Unmanned Aerial Vehicles (UAV) can provide mobile edge computing (MEC) service for resource-limited devices in Internet of Things (IoT). In such scenario, partial offloading can be used to balance the computing task between the UAV and the IoT devices for higher efficiency. However, traditional partial offloading is not suitable for training deep neural network (DNN), since DNN models cannot be portioned with a continuous ratio. In this paper, we introduce a split offloading scheme, which can flexibly split the DNN training task into two parts based on the DNN layers, and allocate them to the IoT device and UAV respectively. We present a scheme to synchronize the training and communicating period of DNN layers in the UAV and IoT device, and thus reduce the model training time. Based on this scheme, an optimization model is proposed to minimize the UAV energy consumption, which jointly optimizes the UAV trajectory, the DNN split position and the service time scheduling. We divide the problem into two subproblems and solve it with an iterative solution. Simulation results show the proposed scheme can reduce the model training time and the UAV energy consumption by up to 25% and 14.4% compared with benchmark schemes, respectively.