Official implementation is released by the authors.
Note that this is an ongoing re-implementation and I cannot fully reproduce the results. Suggestions and PRs are welcome!
This is an unofficial pytorch implementation of a paper, Image Inpainting for Irregular Holes Using Partial Convolutions [Liu+, arXiv2018].
- Python 3.6+
- Pytorch 0.4.1+
pip install -r requirements.txt
- download Places2 and place it somewhere. The dataset should contain
test_largeas the subdirectories. Don't forget to specify the root of the dataset by
--root ROOTwhen using
- Generate masks by following  (saved under
./masksby default). Note that the way of the mask generation is different from the original work
CUDA_VISIBLE_DEVICES=<gpu_id> python train.py
CUDA_VISIBLE_DEVICES=<gpu_id> python train.py --finetune --resume <checkpoint_name>
CUDA_VISIBLE_DEVICES=<gpu_id> python test.py --snapshot <snapshot_path>
Here are some results from the test set after the training of 500,000 iterations and fine-tuning (freezing BN in encoder) of 500,000 iterations. The model is available here, but I don't ensure the quality. (Top to bottom: input, mask, image generated by the network, image which is combined with the original non-masked region of image, ground truth)