Residual Inpainting Using Selective Free-Form Attention

Prerequisites

• Python 3.6
• PyTorch 1.2
• NVIDIA GPU + CUDA cuDNN
• Some other frequently-used python packages, like opencv, numpy, imageio, etc.

Datasets prepration

We use CelebA, Paris Street View and Places2 datasets. The irregular mask dataset is available from here. After Downloading images and masks, create .filst file containing the dataset path in ./datasets/ (some examples have been given, refer to so).

Training

To continue training, first download pretrained models from my OneDrive, and place .pth files in the ./checkpoints directory.

Please edit the config file ./config.yml for your training setting. The options are all included in this file, see comments for the explanations.

Once you've set up, run the ./train.py script to launch the training.

python train.py

Testing

Please download pretrained models from my OneDrive, and place .pth files in the ./checkpoints directory.

Use .test.py for testing, you can directly run this script without any arguments:

python test.py

By default, this will inpaint the example images under the examples/celeba/images with the masks examples/celeba/masks. The output results will be saved in ./results.

Note that please use the original image rather than masked image as the input, our model will do the masking operation. Using masked image as input will introduce corss-color artifact since our model contains downsampling process. This issue will be fixed in the future.

For customized path, here are some args：

• --G1 path to generator 1
• --G2 path to generator 2
• --input path to input images
• --mask path to masks
• --output path to results directory

Alternatively, you can also edit these options in the config file ./config.yml.

Acknowledgement

This project is modified based on the Edge-Connect Model proposed by Nazeri et al.