Pix2Pix

⭐ Star this project on GitHub — it helps!

Pix2Pix is an example of image-to-image translation GAN.

Install

git clone https://github.com/akanametov/pix2pix
cd ./pix2pix

Usage

Datasets

This project allows to train Pix2Pix GAN on three datasets:

• Cityscapes
• Facades
• and Maps ,so that each of them is going to downloaded automatically by following:

from dataset import Cityscapes, Facades, Maps

dataset = Facades(root='.',
                  transform=None, # if transform is None, then it returns PIL.Image
                  download=True,
                  mode='train',
                  direction='B2A')

Transforms

For simple conversion of imageA and imageB from PIL.Image to torch.Tensor use dataset.transforms:

from dataset import transforms as T
# it works almost like `torchvision.transforms`
transforms = T.Compose([T.Resize(size=(..., ...)), # (width, height)
                        T.CenterCrop(size=(..., ...), p=...), # (width, height); probability of crop/else resize
                        T.Rotate(p=...), # probability of rotation on 90'
                        T.HorizontalFlip(p=...), # probability of horizontal flip
                        T.VerticalFlip(p=...), # probability of vertical flip
                        T.ToTensor(), # to convert PIL.Image to torch.Tensor
                        T.Normalize(mean=[..., ..., ...],
                                     std=[..., ..., ...])])

As input, transforms take one/or two arguments (imageA/or imageA and imageB):

imgA_transformed = transforms(imgA)
# or
imgA_transformed, imgB_transformed = transforms(imgA, imgB)

There are also other transforms:

T.ToImage() # to convert torch.Tensor to PIL.Image 
T.DeNormalize(mean=[..., ..., ...],
               std=[..., ..., ...])])

Training

To train Pix2Pix on one of available datasets use train.py:

./pix2pix python train.py [--epochs EPOCHS] [--dataset DATASET] [--batch_size BATCH_SIZE] [--lr LR]

Train Pix2Pix

optional arguments:
  -h, --help            show this help message and exit
  --epochs EPOCHS       Number of epochs
  --dataset DATASET     Name of the dataset: ['facades', 'maps', 'cityscapes']
  --batch_size BATCH_SIZE
                        Size of the batches
  --lr LR               Adams learning rate

Results

Facades:

Maps

License

This project is licensed under MIT.

Links

• Image-to-Image Translation with Conditional Adversarial Networks (arXiv)