This paper proposes a kind of deep learning (DL) based signal detection in dual mode generalized spatial modulation (DM-GSM) system, which aims to balance the detection performance and the complexity. Specifically, two neural networks, deep neural network (DNN) and convolutional neural network (CNN), are utilized to gain the mapping relationship among the received symbols, the channel matrix, and the transmitted bits. After offline training, the trained network is deployed for the online signal detection according to the input feature vector. Simulation results illustrate that the proposed DL detection can obtain the approximate performance of maximum likelihood (ML) detection at lower complexity and can provide better robustness compared with the conventional detection algorithms in the presence of various noise deviating from the standard Gaussian distribution.