Deep Fusion Network for Image completion

arxiv: http://xxx.itp.ac.cn/pdf/1904.08060v1
the official code: https://github.com/hughplay/DFNet

Here we add training code to reproduce the author's results

Prerequisites

• Python 3
• PyTorch 1.1+
• OpenCV
• TensorboardX
• apex

preparation

• download your datasets (such as Celeba, Celeba-HQ, places365, paris)
• make flist for dataloader

$ cd core
$ python flist.py  # something params in the .py can be modified by yourself

Training

训练涉及到的超参均在config/xx.yaml下
如果想训练自己的数据集，可以视情况自行更改

# 以Celeba-HQ为例
python train_net ./config/celeahq.yaml

Testing

基于opencv写了一个简单的实时界面交互来进行修复测试
鼠标左键用于涂抹，radius控制线条的粗细

python ui.py <ckpt_path> <img_path>

<p align="center"> <img width="600" src="imgs/ui.jpg"> </p>

Tensorboard

loss和中间结果可视化

<p align="center"> <img width="600" src="imgs/tensorboardLine.jpg"> </p> <p align="center"> <img width="600" src="imgs/tensorboardImg.jpg"> </p>

讨论

模型结构未改动，loss也是参照作者提供的，实验过程中发现tv loss是递增的，尝试剔除它，结果影响不大