Attempt to use neural networks for classifying socal actions of fruit flies from pose tracking data using the fly-vs-fly dataset. I used neon for training the neural networks.

First download the fly-vs-fly dataset and install Neon. Model setup and training parameters are set in train.py. The dataset classes for training a single class and multiclass model are in flyvfly.py and flyvflymulticlass.py respectively. Specify the path of the dataset in each of these files. To train a single class model, import flyvfly as as Fly and set NUM_CLASSES constant to 1. Set the action number in flyvfly.py. To train a multiclass model, import flyvflymulticlass as Fly and set the NUM_CLASSES constant to 5. Run train.py with the output save file for the model as the first input argument.

Default videos for training are 1-5 and for testing are 6-10. Testing a single class model is done in test.py and testing a multiclass model is done in test_multiclass.py where the input argument to each is the name of the saved model file. The saved model file should end in a two digit number (00 to 99) for labeling the plots though this is not necessary.