Using deep learning architectures to recognize human emotions from electroencephalography (EEG) signals, thereby deploying practical applications, is a problem that is currently of great interest to many scientists. Depending on the application, the EEG signal can be received from devices with many electrodes (32, 64) or small devices (14, 2). For each application, to ensure good recognition efficiency, determining the appropriate deep learning architecture as well as the corresponding input features is an important issue that needs to be solved. In this paper, based on the DEAP emotion database, we evaluate the performance of several different deep learning architectures, including CNN, BiLSTM, and a combination of CNN and BiLSTM, with different input features such as FFT and Welch extracted from 32-electrode to 14-electrode EEG signals. The experimental results show that the CNN network with FFT features has the best performance with the highest average recognition accuracy and the lowest average loss value. However, the CNN network architecture with Welch features gives the best stability when switching from 32 electrodes to 14 electrodes.