Home
Download
News
Mastodon
Bluesky
X
Summary - 18 Years
Summary - 17 Years
Summary - 16 Years
Summary - 15 Years
Summary - 13 Years
Summary - 11 Years
Summary - 10 Years
Resources
Technical Reference
Scripting Tutorial
Video Tutorials
Wiki Pages
Image Gallery
Color Presets
Using libgmic
G'MIC Online
Community
Discussion Forum (Pixls.us)
GimpChat
IRC
Report Issue| Table of Contents ▸ List of Commands ▸ Neural Networks ▸ nn_normalize | ◀ nn_norm | nn_parameters ▶ |
nn_normalize
Arguments:
- out,in,axes,_normalization_type,_group_x>=0,_group_y>=0,_group_z>=0,_group_c>=0,_learning_mode
Description:
Add a normalize layer to the current network.This normalization layer is generic and can simulate the behavior of a LayerNorm, an InstanceNorm, a GroupNorm, or an RMSNorm.
axes defines the shape used to compute the mean/std of the normalization.
axes can be { x | y | z | c | xy | xz | xc | yz | yc | zc | xyz | xyc | xzc | yzc | xyzc }.
For instance:
axes=c means the vector-valued pixels of the input image are normalized independently, according to the channel values.
axes=xyz means each input channel of the input image is normalized independently, according to all (x,y,z) pixel values.
normalization_type can be { 0:Avg_std | 1:RMS | 2:Affine | 3:Linear } :
0: Normalization by mean/std, followed by the affine transform gamma*x + beta.
1: RMS normalization, followed by the linear transform gamma*x.
2: No normalization, only the affine transform gamma*x + beta is applied.
3: No normalization, only the linear transform gamma*x is applied.
For types 2 and 3, axes (and group_*) only define along which dimensions the gamma (and beta) parameters are shared.When different from 0, a group_* argument forces the corresponding complementary axis to be split into N groups. N must divide the corresponding input dimension, and a group_* argument cannot be set for an axis that is already listed in axes.
learning_mode can be { 0:No learning | 1:Gamma only | 2:Beta only | 3:Gamma+beta }
Default values:
normalization_type=0, group_x=group_y=group_z=group_c=0 and learning_mode=3.