A PyTorch reimplementation of the paper Free-Form Image Inpainting with Gated Convolution (DeepFillv2) (https://arxiv.org/abs/1806.03589) based on the original TensorFlow implementation.

Example images (raw | masked | inpainted):

The models in networks_tf.py can be used with the weights from the official repository, which I have converted to PyTorch state dicts.

Download converted weights: Places2 | CelebA-HQ (for networks_tf.py)

The networks in networks_tf.py use TensorFlow-compatibility functions (padding, down-sampling), while the networks in networks.py do not. In order to adjust the weights to the different settings, the model was trained on Places2/CelebA-HQ for some time using the pretrained weights as initialization.

Download fine-tuned weights: Places2 | CelebA-HQ (for networks.py)

Before running the following commands make sure to put the downloaded weights file into the pretrained folder.

python test.py --image examples/inpaint/case1.png --mask examples/inpaint/case1_mask.png --out examples/inpaint/case1_out_test.png --checkpoint pretrained/states_tf_places2.pth

The Jupyter notebook test.ipynb shows how the model can be used.

Train with options from a config file:

python train.py --config configs/train.yaml

Run tensorboard --logdir <your_log_dir> to see the TensorBoard logging.

The web app uses a JS/React frontend and a FastAPI backend. To run it you need the following packages:

• fastapi, python-multipart: for the backend api
• uvicorn: for serving the app

Install with: pip install fastapi python-multipart "uvicorn[standard]"

Run with: python app.py

New models can be added in app/models.yaml

• python3
• pytorch
• torchvision
• numpy
• Pillow
• tensorboard
• pyyaml