This implementation of Batch Normalization is based on MVNLayer in Caffe.

To add this layer, you have to modify common_layers.hpp, layer_factory.cpp and caffe.proto. See Caffe wiki

NOTE

This implementation is very basic which just performs batch normalization computation. Two pieces in the paper are still missing here:

• fixed mean & variance for inference
• per batch shuffling for thorough randomness

Update log

2016/01/18 An improved version is provided here with moving average statistics for inference.