MDMA-style-transfer

Using https://github.com/jfilter/universal-style-transfer-pytorch to transform images.

The main repo is https://github.com/jfilter/MDMA.

Usage

We use a fresh Ubuntu 16.04. installation and use root for everything. Feel free to change to a proper user.

1. Install pipenv
2. git clone https://github.com/jfilter/universal-style-transfer-pytorch
3. cd universal-style-transfer-pytorch && pipenv install && ./setup.sh && cd ..
4. git clone https://github.com/jfilter/MDMA-style-transfer && cd MDMA-style-transfer
5. Create a secrets.py for with the basic auth credentials to connect to MDMA:

username = 'XXX'
password = 'XXX'

5. Create a script do.sh that does the actual transfer:

cd /root/MDMA-style-transfer && /usr/local/bin/pipenv run python run.py >> /root/log.txt 2>&1

6. To make sure the bash script only runs once at the same time: apt-get install run-one
7. add to your crontab (crontab -e) to run e.g. every minute

* * * * * run-one /root/do.sh

Caveats

Right now, you can only transform one image in a batch. This is because the style weight has to be set for a batch. We allow to set the style weight individual for each image. Thus, we process only one.

License

MIT.