Shift-Net_pytorch

This repositity is our Pytorch implementation for Shift-Net, it is just for those who are interesting in our work and want to get a skeleton Pytorch implemention. The original code is https://github.com/Zhaoyi-Yan/Shift-Net. I will upload pytorch models in months(Sorry for the delay).

Prerequisites

Linux or OSX.
Python 2 or Python 3.
CPU or NVIDIA GPU + CUDA CuDNN.
Tested on pytorch 0.3

Getting Started

Installation

Install PyTorch and dependencies from http://pytorch.org/
Install python libraries visdom and dominate.

pip install visdom
pip install dominate

Clone this repo:

git clone https://github.com/Zhaoyi-Yan/Shift-Net_pytorch
cd Shift-Net_pytorch

tain and test

Download your own inpainting datasets.
Train a model:

python train.py

To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.
Test the model

python test.py

The test results will be saved to a html file here: ./results/.

If you find this work useful, please cite:

@InProceedings{Yan_2018_Shift,
author = {Yan, Zhaoyi and Li, Xiaoming and Li, Mu and Zuo, Wangmeng and Shan, Shiguang},
title = {Shift-Net: Image Inpainting via Deep Feature Rearrangement},
booktitle = {The European Conference on Computer Vision (ECCV)},
month = {September},
year = {2018}
}

A bug when training with single GPU

This verison of code makes that InnerCos DO NOT support single gpu training. It is weried and I have not idea to solve it now. If you train our model in a single GPU, please set skip=1 in options/base_options. Otherwise, when training with InnerCos working on a single GPU, please refer the code before commit aa2382b

Kindly remindier

If you find it a little hard to read the code, you may read Guides.

New things that I want to add

Make U-Net handle with inputs of any sizes. (By resizing the size of features of decoder to fit that of the corresponding features of decoder.
Update the code for pytorch >= 0.4.
Clean the code and delete useless comments.
Lots of more crazy operations that help in performance.

What I am doing now

Guides of our code, we hope it helps you understand our code more easily.
Directly resize the mask to save computation.
Guidance loss seems defined on the global region. Need make it work only in masked region.

Acknowledgments

We benefit a lot from pytorch-CycleGAN-and-pix2pix

Naruto-Sasuke/Shift-Net_pytorch