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 data_large, val_large, and test_large as the subdirectories. Don't forget to specify the root of the dataset by --root ROOT when using train.py or test.py

• Generate masks by following [1] (saved under ./masks by default). Note that the way of the mask generation is different from the original work

python generate_data.py

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)

• [1]: Unofficial implementation in Chainer