To improve the accuracy of pulse wave recognition, MIRNet2 is proposed, which is a kind of modified deep hybrid neural networks. Firstly, processable data sets of Caffe are obtained by main pulse extraction, segmenting cycle and making hdf5 data sets. Secondly, deep hybrid neural networks are designed. Inception-ResNet (IRNet) is consisted of inception modules and residual modules, containing IRNet1, IRNet2 and IRNet3. Subsequently, Modified Inception-ResNet (MIRNet) composed of modified Inception modules, residual modules and pooling modules (or reduction modules) is proposed, including MIRNet1 and MIRNet2. Compared with other neural networks in the study, MIRNet2 is the best one, with the specificity of 87.85%, the sensitivity of 88.05% and the accuracy of 87.84%, respectively. In addition, parameters and operations of MIRNet2 are also less than that of IRNet3.