Finding the mounting angle of a smartphone inside a car is crucial for navigation, motion detection, activity recognition, and other applications. It is a challenging task in several aspects: (i) the mounting angle at the drive start is unknown and may differ significantly between users; (ii) the user, or bad fixture, may change the mounting angle while driving; (iii) a rapid and computationally efficient real-time solution is required for most applications. To tackle these problems, a data-driven approach using deep neural networks (DNNs) is presented to learn the yaw mounting angle of a smartphone equipped with an inertial measurement unit (IMU) and strapped to a car. The proposed model, MountNet, uses only IMU readings as input and, in contrast to existing solutions, does not require inputs from global navigation satellite systems (GNSS). IMU data is collected for training and validation with the sensor mounted at a known yaw mounting angle and a range of ground truth labels is generated by applying a prescribed rotation to the measurements. Although the training data did not include recordings with real sensor rotations, tests on data with real and synthetic rotations show similar results. An algorithm is formulated for real-time deployment to detect and smooth transitions in device mounting angle estimated by MountNet. MountNet is shown to find the mounting angle rapidly which is critical in real-time applications. Our method converges in less than 30 seconds of driving to a mean error of 4 degrees allowing a fast calibration phase for other algorithms and applications. When the device is rotated in the middle of a drive, large changes converge in 5 seconds and small changes converge in less than 30 seconds.