Generate icons using recurrent neural networks
Python
Latest commit 5ebeed7 Jan 17, 2017 @DOsinga Initial commit
Permalink
Failed to load latest commit information.
data/4x4 Initial commit Jan 17, 2017
.gitignore Initial commit Jan 17, 2017
LICENSE Initial commit Jan 17, 2017
README.md Initial commit Jan 17, 2017
draw_emoiji.py Initial commit Jan 17, 2017
icon_rnn.py Initial commit Jan 17, 2017
icon_rnn_test.py Initial commit Jan 17, 2017
requirements.txt Initial commit Jan 17, 2017

README.md

mandylion

Generate icons using recurrent neural networks.

icon_rnn.py expects a data directory containing one or more datasets. A dataset should contain a training subdirectory containing icons to train on. If you don't have one, you can run draw_emoiji.py which will render the emoiji's as bitmaps. That's only around 1000 images though, which is not very much.

Once you have a dataset in place, train with:

python icon_rnn.py --dataset=<your set> --mode=train

When done, you can create a poster of your result with

python icon_rnn.py --dataset=<your set> --mode=poster

Or you can pipe generated icons into an image classifier. You need to have inception5h downloaded in the incpetion5h directory for this to work. You can then run:

python icon_rnn.py --dataset=<your set> --mode=classify

The results of these steps are stored in the dataset directory.