Size In Covolutional Layer: l
\text{Size In Covolutional Layer: l}\\
Size In Covolutional Layer: l
F
i
l
t
e
r
S
i
z
e
:
f
[
l
]
Width and hight of a filter
P
a
d
d
i
n
g
:
p
[
l
]
The extra edge added around
S
t
r
i
d
e
:
s
[
l
]
The step for filter to move
C
h
a
n
n
e
l
:
n
c
[
l
]
Equal to the numbers of filters
F
i
l
t
e
r
:
f
[
l
]
×
f
[
l
]
×
n
c
[
l
−
1
]
I
n
p
u
t
:
n
h
[
l
−
1
]
×
n
w
[
l
−
1
]
×
n
c
[
l
−
1
]
e.g.An image shaped(32,32,3)
O
u
t
p
u
t
:
n
h
[
l
]
×
n
w
[
l
]
×
n
c
[
l
]
n
n
/
w
[
l
]
=
⌊
n
n
/
w
[
l
−
1
]
+
2
p
[
l
]
−
f
[
l
]
s
[
l
]
+
1
⌋
N
o
t
e
1
:
Deeper,
n
h
/
w
[
l
]
↑
while
n
c
[
l
]
↓
N
o
t
e
2
:
I
n
p
u
t
→
C
o
n
v
→
A
d
d
B
i
a
s
→
R
e
L
U
→
O
u
t
p
u
t
\begin{aligned} && Filter\ Size&: & f^{[l]} &\quad \text{Width and hight of a filter}\\ && Padding &: &p^{[l]} &\quad \text{The extra edge added around}\\ && Stride & : &s^{[l]} &\quad \text{The step for filter to move}\\ && Channel &: &n_c^{[l]}&\quad \text{Equal to the numbers of filters}\\ && Filter&: & f^{[l]} \times f^{[l]} \times n_c^{[l-1]}\\ && Input&: & n_h^{[l-1]} \times n_w^{[l-1]} \times n_c^{[l-1]} &\quad \text{e.g.An image shaped(32,32,3)}\\ && Output&: & n_h^{[l]} \times n_w^{[l]} \times n_c^{[l]}\\ && n_{n/w}^{[l]} &= & \biggl \lfloor\frac{n_{n/w}^{[l-1]} + 2p^{[l]}-f^{[l]}}{s^{[l]}}+1 \biggr \rfloor\\ && Note1 & : & \text{Deeper, } n_{h/w}^{[l]} \uparrow \text{while } n_c^{[l]} \downarrow\\ && Note2 & : &Input \rightarrow Conv \rightarrow Add\ Bias & \rightarrow ReLU \rightarrow Output \end{aligned}
Filter SizePaddingStrideChannelFilterInputOutputnn/w[l]Note1Note2:::::::=::f[l]p[l]s[l]nc[l]f[l]×f[l]×nc[l−1]nh[l−1]×nw[l−1]×nc[l−1]nh[l]×nw[l]×nc[l]⌊s[l]nn/w[l−1]+2p[l]−f[l]+1⌋Deeper, nh/w[l]↑while nc[l]↓Input→Conv→Add BiasWidth and hight of a filterThe extra edge added aroundThe step for filter to moveEqual to the numbers of filterse.g.An image shaped(32,32,3)→ReLU→Output
【助记】CNN的维度变化
最新推荐文章于 2024-05-15 16:00:00 发布