🖌️ Automatic Image Colorization

Implementation of Let there be Color! by Satoshi Iizuka, Edgar Simo-Serra and Hiroshi Ishikawa.

Colorized 100 years old photo of Książ Castle:

More images

Other colorized images:

First run

Places365-Standard dataset will be downloaded and split into train/dev/test subsets. By default only 10 arbitrary categories will be considered.

$ git clone https://github.com/kainoj/colnet.git
$ cd colnet
$ make dataset
$ make split

Requirements

Code is written in Python 3.6.3. Here are all requirements, hit the following to install:

pip3 install -r requirements.txt

Network training

Simply run:

$ python3 loader.py config/places10.yaml

places10.yaml is a sample configuration file – i.e. specifies total number of epoch, learning rate, output directories etc. To see full .yaml configuration, run python3 loader.py config/places10.yaml

Checkpoints of models are saved on every epoch. Training can be interrupted and resumed anytime. Resume by executing:

$ python3 loader.py config/places10.yaml --model model.pt

where model.pt is a previously saved model checkpoint.

Colorize!

Choose the most favourite model and hit:

$ python3 colorize.py img.jpg ./models/places.pt