Awesome

waifu2x-chainer

This is a Chainer implementation of waifu2x [1]. Note that the training procedure of waifu2x-chainer may be slightly different from original waifu2x.

Summary

• 2D character picture (Kagamine Rin) is licensed under CC BY-NC by piapro [2].

Requirements

• Chainer
• CuPy (for GPU support)
• Matplotlib (for benchmark)
• ONNX-Chainer (for ONNX model export)
• Pillow
• Wand (for training)

Installation

Install Python packages

pip install chainer
pip install pillow

Enable GPU support

Install CuPy precompiled binary package which includes the latest version of cuDNN library.
See: CuPy Installation Guide

Getting waifu2x-chainer

git clone https://github.com/tsurumeso/waifu2x-chainer.git

Testing

cd waifu2x-chainer
python waifu2x.py

Usage

Omitting --gpu (-g) option run on CPU.

Noise reduction

python waifu2x.py --method noise --noise_level 1 --input path/to/image/or/directory --arch VGG7 --gpu 0

python waifu2x.py -m noise -n 0 -i path/to/image/or/directory -a 0 -g 0
python waifu2x.py -m noise -n 2 -i path/to/image/or/directory -a 0 -g 0
python waifu2x.py -m noise -n 3 -i path/to/image/or/directory -a 0 -g 0

2x upscaling

python waifu2x.py --method scale --input path/to/image/or/directory --arch VGG7 --gpu 0

python waifu2x.py -m scale -i path/to/image/or/directory -a 0 -g 0

Noise reduction + 2x upscaling

python waifu2x.py --method noise_scale --noise_level 1 --input path/to/image/or/directory --arch VGG7 --gpu 0

python waifu2x.py -m noise_scale -n 0 -i path/to/image/or/directory -a 0 -g 0
python waifu2x.py -m noise_scale -n 2 -i path/to/image/or/directory -a 0 -g 0
python waifu2x.py -m noise_scale -n 3 -i path/to/image/or/directory -a 0 -g 0

Train your own model

Install Wand

sudo apt install libmagickwand-dev
pip install wand

For more details, please refer template training script at <a href="https://github.com/tsurumeso/waifu2x-chainer/tree/master/appendix/linux">appendix/linux</a> or <a href="https://github.com/tsurumeso/waifu2x-chainer/tree/master/appendix/windows">appendix/windows</a> . In my case, 5000 JPEG images are used for pretraining and 1000 noise-free-PNG images for finetuning.

Convert Chainer models to ONNX and Caffe models

Install ONNX-Chainer

pip install onnx-chainer

Run script

cd appendix
python convert_models.py

Results are saved at the same directory of the original models (e.g. models/vgg7/anime_style_scale_rgb.npz to models/vgg7/anime_style_scale_rgb.caffemodel).
Note: Since chainer.CaffeFunction does not currently support Slice layer, some models are not converted to caffemodel.

References

• [1] nagadomi, "Image Super-Resolution for Anime-Style Art", https://github.com/nagadomi/waifu2x
• [2] "For Creators", http://piapro.net/en_for_creators.html