Group_Normalization_pytorch

• Implementation of Group Normalization.
• Do the Experiments on MNIST.

Statement

• Class GroupNorm1D(in_channels, G, channels_per_group, eps=1e-5) for 1D features.
• Class GroupNorm2D, GroupNorm3D for 2D and 3D features.
• If you want to group normalization to process higher dimension features, you can Class GroupNormND(ND, in_channels, G, channels_per_group, eps=1e-5). For example, **GroupNormND(4, ...) for 4D features.
• Parameter G means group number.
• Parameter channels_per_group means channel number in each group.
• Only can assign a integer to one parameter and assign None to another.
• You can find the the code of group normalization in lib/group_normalization.py.
• If there is something wrong in my code, please contact me, thanks!

Environment

• python 3.6
• pytorch 0.4.0

Visualization of Loss

• Train loss of batch size 128.

• Test loss of batch size 128.

• Test loss of batch size 2.