https://openaccess.thecvf.com/content_cvpr_2017/papers/Chollet_Xception_Deep_Learning_CVPR_2017_paper.pdf

 

์•ˆ๋…•ํ•˜์„ธ์š”.

 

Xception ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ 2ํšŒ์ฐจ์ž…๋‹ˆ๋‹ค. 

 

1ํšŒ์ฐจ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ๋กœ An Extreme version of Inception module์ด Depthwise Separable Convolution๊นŒ์ง€ ์†Œ๊ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.

 

์•„๋ž˜ ์ „ ํŽธ์˜ ๊ธ€์„ ์ฝ๋Š” ๊ฒƒ์„ ์ถ”์ฒœ๋“œ๋ฆฝ๋‹ˆ๋‹ค.

 

https://coding-yoon.tistory.com/78?category=825914

 

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Xception: Deep Learning with Depthwise Separable Convolutions (feat.Pytorch)(1)

์•ˆ๋…•ํ•˜์„ธ์š”. ์ €๋ฒˆ Depthwise Separable Convolution ๊ธฐ๋ฒ•์— ๋Œ€ํ•ด ๊ธ€์„ ์˜ฌ๋ ธ์Šต๋‹ˆ๋‹ค. ์˜ค๋Š˜์€ ์ด ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•œ Xception ๋…ผ๋ฌธ์— ๋Œ€ํ•ด ๋ฆฌ๋ทฐํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. https://openaccess.thecvf.com/content_cvpr_2017/html..

coding-yoon.tistory.com

 

An Extreme version of Inception module๊ณผ Depthwise Separable Convolution๋Š” ๊ต‰์žฅํžˆ ๋น„์Šทํ•œ ํ˜•ํƒœ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

 

๋…ผ๋ฌธ์—์„œ ์†Œ๊ฐœํ•˜๋Š” ๋‘ ๊ฐœ์˜ ์ฐจ์ด์ ์„ ์„ค๋ช…ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

 

Two minor differences between and โ€œextremeโ€ version of an Inception module and a depthwise separable 
convolution would be:

โ€œextremeโ€ version of an Inception module๊ณผ depthwise separable 
convolution์˜ ๋‘ ๊ฐ€์ง€ ์‚ฌ์†Œํ•œ ์ฐจ์ด์ ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

 

 

 

1. Order(์ˆœ์„œ)

โ€ข The order of the operations: 
depthwise separable convolutions as usually implemented (e.g. in TensorFlow)perform first channel-wise 
spatial convolution and then perform 1x1 convolution, whereas Inception performs
the 1x1 convolution first.

Depthwise Separable Convolution : Depthwise Convolution(3x3, 5x5 ...), Pointwise Convolution 

An Extreme version of Inception module : Pointwise Convolution , Depthwise Convolution(3x3, 5x5 ...)

 

Convolution์˜ ์ˆœ์„œ๊ฐ€ ๋‹ค๋ฅด๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.

We argue that the first difference is unimportant, in particular because these operations are meant
to be used in a stacked setting.

 

์Šคํƒ ์„ค์ •์—์„œ ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์ˆœ์„œ์˜ ์ฐจ์ด์ ์€ ์ค‘์š”ํ•˜์ง€ ์•Š๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.

 

 

 

2. Non-Linearity(๋น„์„ ํ˜•์„ฑ)

โ€ข The presence or absence of a non-linearity after the first operation.
In Inception, both operations are followed by a ReLU non-linearity, however depthwise separable 
convolutions are usually implemented without non-linearities.

ReLU(๋น„์„ ํ˜•)์˜ ์œ ๋ฌด.

 

Inception์€ Convolution ์ˆ˜ํ–‰์ด ํ›„ ReLU๊ฐ€ ๋ถ™๋Š” ๋ฐ˜๋ฉด, 

 

์ผ๋ฐ˜์ ์œผ๋กœ Depthwise Separable Convolution๋Š” ReLU๊ฐ€ ์—†์ด ๊ตฌํ˜„๋ฉ๋‹ˆ๋‹ค. 

 

The second difference might matter, and we investigate it in the experimental section 
(in particular see figure 10).

 

ReLU์˜ ์œ ๋ฌด๋Š” ์ค‘์š”ํ•˜๋ฉฐ, Figure 10์—์„œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

 

Depthwise Separable Convolution์—์„œ ReLU๋ฅผ ์‚ฌ์šฉํ•˜์ง€ ์•Š์•˜์„ ๋•Œ ๋” ๋†’์€ Accruacy๋ฅผ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

 

It may be that the depth of the intermediate feature spaces on which spatial convolutions are applied 
is critical to the usefulness of the non-linearity:

 for deep feature spaces (e.g. those found in Inception modules) the non-linearity is helpful, 
 but for shallow ones (e.g. the 1-channel deep feature spaces of depthwise separable convolutions) it becomes harmful, 
 possibly due to a loss of information.

 

๊นŠ์€ ํŠน์ง• ๊ณต๊ฐ„์˜ ๊ฒฝ์šฐ, ReLU๊ฐ€ ๋„์›€์ด ๋˜์ง€๋งŒ  ์–•์€ ๊ณต๊ฐ„( 1x1 Convolution )์—์„œ๋Š” ReLU์— ์˜ํ•ด ์ •๋ณด ์†์‹ค์ด ์ƒ๊ธธ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š” ๊ฒƒ์ด ์ข‹๋‹ค๋Š” ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.

 

๋‹ค์Œ ๊ธ€์€ Xception์˜ Architecture์— ๋Œ€ํ•ด ์„ค๋ช…ํ•˜๊ณ , Pytorch๋กœ ๊ตฌํ˜„ํ•จ์œผ๋กœ ๊ธ€์„ Xception ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ๋งˆ์น˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. ํ”ผ๋“œ๋ฐฑ์ฃผ์‹œ๋ฉด ๊ฐ์‚ฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

728x90
๋ฐ˜์‘ํ˜•
18์ง„์ˆ˜