 We compared five motor imagery, MI, based EEG deep learning models on two large data sets with 42 and 62 subjects respectively. The models were EEG net, shallow and deep CONV net, MB3D and PAR-ATT. We found that EEG net performed the best on the second data set with a relatively low training cost. This suggests that EEG net may be a good choice for future MI-based BCI applications. This article was authored by Hal Zhu, Dylan Ferenzo and Ben He.