bw2color's Introduction

bw2color

Project for the course Artificial Intelligence in Practice. Use cGAN to colorize Anime images. Reference: https://arxiv.org/abs/1611.07004

~~这个 README 暂时很不准确，与代码结构不匹配，需要重构~~

Dataset

nico-opendata

400k 张动漫图片，需要申请下载，质量一般
danbooru2017 512x512

主要是女性动漫角色的预处理好的 512x512 大小的图片，质量较高，但风格不统一
海贼王漫画图片

海贼王的漫画图片数据 ~~到处都可以下到，提供一个 github 上的~~

project structure

├── data/               // original dataset
├── model/              // saved model
├── out/                // preprocessed image
├── test_result/        // test result image
├── training_result/    // training result image 
├── tfrecord/           // tfrecord folder
├── backward.py         // nn backward
├── forward.py          // nn forward
├── generateds.py       // tfrecord generate and get
├── test.py             // result test
└── tools           
    ├── preprocess.py   // cut image
    ├── run.sh          // py run script
    ├── upload.sh       // upload py files to server
    └── wallpaper.py    // spiders

Usage

download wall.alphacoders.com thumb image

cd bw2color
python tools/wallpaper.py

crop downloaded image to 256x256 and convert to grey

cd bw2color
# basic use
python tools/preprocess.py
# set data dir
python tools/preprocess.py --data [data dir] --save  [output dir]
# set process method
python tools/preprocess.py --method gray
python tools/preprocess.py --method sketch --mod_path [mod file path]

generate tfrecord