake and real images at some steps.

Methodology

A detailed explanation of the proposed method can be found in the official CycleGAN paper Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

Acknowledgements

Very thanks to authors who have implementations of CycleGAN:

1. Official CycleGAN by @junyanz
2. CycleGAN in pytorch by @aitorzip
3. GAN specialization tutorial by Coursera