This repo is trying to reproduce some results of awesome Deepfill v2 (Project & Code) because I personally prefer pytorh. Besides, I abuse the awesome detectron2 lib to implement although it is originally designed for object detection, because I appreciate its well-organized codes. If you know more suitable tools, welcome to give recommendations.

This repo is yet to be finished and tested.

• Build up the model.
• Translate the pretrained tensorflow model into pytorch.
• Fix the bug of converting tensorflow pretrained model to pytorch. Tensorflow behaves slightly different from Pytorch on Conv2d when stride is greater than 1 (e.g. 2). Hence, I deal with this issue by manually striding the convolutional feature map. Moreover, original tensorflow requires nearest neighbor downsample with align_corners=True while official pytorch interpolate does not support align_corners=True when mode="nearest". Therefore, I write my own downsampling functions enabling the align_corners.
• Evaluate the pretrained model on Places2 and CelebA-HQ.
• Train the model on Places2 and CelebA-HQ.

Now we can reproduce the demo results given in the original repo.

• Python==3.6
• Pytorch==1.3.0 (yet not tested for higher version)
• detectron2==0.1

The pretrained model is converted from tensorflow to pytorch using param_convertor.py. You can download the tensorflow pretrained model Places2 and convert the parameters or directly download the converted model. Make sure the folder that contains the pretrained model is like

./output
./output/pretrained/
./output/pretrained/places2_256_deepfill_v2.pth 

run the jupyter notebook file ./inpaint_demo.ipynb. The results are dumped in the folder ./demo_outputs.

TO BE COMPLETED

TO BE COMPLETED