I found that both in solver.py and joint_solver.py, when doing buid_model(), you just set self.net.eval(), but when we do train(), there is no command to set self.net.train(), which is commented in build_model().