Summary

We began this chapter by discussing the need for mobile neural networks to make CNNs work in real-time applications. We discussed the two benchmark MobileNet architectures that were introduced by Google—MobileNet and MobileNetV2. We looked at how modifications such as depth-wise separable convolution work and replaced the standard convolutions, enabling the network to achieve the same results with significantly fewer parameters. With MobileNetV2, we looked at the possibility of reducing the network even further with expansion layers and bottleneck layers. We also looked at the implementation of both the networks in Keras and compared both the networks in terms of the number of parameters, MACs, and memory required. Finally, we discussed the successful combination of MobileNets with object detection networks, such as SSD, to achieve object detection on mobile devices...