<p>The complexity of multi-dimensional channels makes accurate channel estimation difficult in intelligent reflecting surface (IRS) assisted mmWave MIMO systems. To improve the estimation accuracy, this paper presents a channel estimation algorithm based on a depthwise separable residual shrinkage network (DSRSN), named as DSRSN algorithm. Firstly, the channel matrix estimated by the least square method and the real channel matrix are given as the training input and the training label, separately. Then, a depthwise separable residual shrinkage frame is constructed to improve the accuracy of channel matrix estimation. Specially, to decrease the network parameters number, depthwise separable convolution is introduced. Next, the input features are processed by a soft threshold module to suppress channel noise and extract channel characteristics more accurately. Finally, the well-trained DSRSN model can exactly output the estimated channel matrix. According to the simulation results, the suggested DSRSN approach increases the estimation accuracy further while requiring less pilot overhead. For example, the simulation results verify that the normalized mean square error reduces by more than 2.94 times in comparision to the related methods, when the signal-to-noise ratio is 15dB.</p>