Reproducing "Recurrent Feature Reasoning For Image Inpainting" of CVPR 2020 by tensorflow
About half a day of training in NVIDIA V100 (32G)：
| Inputs | Outputs |
|---|---|
 |
 |
 |
 |
- tensorflow 2.0
- Lesser useless code
- Simpler core network architecture
- Faster configuration and running
network.py : Core network structure
utils.py : Other core utils
config.py : Parameter Configuration
run.py : Run the network
test.py : Test the network
train
python run.py /root/image_root_path/ /root/mask_root_path
test
python test.py /root/image_root_path/ /root/mask_root_path
All parameters are set in config.py.
CelebA : At Once
Places2 : TODO
Paris Street View ：TODO
Unlike the original version, I normalize all inputs to between 0 and 1, and use the sigmoid function for the output. Because I find the author's original code hard to converge.