2.1.1. MobileNet V2 differences between Caffe and TensorFlow models
There are two inverted bottlenecks (group of expand, depthwise, projection) in which TensorFlow has already gone down to 14x14 while Caffe is still at 28x28. This is the only place where the structure of the graph differs. TensorFlow also uses ReLU6, implemented with a clamp in the Intel® FPGA AI Suite IP, while Caffe uses ReLU.