0️⃣ Paper

Very Deep Convolutional Networks for Large-Scale Image Recognition

1️⃣ Architecture Point

• 224 x 224의 RGB 이미지를 입력으로 사용함

  • 이미지의 전처리는 Training Set에서 계산된 평균 RGB 값을 각 픽셀에서 뺌
  • 입력 데이터의 Spatial Padding은 Convolution 후에도 공간 해상도가 유지되도록 설정함

• Convolutional Layer에서는 3 x 3 필터를 사용하며, Stride는 항상 1px 고정

  • Padding은 공간 해상도 유지를 위해 1px 고정

• 특정 구성에서는 1 x 1 필터를 사용함

  • 입력 채널에 대한 선형 변환(뒤따르는 비선형성을 포함)

• Max Pooling은 2 x 2 필터를 사용하며, Stride는 2px

• FC Layer, Activation Function(use ReLU)는 AlexNet과 동일함

2️⃣ VGGNet Architecture Visualization

3️⃣ Architecture Summary

• VGGNet-A

  Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map	Activation Function
  Input	-	-	-	-	224 x 224 x 3	-
  Convolution 1	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Max Pool 1	-	2 x 2	2	-	112 x 112 x 64	-
  Convolution 2	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Max Pool 2	-	2 x 2	2	-	56 x 56 x 128	-
  Convolution 3	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Convolution 4	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Max Pool 3	-	2 x 2	2	-	28 x 28 x 256	-
  Convolution 5	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Convolution 6	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Max Pool 4	-	2 x 2	2	-	14 x 14 x 512	-
  Convolution 7	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Convolution 8	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Max Pool 5	-	2 x 2	2	-	7 x 7 x 512	-
  Dense(FC) 1	4096	-	-	-	1 x 4096	-
  Dense(FC) 2	4096	-	-	-	1 x 4096	-
  Softmax	1000	-	-	-	1 x 1000	-

• VGGNet-B

  Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map	Activation Function
  Input	-	-	-	-	224 x 224 x 3	-
  Convolution 1	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Convolution 2	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Max Pool 1	-	2 x 2	2	-	112 x 112 x 64	-
  Convolution 3	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Convolution 4	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Max Pool 2	-	2 x 2	2	-	56 x 56 x 128	-
  Convolution 5	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Convolution 6	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Max Pool 3	-	2 x 2	2	-	28 x 28 x 256	-
  Convolution 7	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Convolution 8	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Max Pool 4	-	2 x 2	2	-	14 x 14 x 512	-
  Convolution 9	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Convolution 10	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Max Pool 5	-	2 x 2	2	-	7 x 7 x 512	-
  Dense(FC) 1	4096	-	-	-	1 x 4096	-
  Dense(FC) 2	4096	-	-	-	1 x 4096	-
  Softmax	1000	-	-	-	1 x 1000	-

• VGGNet-C

  Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map	Activation Function
  Input	-	-	-	-	224 x 224 x 3	-
  Convolution 1	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Convolution 2	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Max Pool 1	-	2 x 2	2	-	112 x 112 x 64	-
  Convolution 3	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Convolution 4	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Max Pool 2	-	2 x 2	2	-	56 x 56 x 128	-
  Convolution 5	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Convolution 6	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Convolution 7	256	1 x 1	1	1	58 x 58 x 256	ReLU
  Max Pool 3	-	2 x 2	2	-	29 x 29 x 256	-
  Convolution 8	512	3 x 3	1	1	29 x 29 x 512	ReLU
  Convolution 9	512	3 x 3	1	1	29 x 29 x 512	ReLU
  Convolution 10	512	1 x 1	1	1	31 x 31 x 512	ReLU
  Max Pool 4	-	2 x 2	2	-	15 x 15 x 512	-
  Convolution 11	512	3 x 3	1	1	15 x 15 x 512	ReLU
  Convolution 12	512	3 x 3	1	1	15 x 15 x 512	ReLU
  Convolution 13	512	1 x 1	1	1	17 x 17 x 512	ReLU
  Max Pool 5	-	2 x 2	2	-	8 x 8 x 512	-
  Dense(FC) 1	4096	-	-	-	1 x 4096	-
  Dense(FC) 2	4096	-	-	-	1 x 4096	-
  Softmax	1000	-	-	-	1 x 1000	-

• VGGNet-D

  Layer	# Filters / neurons	Filter Size	Stride	Padding	Size of Feature Map	Activation Function
  Input	-	-	-	-	224 x 224 x 3	-
  Convolution 1	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Convolution 2	64	3 x 3	1	1	224 x 224 x 64	ReLU
  Max Pool 1	-	2 x 2	2	-	112 x 112 x 64	-
  Convolution 3	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Convolution 4	128	3 x 3	1	1	112 x 112 x 128	ReLU
  Max Pool 2	-	2 x 2	2	-	56 x 56 x 128	-
  Convolution 5	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Convolution 6	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Convolution 7	256	3 x 3	1	1	56 x 56 x 256	ReLU
  Max Pool 3	-	2 x 2	2	-	28 x 28 x 256	-
  Convolution 8	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Convolution 9	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Convolution 10	512	3 x 3	1	1	28 x 28 x 512	ReLU
  Max Pool 4	-	2 x 2	2	-	14 x 14 x 512	-
  Convolution 11	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Convolution 12	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Convolution 13	512	3 x 3	1	1	14 x 14 x 512	ReLU
  Max Pool 5	-	2 x 2	2	-	7 x 7 x 512	-
  Dense(FC) 1	4096	-	-	-	1 x 4096	-
  Dense(FC) 2	4096	-	-	-	1 x 4096	-
  Softmax	1000	-	-	-	1 x 1000	-

• VGGNet-E