Some investigations on robustness of deep learning in limited angle tomography