精品九九热免费,欧美久久久大香蕉,精品av国

Chapter 1. Introduction

1.1 Discrete convolutions

$N$ : N-D
$n$ : number of output feature maps
$m$ : number of input feature maps
$k_j$ : kernel size along axis j
$i_j$ : input size along axis j
$s_j$ : stride (distance between two consecutive positions of the kernel) along axis j
$p_j$ : zero padding (number of zeros concatenated at the beginning and at the end of an axis) along axis j

1.2 Pooling

Pooling operations reduce the size of feature maps by using some function to summarize subregions, such as taking the average or the maximum value.

Chapter 2. Convolution arithmetic

The analysis of the relationship between convolutional layer properties is eased by the fact that they don’t interact across axes. Because of that, this chapter will focus on the following simplified setting:

2-D discrete convolutions ( $N = 2$ )
square inputs ( $i_1 = i_2 = i$ ),
square kernel size ( $k_1 = k_2 = k$ ),
same strides along both axes ( $s_1 = s_2 = s$ ),
same zero padding along both axes ( $p_1 = p_2 = p$ ).

Note: the results outlined here also generalize to the $N$ -D and non-square cases.

2.1 No zero padding, unit strides ( $p=0, s=1$ )

Relationship 1. For any $i$ and $k$ , and for $s = 1$ and $p = 0$ , $o = (i - k) + 1$ .

2.2 Zero padding, unit strides ( $p>0, s=1$ )

Relationship 2. For any $i$ , $k$ and $p$ , and for $s = 1$ ,
$o = (i - k) + 2p + 1$ .

2.2.1 Half (same) padding ( $p=\frac{k-1}{2}$ )

Relationship 3. For any $i$ and for $k$ odd ( $k = 2n + 1$ , $n\in N$ ), $s = 1$ and $p = n$ , $o= (i+2p)-k+1=(i+2n)-(2n+1)+1=i$

2.2.2 Full padding ( $p=k-1$ )

Relationship 4. For any $i$ and $k$ , and for $p = k - 1$ and $s = 1$ , $o = i + 2(k - 1) - (k - 1)= i + (k - 1)$ .

2.3 No zero padding, non-unit strides ( $p=0, s>1$ )

Relationship 5. For any $i$ , $k$ and $s$ , and for $p = 0$ ,
$o =\lfloor\frac{i - k}{s}\rfloor+1$ .

2.4 Zero padding, non-unit strides ( $p>0, s>1$ )

Relationship 6. For any $i$ , $k$ , $p$ and $s$ ,
$o =\lfloor\frac{i+2p-k}{s}\rfloor+1$ .

Chapter 3. Pooling arithmetic

Pooling does not involve zero padding ( $p=0$ ).

Relationship 7. For any $i$ , $k$ and $s$ ,
$o =\lfloor\frac{i - k}{s}\rfloor+1$ .

Chapter 4. Transposed convolution arithmetic

The need for transposed convolutions generally arises from the desire to use a transformation going in the opposite direction of a normal convolution.
Note: transposed convolution properties don’t interact across axes
We still use the same settings as chapter 2 in the following.

4.1 Convolution as a matrix operation

4.2 Transposed convolution

4.3 No zero padding, unit strides, transposed ( $p=0, s=1, C^T$ )

Relationship 8. A convolution described by $s = 1$ , $p = 0$ and $k$ has an associated transposed convolution described by $k' = k$ , $s' = s$ and $p' = k - 1$ and its output size is $o' =i' + (k - 1)$ :

$i\xrightarrow[]{\quad k, s=1, p=0\quad} o=i-k+1$
$i'=o=i-k+1\xrightarrow[]{\; k'=k, s'=s, p'=k-1\;} o'=i'+2p'-k'+1=i$

4.4 Zero padding, unit strides, transposed ( $p>0, s=1, C^T$ )

Relationship 9. A convolution described by $s = 1$ , $k$ and $p$ has an associated transposed convolution described by $k' = k$ , $s' = s$ and $p' = k - p - 1$ and its output size is
$o' = i + (k - 1) - 2p$

$i\xrightarrow[]{\quad k, s=1, p\quad} o=i+2p-k+1$
$i'=o\xrightarrow[]{\; k'=k, s'=s, p'\;} o'=i'+2p'-k'+1=i\implies p'=k-p-1$

4.4.1 Half (same) padding, transposed ( $p=\frac{k-1}{2}, C^T$ )

Relationship 10. A convolution described by $k = 2n+1,n\in N$ , $s = 1$ and $p = n$ has an associated transposed convolution described by $k'= k$ , $s'= s$ and $p' = k-p-1=(2p+1)-p-1=p$ and its output size is $o'=i'$ .

