Play Rock Paper Scissors Lizard Spock against your browser.
By using a convolutional neural network, this demo can regonize which move is played. The neural network runs right in a browser, powered by deeplearn.js.
- deeplearn.js to run a neural net in a browser.
- TypeScript with React for the UI.
- Keras with keras-squeezenet to train the model.
yarn start
yarn train
This allows you to take images for training and save them in a folder per category.
If you do this multiple times (with page reloads) there will be conflicting names. To solve this use model-training/merge-dirs.py.
E.g. if you have the folders train1 and train2 each containing a folder per category, you can do:
python merge-dirs.py train2 train1
This will move all files from train2 to train1 and resolve naming conflicts.
After this you can delete train2 (which by now should not contain any files):
rm -rf train2
The jupyter notebook model-training/rock-paper-scissors-lizard-spock.ipynb walks you through the process of training the neural net using keras. It also shows how to save the model for use with deeplearn.js.
The resulting weights can be downloaded here.
To run the notebook you need:
- Python 3
- numpy
- Jupyter
- Keras
- keras-squeezenet
- Tensorflow
MIT