<p>Intelligent reflecting surface (IRS) improve coverage and throughput in millimeter-wave (mmWave) systems. However, existing channel estimation schemes require a large number of pilots to accurately obtain the channel matrix, especially in the case of high-dimensional cascade channels and passive reflecting elements of IRS, which challenges the efficiency of the system. To address this problem, this paper proposes a channel estimation method based on residual attention conditional generative adversarial networks (RA-CGAN). The scheme adopts a hybrid active and passive IRS architecture and uses the simultaneous orthogonal matching pursuit (SOMP) algorithm to obtain preliminary channel estimates. After preprocessing, the preliminary channel estimate is used as the input of the neural network, and the RA-CGAN model is used to perform precise channel estimation to improve the accuracy of the estimation. The generator of the RA-CGAN model takes the residual attention module as the main body. The residual attention module combines the advantages of residual connection and attention mechanism, which can remove noise interference, fully extract channel features, and slow down the gradient disappearance problem. The simulation results show that the RA-CGAN model has a small number of pilots. Its normalized mean squared error (NMSE) is superior to least square (LS), SOMP, spatial-frequency convolution neural network (SFCNN), and semi-super-resolution generative adversarial network (SSRGAN) algorithms. The feasibility of the scheme is proved.</p>