The parameters used for batch normalization are not specified in the original paper.
I don't know how batch norm works in Matlab (I have tried to read the codes but it's very difficult to me), so I tried looking in keras. The parameters used seemed very odd (in addition the batch normalisation is involved in a code mess), so I looked in pytorch, and saw that they are different.
In keras the momentum is 0.0 or 0.1 depending on where you look, and in pytorch the momentum is 0.95.
The parameters used for batch normalization are not specified in the original paper.
I don't know how batch norm works in Matlab (I have tried to read the codes but it's very difficult to me), so I tried looking in keras. The parameters used seemed very odd (in addition the batch normalisation is involved in a code mess), so I looked in pytorch, and saw that they are different. In keras the momentum is
0.0
or0.1
depending on where you look, and in pytorch the momentum is0.95
.Note that this differs from the original unofficial keras implementation.