This is the Google Colab implementation of tsurumeso's chainer implementation of waifu2x, for people like me that do not have access to NVIDIA GPUs. Google Colab has CUDA version 10.00, so it is neccessary that the CuPy installed is of the same version.

The following is taken from tsurumeso's documentation -

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

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

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

pip install chainer
pip install pillow

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

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

cd waifu2x-chainer
python waifu2x.py

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

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

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

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

sudo apt install libmagickwand-dev
pip install wand

For more details, please refer template training script at appendix/linux or appendix/windows . In my case, 5000 JPEG images are used for pretraining and 1000 noise-free-PNG images for finetuning.

pip install onnx-chainer

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 skip conversion to caffemodel.

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