VGG-16

13 Convolution layers + 3 Fully-connected layers = 16 layers
$3\times3$ convolution filters (stride = $1$, padding = $1$)
$2\times2$ max pooling (stride = $2$)

---

Architecture

Input
$224\times224\times3$ image.

1) Conv 1_1
convolution : 64개의 $3\times3\times3$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 64개의 $224\times224$ feature maps

2) Conv 1_2
convolution : 64개의 $3\times3\times64$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 64개의 $224\times224$ feature maps
$2\times2$ max pooling (stride = $2$)
⇒ 64개의 $112\times112$ feature maps

3) Conv 2_1
convolution : 128개의 $3\times3\times64$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 128개의 $112\times112$ feature maps

4) Conv 2_2
convolution : 128개의 $3\times3\times128$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 128개의 $112\times112$ feature maps
$2\times2$ max pooling (stride = $2$)
⇒ 128개의 $56\times56$ feature maps

5) Conv 3_1
convolution : 256개의 $3\times3\times128$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 256개의 $56\times56$ feature maps

6) Conv 3_2
convolution : 256개의 $3\times3\times256$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 256개의 $56\times56$ feature maps

7) Conv 3_3
convolution : 256개의 $3\times3\times256$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 256개의 $56\times56$ feature maps
$2\times2$ max pooling (stride = $2$)
⇒ 256개의 $28\times28$ feature maps

8) Conv 4_1
convolution : 512개의 $3\times3\times256$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 512개의 $28\times28$ feature maps

9) Conv 4_2
convolution : 512개의 $3\times3\times512$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 512개의 $28\times28$ feature maps

10) Conv 4_3
convolution : 512개의 $3\times3\times512$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 512개의 $28\times28$ feature maps
$2\times2$ max pooling (stride = $2$)
⇒ 512개의 $14\times14$ feature maps

11) Conv 5_1
convolution : 512개의 $3\times3\times512$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 512개의 $14\times14$ feature maps

12) Conv 5_2
convolution : 512개의 $3\times3\times512$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 512개의 $14\times14$ feature maps

13) Conv 5_3
convolution : 512개의 $3\times3\times512$ kernels (stride = $1$, padding = $1$)
activate by ReLU
⇒ 512개의 $14\times14$ feature maps
$2\times2$ max pooling (stride = $2$)
⇒ 512개의 $7\times7$ feature maps

14) Fc1
$7\times7\times512$ feature maps을 flatten
⇒ $25088\times1$ neurons
Fc1의 4096개 neurons와 fully-connected.
Dropout 적용.
⇒ 4096개 neurons

15) Fc2
Fc2의 4096개 neurons와 fully-connected.
Dropout 적용.
⇒ 4096개 neurons

16) Fc3
Fc3의 1000개 neurons와 fully-connected. (※ $1000$ = # of classes)
activate by softmax

---

Reference
[CNN 알고리즘들] VGGNet의 구조 (VGG16)
VGG16 – Convolutional Network for Classification and Detection