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