Berü is a set of tools for training and detecting gestures made on the chassis of a laptop. It accomplishes this by using a Time Delay Neural Network (TDNN) trained on a large set of data.

This project used Anaconda3 to manage dependencies. It is recommended that you use it. If you choose not to, you will need to manually install numpy and scipy.

You will also need the pyaudio package. Before you can install it however, you will need portaudio installed. To do this, run

brew install portaudio

sudo apt-get install portaudio19-dev

Then, run pip install pyaudio. As of writing, Anaconda does not have the latest version of pyaudio, so it is recommended that you install via pip.

To acquire more testing or training samples, run python sample. You will then be prompted to add testing or training data, the name of samples, the number to record, and their duration.

To configure the Neural Network, edit neural-net/load_data.py. The parameters that you will need to change are

• NUM_FQS => The total number of samples
• NUM_TIME => The number of time buckets
• VERSION => Either sample.FREQUENCY or sample.AUTOCORRELATION depending on which method you want to use

To create a new neural net to train on, run python neural-net init <name> <layer1> [...<layerN>] where name can be anything, each layer should be of the form <NODES>,<TIME_OFFSETS>. Furthermore, layer1 should always have NUM_FREQ nodes, and NUM_TIME time offsets, and layern should have nodes equal to the number of samples, and one time offset.

To train the newly created neural network, run python neural-net init <input_name> <output_name> where <input_name> is the name specified above, and <output_name> is the file to store the resulting weights. You can also specifiy the -t flag to also test the net, -i N to perform testing every N rounds, -l to continue training instead of stopping after N rounds, and -r RATE to set the learning rate. This will display a graphical interface that shows the progress that's being made, with a breakdown of wherein error lies. You can hit ctrl-c to save and exit at any time.

Finally, to test the trained data against a live input stream, run python neural-net classify <weight_file> <FRQ|COR> where weight_file is the output file from training, and they type is FRQ if the net was trained with sample.FREQUENCY and COR if it was trained with sample.AUTOCORRELATION.

This was designed and implemented by Matthew Savage (@bluepichu) and Zachary Wade (@zwade).