This repository is deprecated and does not work as planned. The new version is released on the Inpainting-using-Context-Encoder-in-Generative-Adversarial-Network-V2 repo.
I was excited about the inpainting project from here, which in turn is based on the paper Context Encoders: Feature Learning by Inpainting. Therefore I wanted to try my own implementation. Some parts are from the other project, so have a look at the other project and the original paper. In addition, I also added some own modifications, e.g. an additional layer in the decoder or the tiling of the pictures.
The following result needed 1:10h with batch size of 1 and 30k epochs on a GTX1080TI.
First I created the directory weights. Then I copied the training data into the directory training.
For the upper training I used the following command:
python3 train.py -mar 3 -mal 0 -mac 0 -bs 1
The arguments -mar define the max. ammount of random ractangles for the random mask.
In addition I disable the random drawing of lines and circles by setting -mal and -mac to 0.
But you can play around.
Then I run the following command to perform a prediction:
python3 predict.py -i ./media/0.jpg -m ./media/0_mask.jpg -o ./media/0_prediction.jpg
The Input images and masks looked like the Masked-Image and Mask from the upper figure.
Where the white area characterised the region to inpaint.
You can get more information about available arguments by using the argument -h.