PSR-Net (WWW 2021 )

Progressive Semantic Reasoning for Image Inpainting

Requirements

• tensorflow 2.0 （required）
• numpy 1.19.5 (optional)
• loader 1.0 (required)
• opencv 4.1.1.26 (optional)
• Pillow 6.0.0 (optional)
• pathlib 1.0.1 (optional)

Usage

run model

  python train.py

test model

  python test.py

All configuration option in config.py.

Training and Fine Tuning

training (Default)

  generator = build_model(mode = "training")

fine tuning (Option)

  generator = build_model(mode = "tuning")

Mask Dataset

We provide two ways to support loading the Mask dataset.

1. Existing Mask Dataset
   set train_mask_path and val_mask_path in config.py.
2. Generating Random Mask (we used the strategy of "gated conv" paper)
   set generated_mask=False in config.py

Pretrained model

• Places2
• CelebA
• Paris StreetView

Reference

Junjie Jin, Xinrong Hu, Kai He, Tao Peng, Junping Liu, and Jie Yang. 2021.Progressive Semantic Reasoning for Image Inpainting. InProceedings of theWeb Conference 2021 (WWW ’21 Companion), April 19–23, 2021, Ljubljana,Slovenia.ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3442442.3451142