Chinese-Painting-Generator

Project for SJTU-AU332 (AI3603).

Implementing transferring a real photo containing natural scenery into Chinese painting style using CycleGAN and Neural Style Transfer.

Code Structure

• CycleGAN/: source code of CycleGAN
  • train.py: training script
  • test.py: test script
  • models/: model scripts
  • util/: util scripts
  • data/: dataset scripts
• NeuralStyleTransfer/: source code of Neural Style Transfer
  • main.py: main script to run neural style transfer
  • LossFunction.py: loss function script
  • data/: style images and content images
  • results/: some sample results

How to run the experiments

• CycleGAN

  To run training script, first, change directory to where the script locates, and put trainA and trainB into directory dataroot. Then use command python train.py --dataroot dataroot in terminal. Use command python train.py --help for more instructions.

  To run testing script, first, change directory to where the script locates, put test data into directory dataroot and put pretrained model into directory ./checkpoints/xxx where xxx is the name you defined in terminal. Then use command python test.py --dataroot dataroot --name name in terminal. Use command python test.py --help for more instructions.

  Our pretrained Model:

  Google Drive

  Baidu NetDisk (key: eplt)

• NeuralStyleTransfer

  First, change directory to where the main.py script locates, put content images into directory content_img_dir, and put style images into path style_img_path. Then use command python main.py --content_img_dir content_img_dir --style_img_path style_img_path in terminal. Use command python main.py --help for more instructions.

Requirements

The code requires only common Python environments for machine learning; Basically, it was tested with

• Python 3 (Anaconda 3.6.3 specifically)
• PyTorch==0.3.1
• numpy==1.18.5
• tqdm
• pillow
• matplotlib
• argparse

Higher (or lower) versions should also work (perhaps with minor modifications).

Dataset

download:

Google Drive

Baidu NetDisk (key: 1t61)

Adapted from :

Traditional Chinese Landscape Painting Dataset

Landscape Dataset

Acknowledgements

This repo borrows a lot from junyanz/pytorch-CycleGAN-and-pix2pix and pytorch tutorials.