4.4.2 Full padding, transposed ( $p=k-1, C^T$ )

Relationship 11. A convolution described by $s = 1$ , $k$ and $p =k-1$ has an associated transposed convolution described by $k' = k$ , $s' = s$ and $p' = k-p-1=0$ and its output size is $o'=i'-(k-1)$ .

4.5 No zero padding, non-unit strides, transposed ( $p=0, s>1, C^T$ )

Relationship 12. A convolution described by $p=0$ , $k$ and $s$ and whose input size is such that $i-k$ is a multiple of $s$ , has an associated transposed convolution described by $\tilde{i'}$ , $k' = k$ , $s' = 1$ and $p' = k-1$ , where $\tilde{i'}$ is the size of the stretched input obtained by adding $s-1$ zeros between each input unit, and its output size is $o'= s(i'-1)+k$ .

$i\xrightarrow[]{\quad k, s, p=0\quad} o=\lfloor\frac{i-k}{s}\rfloor+1$
$\tilde{i'}=i'+(s-1)(i'-1)=s(i'-1)+1\xrightarrow[]{\; k'=k, s'=1, p'=k-1\;} o'=\lfloor\frac{\tilde{i'}+2p'-k'}{s'}\rfloor+1=s(i'-1)+k$

4.6 Zero padding, non-unit strides, transposed ( $p>0, s>1, C^T$ )

Relationship 13. A convolution described by $p$ , $k$ and $s$ and whose input size is such that $i+2p-k$ is a multiple of $s$ , has an associated transposed convolution described by $\tilde{i'}$ , $k' = k$ , $s' = 1$ and $p' = k-p-1$ , where $\tilde{i'}$ is the size of the stretched input obtained by adding $s-1$ zeros between each input unit, and its output size is $o'= s(i'-1)+k-2p$ .

$i\xrightarrow[]{\quad k, s, p\quad} o=\lfloor\frac{i+2p-k}{s}\rfloor+1$
$\tilde{i'}=i'+(s-1)(i'-1)=s(i'-1)+1\xrightarrow[]{\; k'=k, s'=1, p'=k-p-1\;} o'=\lfloor\frac{\tilde{i'}+2p'-k'}{s'}\rfloor+1=s(i'-1)+k-2p$

Relationship 14.A convolution described by $p$ , $k$ and $s$ has an associated transposed convolution described by $a$ , $\tilde{i'}$ , $k' = k$ , $s' = 1$ and $p' = k-p-1$ , where $\tilde{i'}$ is the size of the stretched input obtained by adding $s-1$ zeros between each input unit, and $a=i+2p-k$ mod $s$ represents the number of zeros added to the bottom and right
edges of the input, its output size is $o'= s(i'-1)+a+k-2p$ .

Chapter 5. Miscellaneous convolutions

5.1 Dilated convolutions

Dilated convolutions are used to cheaply increase the receptive field of output units without increasing the kernel size, there are usually d-1 spaces inserted between kernel elements such that d = 1 corresponds to a regular convolution.
A kernel of size k dilated by a factor d has an effective size $\hat{k}=k+(k-1)(d-1)$

Relationship 15. For any $i$ , $k$ , $p$ and $s$ , and for a dilation rate $d$ , $o=\lfloor\frac{i+2p-\hat{k}}{s}\rfloor+1=\lfloor\frac{i+2p-k-(k-1)(d-1)}{s}\rfloor+1$ .

