Abstract: Intelligent reflecting surface (IRS) aided communication systems have high channel dimensionality and the existing channel estimation methods require a lot of pilots to obtain an accurate channel matrix. To address this problem, a hybrid loss based residual generative adversarial network (H-ResGAN) model is proposed. H-ResGAN uses multiple residual blocks to deepen the network, which can fully extract channel features and mitigate the gradient disappearance problem. At the same time, a hybrid loss combining least squares loss and L1 loss is adopted as the objective function to improve the stability of the training. Simulation experiments demonstrate that H-ResGAN is more robust to environmental noise and has significantly lower estimation errors than traditional methods. In addition, H-ResGAN can obtain accurate estimation results by sending only few pilots compared to traditional estimation algorithms.