GAN-based Chinese Painting Generation

Author

黄哲昊 叶瑜超

Source code

pytorch-CycleGAN-and-pix2pix

Environment

• torch>=1.4.0
• torchvision>=0.5.0
• dominate>=2.4.0
• visdom>=0.1.8.8

Our datasets

• cp2photo_v1: contains 5382 photosincluding landscape photos, plant photos and animal pho-tos, 1070 Chinese paintings including Chinese landscape painting, Chinese bird-and-flower painting and Chinesefigure painting.
• cp2photo_v2: contains 826 photos including almostall landscape photos and 522 traditional Chinese paintingsincluding almost all landscape paintings.
• cp2photo_v3: contains 2194 landscapephotos and 2686 Chinese landscape paintings.

Baiduyunpan link for these three datasets: https://pan.baidu.com/s/1cEgMAYsL8wLR17XnGa-vGA, password: 03ci

开启一次训练完整流程

• 登录Xshell 6连接服务器au332p10

• tmux new -s visdom用于开启visdom训练可视化终端链接
  在tmux窗口内运行 python -m visdom.server
  然后ctrl b + d 退出tmux窗口

• 在自己电脑windows cmd终端使用ssh连接服务器
  ssh -L 8097:127.0.0.1:8097 -p 3333 [email protected]
  然后使用本地浏览器访问 http://localhost:8097 就可以实时监控训练情况了

• 使用tmux另开一个窗口运行训练程序tmux new -s session_name session_name为窗口名

• 在tmux窗口内运行 conda activate Pytorch

• cd pytorch-CycleGAN-and-pix2pix

• 开始训练
  

 --dataroot: path to images (should have subfolders trainA, trainB, valA, valB, etc)
 --name: name of the experiment. It decides where to store samples and models
 --model: chooses which model to use, we use 'cycle_gan'
 --gpu_ids: gpu ids: e.g. 0  0,1,2, 0,2. use -1 for CPU
 --display_env: visdom display environment name (default is "main")
 --netG: specify generator architecture, choose from [resnet_9blocks | unet_256 ]
 --n_epochs: number of epochs with the initial learning rate
 --n_epochs_decay: number of epochs to linearly decay learning rate to zero
 --n_epochs_D: number of epochs to train the D per training the G, which must > 0
 the total epochs = n_epochs * n_epochs_D + n_epochs_decay
 --lr_policy: learning rate schedular, choose from [linear | step | plateau | cosine]
 --addnoise: add which noise data augmentation to the training img or none, choose from ['SaltPepper', 'Gaussian', 'SaltPepper_and_Gaussian', None]

 python train_more_D.py --dataroot ./datasets/datasets_name --name save_name --model cycle_gan --gpu_ids 0 --display_env env --netG resnet_9blocks --n_epochs 100 --n_epochs_decay 200 --lr_policy cosine --n_epochs_D 3

• 训练结果保存在checkpoints文件夹下
• ctrl b + d 退出tmux窗口
• tmux attach -t session_name重新进入tmux窗口

使用已经训练好的模型得到测试结果

• 源代码中test.py使用epoch次的保存模型进行测试，得到Domain x 和Domain y的双向测试结果
  

 --epoch: which epoch to load?
 --num_test: how many test images to run
 python test.py --dataroot ./datasets/datasets_name --name save_name --model cycle_gan --gpu_ids 0 --epoch 50 --netG resnet_9blocks

• 我们的test_img.py只得到从原图转换为水墨画

 --save_path: where to save the result
 python test_img.py --dataroot ./datasets/datasets_name --name save_name --model cycle_gan --epoch 50 --netG resnet_9blocks --save_path ./result/test

毛笔建模Demo

 python brush/brushwritev2.py