DeepFill v2 PyTorch

WIP Unofficial implementation of "Free-Form Image Inpainting with Gated Convolution".

Setup

• Install conda environment

  conda env create -f environment.yml
  conda activate iminpaint
  

• Setup data

  • Put the training dataset into one folder of images, e.g., data/training_imgs

  • Run the edge extraction

    python iminpaint/data/scripts/generate_edge_masks --input_folder data/training_imgs --output_folder some_folder --prototxt path_to_prototxt --caffemodel path_to_model
    

    Download the prototxt from here and the caffe model from here.
    Optionally, one can add the --visualize flag to plot an example edge output.

  • Adjust the paths to the folder with training images and edges in the training script train.py in the Data dataclass.

Training

Hydra together with PyTorch-Lightning is used for implementation. To start default training:

python train.py

and a folder will be created with Tensorboard logs and checkpoints in ./outputs.

TODOs

• Implement Contextual Attention & respective network branch
• Add support for Places2 and CelebA-HQ faces datasets used in the paper
• Add pre-trained models