I am currently refactoring the codes for the latest version of PyTorch. I will update the codes and upload the pretrained models (e.g. Places and CelebA) soon. Apologies for the inconvenience.
This is the official implementation of the paper "Global and Local Attention-Based Free-Form Image Inpainting" published in Sensors (paper). Currently we are reformatting the codes. We will upload the pretrained models soon.
- Python3
- PyTorch 1.0+ (The code works up to PyTorch 1.4. There seems to be an auto-grad problem with PyTorch 1.5. I will update the code for PyTorch 1.5 after finding the underlying issue.)
- Torchvision 0.2+
- PyYaml
If you find our paper and code beneficial for your work, please consider citing us!
@article{uddin2020global,
title={Global and Local Attention-Based Free-Form Image Inpainting},
author={Uddin, SM and Jung, Yong Ju},
journal={Sensors},
volume={20},
number={11},
pages={3204},
year={2020},
publisher={Multidisciplinary Digital Publishing Institute}
}
- Set directory path in "configs/config.yaml". -- Set dataset name, if needed. -- If the dataset has subfolders, set "data_with_subfolder" to "True".
- Run
python train.py --config configs/config.yaml - To resume, set "resume" to True in "configs/config.yaml". Currently it overwrites the previous checkpoints. Updated code will have checkpoints listed.
- To view training, run
tensorboard --logdir checkpoints/DATASET_NAME/hole_benchmark
- Modify "test_single.py" as per need and run.
- Bulk testing code will be uploaded soon.
- Pretrained models will be uploaded soon.
- Places dataset
- ImageNet dataset
- CelebA dataset
- Ablation study of the modules
- Code base: This code is heavily relied on this repo. Kudus!!!