We present open-source deep learning based classifiers for predicting physical therapy exercises using inertial measurement units, and a comprehensive analysis of impacts of sensor density, location, type, state estimation, and sample size on the performance.