【AI基础】图解手算BatchNorm、LayerNorm和GroupNorm
作者：互联网
这几天整理对比了一下网络中几个常用的Norm的方法，之前也看过，网上有很多讲的非常详细的资料，以前看一下理解了就过了，时间长了就模糊了，此次自己亲手算了一遍，加深了印象，特此整理一下，以便之后的回顾。
设置一个Tensor，其Size为[3,4,2,2]，便于之后的理解
一、BatchNorm
BatchNorm详解
所有Norm方法无非都是减均值再除以标准差，无非是在哪个尺度上进行该操作的差异，而BatchNorm是在一个batch上，同一个通道上面进行Norm，那么有多少个通道就会计算多少个均值和标准差。
Mean=所有batch同一个通道上的均值
Std=所有batch同一个通道上的标准差
图中红色虚线框为Norm的尺度，黄色矩形框中为更新后的值
n
n
.
B
a
t
c
h
N
o
r
m
2
d
(
n
u
m
_
f
e
a
t
u
r
e
s
=
4
)
{nn.BatchNorm2d(num\_features=4)}
nn.BatchNorm2d(num_features=4)其中
n
u
m
_
f
e
a
t
u
r
e
s
{num\_features}
num_features为Channel数
代码验证：
import torch
import torch.nn as nn
a = torch.tensor([[[[1.,1],[1,1]],[[0,1],[1,0]],[[0,0],[0,1]],[[1,1],[0,0]]],
[[[2.,2],[0,0]],[[2,0],[1,1]],[[1,0],[0,2]],[[2,1],[1,0]]],
[[[3.,1],[2,2]],[[3,0],[0,2]],[[2,3],[1,2]],[[3,3],[2,1]]]])
batch = nn.BatchNorm2d(num_features=4)
b = batch(a)
输出结果与手算一致：
tensor([[[[-0.3922, -0.3922],
[-0.3922, -0.3922]],
[[-0.9611, 0.0874],
[ 0.0874, -0.9611]],
[[-1.0000, -1.0000],
[-1.0000, 0.0000]],
[[-0.2474, -0.2474],
[-1.2372, -1.2372]]],
[[[ 0.7845, 0.7845],
[-1.5689, -1.5689]],
[[ 1.1358, -0.9611],
[ 0.0874, 0.0874]],
[[ 0.0000, -1.0000],
[-1.0000, 1.0000]],
[[ 0.7423, -0.2474],
[-0.2474, -1.2372]]],
[[[ 1.9611, -0.3922],
[ 0.7845, 0.7845]],
[[ 2.1842, -0.9611],
[-0.9611, 1.1358]],
[[ 1.0000, 2.0000],
[ 0.0000, 1.0000]],
[[ 1.7320, 1.7320],
[ 0.7423, -0.2474]]]], grad_fn=<NativeBatchNormBackward>)
二、LayerNorm
LayerNorm详解
LayerNorm可以在3种不同的尺度进行
第一种：
n
n
.
L
a
y
e
r
N
o
r
m
(
n
o
r
m
a
l
i
z
e
d
_
s
h
a
p
e
=
[
4
,
2
,
2
]
)
{nn.LayerNorm(normalized\_shape=[4,2,2])}
nn.LayerNorm(normalized_shape=[4,2,2])
layer1 = nn.LayerNorm(normalized_shape=[4,2,2])
c1 = layer1(a)
输出结果与手算一致：
tensor([[[[ 0.8819, 0.8819],
[ 0.8819, 0.8819]],
[[-1.1339, 0.8819],
[ 0.8819, -1.1339]],
[[-1.1339, -1.1339],
[-1.1339, 0.8819]],
[[ 0.8819, 0.8819],
[-1.1339, -1.1339]]],
[[[ 1.2851, 1.2851],
[-1.1339, -1.1339]],
[[ 1.2851, -1.1339],
[ 0.0756, 0.0756]],
[[ 0.0756, -1.1339],
[-1.1339, 1.2851]],
[[ 1.2851, 0.0756],
[ 0.0756, -1.1339]]],
[[[ 1.1339, -0.8819],
[ 0.1260, 0.1260]],
[[ 1.1339, -1.8898],
[-1.8898, 0.1260]],
[[ 0.1260, 1.1339],
[-0.8819, 0.1260]],
[[ 1.1339, 1.1339],
[ 0.1260, -0.8819]]]], grad_fn=<NativeLayerNormBackward>)
第二种：
n
n
.
