Xception | Notion

Depthwise Separable Convolution Layer로만 구성된 CNN 구조 제안
가설) CNN의 Feature Map에서 채널 간 상관성과 공간 상관성의 Mapping을 완전히 분리할 수 있음
- Inception Architecture의 기본 가설 확장
- Extreme Inception을 의미하는 Xception이라 명명함
Redisual 연결이 있는 Depthwise Separable Convolution Layer의 선형(Linear) 스택
Separable Convolution → Depthwise Convolution + Pointwise Convolution
- Depthwise Convolution: Convolution(in_channels → in_channels / kernel_size = $\alpha$)
- Pointwise Convolution: Convolution(in_channels → out_channels / kernel_size = 1)

Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map
Input	-	-	-	-	299 x 299 x 3
Convolution 1	32	3 x 3	2	1	150 x 150 x 32
Convolution 2	64	3 x 3	1	1	150 x 150 x 64
Separable Convolution 1	128	3 x 3	1	1	150 x 150 x 128
Separable Convolution 2	128	3 x 3	1	1	150 x 150 x 128
Max Pool 1	-	3 x 3	2	1	75 x 75 x 128
Convolution - Residual 1
(Matrix Sum: Max Pool 1)	128	1 x 1	2	-	75 x 75 x 128
Separable Convolution 3	256	3 x 3	1	1	75 x 75 x 256
Separable Convolution 4	256	3 x 3	1	1	75 x 75 x 256
Max Pool 2	-	3 x 3	2	1	37 x 37 x 256
Convolution - Residual 2
(Matrix Sum: Max Pool 2)	256	1 x 1	2	-	37 x 37 x 256
Separable Convolution 5	728	3 x 3	1	1	37 x 37 x 728
Separable Convolution 6	728	3 x 3	1	1	37 x 37 x 728
Max Pool 3	-	3 x 3	2	1	19 x 19 x 728
Convolution - Residual 3
(Matrix Sum: Max Pool 3)
Separable Convolution 7	728	3 x 3	1	1	19 x 19 x 728
Separable Convolution 8	728	3 x 3	1	1	19 x 19 x 728
Separable Convolution 9	728	3 x 3	1	1	19 x 19 x 728
! Repeat
(Separable Convolution 7, 8, 9)	# Times = 8
Separable Convolution 10	728	3 x 3	1	1	19 x 19 x 728
Separable Convolution 11	1024	3 x 3	1	1	19 x 19 x 1024
Max Pool 4	-	3 x 3	2	1	9 x 9 x 1024
Convolution - Residual 4
(Matrix Sum: Max Pool 4)	1024	1 x 1	2	-	9 x 9 x 1024
Separable Convolution 12	1536	3 x 3	1	1	9 x 9 x 1536
Separable Convolution 13	2048	3 x 3	1	1	9 x 9 x 2048
Global Average Pool 1	-	9 x 9	1	-	1 x 1 x 2048
Softmax					1 x 1000