色偷偷精品伊人,欧洲久久精品,欧美综合婷婷骚逼,国产AV主播,国产最新探花在线,九色在线视频一区,伊人大交九欧美,1769亚洲,黄色成人av

[論文筆記]A guide to convolution arithmetic for deep learning

[論文筆記]A guide to convolution arithmetic for deep learning

Chapter 1. Introduction

1.1 Discrete convolutions

1.2 Pooling

Chapter 2. Convolution arithmetic

2.1 No zero padding, unit strides ( $p=0, s=1$ )

2.2 Zero padding, unit strides ( $p>0, s=1$ )

2.2.1 Half (same) padding ( $p=\frac{k-1}{2}$ )

2.2.2 Full padding ( $p=k-1$ )

2.3 No zero padding, non-unit strides ( $p=0, s>1$ )

2.4 Zero padding, non-unit strides ( $p>0, s>1$ )

Chapter 3. Pooling arithmetic

Chapter 4. Transposed convolution arithmetic

4.1 Convolution as a matrix operation

4.2 Transposed convolution

4.3 No zero padding, unit strides, transposed ( $p=0, s=1, C^T$ )

4.4 Zero padding, unit strides, transposed ( $p>0, s=1, C^T$ )

4.4.1 Half (same) padding, transposed ( $p=\frac{k-1}{2}, C^T$ )

4.4.2 Full padding, transposed ( $p=k-1, C^T$ )

4.5 No zero padding, non-unit strides, transposed ( $p=0, s>1, C^T$ )

4.6 Zero padding, non-unit strides, transposed ( $p>0, s>1, C^T$ )

Chapter 5. Miscellaneous convolutions

5.1 Dilated convolutions

相關(guān)閱讀更多精彩內(nèi)容

友情鏈接更多精彩內(nèi)容

色偷偷精品伊人,欧洲久久精品,欧美综合婷婷骚逼,国产AV主播,国产最新探花在线,九色在线视频一区,伊人大交九 欧美,1769亚洲,黄色成人av

[論文筆記]A guide to convolution arithmetic for deep learning

Chapter 1. Introduction

1.1 Discrete convolutions

1.2 Pooling

Chapter 2. Convolution arithmetic

2.1 No zero padding, unit strides ()

2.2 Zero padding, unit strides ()

2.2.1 Half (same) padding ()

2.2.2 Full padding ()

2.3 No zero padding, non-unit strides ()

2.4 Zero padding, non-unit strides ()

Chapter 3. Pooling arithmetic

Chapter 4. Transposed convolution arithmetic

4.1 Convolution as a matrix operation

4.2 Transposed convolution

4.3 No zero padding, unit strides, transposed ()

4.4 Zero padding, unit strides, transposed ()

4.4.1 Half (same) padding, transposed ()

4.4.2 Full padding, transposed ()

4.5 No zero padding, non-unit strides, transposed ()

4.6 Zero padding, non-unit strides, transposed ()

Chapter 5. Miscellaneous convolutions

5.1 Dilated convolutions

相關(guān)閱讀更多精彩內(nèi)容

友情鏈接更多精彩內(nèi)容

色偷偷精品伊人,欧洲久久精品,欧美综合婷婷骚逼,国产AV主播,国产最新探花在线,九色在线视频一区,伊人大交九欧美,1769亚洲,黄色成人av

2.1 No zero padding, unit strides ( $p=0, s=1$ )

2.2 Zero padding, unit strides ( $p>0, s=1$ )

2.2.1 Half (same) padding ( $p=\frac{k-1}{2}$ )

2.2.2 Full padding ( $p=k-1$ )

2.3 No zero padding, non-unit strides ( $p=0, s>1$ )

2.4 Zero padding, non-unit strides ( $p>0, s>1$ )

4.3 No zero padding, unit strides, transposed ( $p=0, s=1, C^T$ )

4.4 Zero padding, unit strides, transposed ( $p>0, s=1, C^T$ )

4.4.1 Half (same) padding, transposed ( $p=\frac{k-1}{2}, C^T$ )

4.4.2 Full padding, transposed ( $p=k-1, C^T$ )

4.5 No zero padding, non-unit strides, transposed ( $p=0, s>1, C^T$ )

4.6 Zero padding, non-unit strides, transposed ( $p>0, s>1, C^T$ )