L
a
y
e
r
N
o
r
m
(
n
o
r
m
a
l
i
z
e
d
_
s
h
a
p
e
=
[
2
,
2
]
)
{nn.LayerNorm(normalized\_shape=[2,2])}
nn.LayerNorm(normalized_shape=[2,2])
layer2 = nn.LayerNorm(normalized_shape=[2,2])
c2 = layer2(a)
输出结果与手算一致：
tensor([[[[ 0.0000, 0.0000],
[ 0.0000, 0.0000]],
[[-1.0000, 1.0000],
[ 1.0000, -1.0000]],
[[-0.5773, -0.5773],
[-0.5773, 1.7320]],
[[ 1.0000, 1.0000],
[-1.0000, -1.0000]]],
[[[ 1.0000, 1.0000],
[-1.0000, -1.0000]],
[[ 1.4142, -1.4142],
[ 0.0000, 0.0000]],
[[ 0.3015, -0.9045],
[-0.9045, 1.5075]],
[[ 1.4142, 0.0000],
[ 0.0000, -1.4142]]],
[[[ 1.4142, -1.4142],
[ 0.0000, 0.0000]],
[[ 1.3471, -0.9622],
[-0.9622, 0.5773]],
[[ 0.0000, 1.4142],
[-1.4142, 0.0000]],
[[ 0.9045, 0.9045],
[-0.3015, -1.5075]]]], grad_fn=<NativeLayerNormBackward>)
第三种：
n
n
.
L
a
y
e
r
N
o
r
m
(
n
o
r
m
a
l
i
z
e
d
_
s
h
a
p
e
=
2
)
{nn.LayerNorm(normalized\_shape=2)}
nn.LayerNorm(normalized_shape=2)
唔，太多了，不想算了…抽检吧
layer3 = nn.LayerNorm(normalized_shape=2)
c3 = layer3(a)
输出：
tensor([[[[ 0.0000, 0.0000],
[ 0.0000, 0.0000]],
[[-1.0000, 1.0000],
[ 1.0000, -1.0000]],
[[ 0.0000, 0.0000],
[-1.0000, 1.0000]],
[[ 0.0000, 0.0000],
[ 0.0000, 0.0000]]],
[[[ 0.0000, 0.0000],
[ 0.0000, 0.0000]],
[[ 1.0000, -1.0000],
[ 0.0000, 0.0000]],
[[ 1.0000, -1.0000],
[-1.0000, 1.0000]],
[[ 1.0000, -1.0000],
[ 1.0000, -1.0000]]],
[[[ 1.0000, -1.0000],
[ 0.0000, 0.0000]],
[[ 1.0000, -1.0000],
[-1.0000, 1.0000]],
[[-1.0000, 1.0000],
[-1.0000, 1.0000]],
[[ 0.0000, 0.0000],
[ 1.0000, -1.0000]]]], grad_fn=<NativeLayerNormBackward>)
三、GroupNorm
GroupNorm详解
n
n
.
G
r
o
u
p
N
o
r
m
(
n
u
m
_
g
r
o
u
p
s
=
2
,
n
u
m
_
c
h
a
n
n
e
l
s
=
4
)
)
{nn.GroupNorm(num\_groups=2, num\_channels=4))}
nn.GroupNorm(num_groups=2,num_channels=4))
其中
n
u
m
_
g
r
o
u
p
s
{num\_groups}
num_groups为分组数，
n
u
m
_
c
h
a
n
n
e
l
s
{num\_channels}
num_channels为通道数
group = nn.GroupNorm(num_groups=2, num_channels=4)
d = group(a)
输出结果与手算一致：
tensor([[[[ 0.5773, 0.5773],
[ 0.5773, 0.5773]],
[[-1.7320, 0.5773],
[ 0.5773, -1.7320]],
[[-0.7746, -0.7746],
[-0.7746, 1.2910]],
[[ 1.2910, 1.2910],
[-0.7746, -0.7746]]],
[[[ 1.1547, 1.1547],
[-1.1547, -1.1547]],
[[ 1.1547, -1.1547],
[ 0.0000, 0.0000]],
[[ 0.1601, -1.1209],
[-1.1209, 1.4411]],
[[ 1.4411, 0.1601],
[ 0.1601, -1.1209]]],
[[[ 1.2376, -0.5625],
[ 0.3375, 0.3375]],
[[ 1.2376, -1.4626],
[-1.4626, 0.3375]],
[[-0.1601, 1.1209],
[-1.4411, -0.1601]],
[[ 1.1209, 1.1209],
[-0.1601, -1.4411]]]], grad_fn=<NativeGroupNormBackward>)
注意：只有BatchNorm与Batch有关，LayerNorm和GroupNorm都不会在Batch上进行计算
标签：nn,AI,1.0000,1.1339,num,手算,0.0000,BatchNorm,LayerNorm 来源： https://blog.csdn.net/qq_43426908/article/details/123119919