For inference the batch normalization layers can be merged into convolutional kernels, to
speed up the network. Both layers applies a linear transformation. For that reason
the batch normalization layers can be absorbed in the previous convolutional layer
by modifying its weights and biases. That is exactly what the script does. In doing so it is possible to speed up SegNet by around 30 %.
alexgkendall
commented
7 years ago
For inference the batch normalization layers can be merged into convolutional kernels, to speed up the network. Both layers applies a linear transformation. For that reason the batch normalization layers can be absorbed in the previous convolutional layer by modifying its weights and biases. That is exactly what the script does. In doing so it is possible to speed up SegNet by around 30 %.