In this paper, we propose a Bayesian channel estimator for intelligent reflecting surface-aided (IRS-aided) millimeter wave (mmWave) massive multiple-input multiple-output (MIMO) systems with semi-passive elements that can receive the signal in the active sensing mode. Ultimately, our goal is to minimize the channel estimation error using the received signal at the base station and additional information acquired from a small number of active sensors at the IRS. Unlike recent works on channel estimation with semi-passive elements that require both uplink and downlink training signals to estimate the UE-IRS and IRS-BS links, we only use uplink training signals to estimate all the links. To compute the minimum mean squared error (MMSE) estimates of all the links, we propose a novel variational inference-sparse Bayesian learning (VI-SBL) channel estimator that performs approximate posterior inference on the channel using VI with the mean-field approximation under the SBL framework. The simulation results show that VI-SBL outperforms the state-of-the-art baselines for IRS with passive reflecting elements in terms of the channel estimation accuracy, training overhead, and spectral efficiency. Furthermore, VI-SBL with semi-passive elements is shown to be more energy-efficient than the baselines with passive reflecting elements while employing a small number of low-cost active sensors.