In this paper, we propose intelligent reconfigurable surface (IRS)-assisted unmanned aerial vehicles (UAVs) networks that can utilise both advantages of agility and reflection for enhancing the network's performance. To aim at maximising the energy efficiency (EE) of the considered networks, we jointly optimise the power allocation of the UAVs and the phaseshift matrix of the IRS. A deep reinforcement learning (DRL) approach is proposed for solving the continuous optimisation problem with time-varying channel gain in a centralised fashion. Moreover, a parallel learning approach is also proposed for reducing the information transmission requirement of the centralised approach. Numerical results show a significant improvement of our proposed schemes compared with the conventional approaches in terms of EE, flexibility, and processing time. Our proposed DRL methods for IRS-assisted UAV networks can be used for real-time applications due to their capability of instant decision-making and handling the time-varying channel with the dynamic environmental setting.