Thoroughly get to know VGGNet-16

VGGNet-16 as a whole can be divided into eight sections (section 8), the first segment is a convolutional network 5, section 3 is fully connected network.
First create a convolutional network first section, this section consists of two convolutional network layers and a convolution maximum cell layers, i.e., a total of 3 layers. For two layers convolution, convolution kernel size is 3x3, while the number of the convolution kernel (output channels) 64 are also steps of 1x1, padding are 1 (and not explicitly mentioned padding = 1, many places do not say so, do not know why) - this is one of several parameters do we care about convolution: the convolution kernel size, convolution kernel depth, step length, padding. Therefore, the first input of a convolution of the input layer size of 224x224x3, output size of 224x224x64; size of the input layer is the second convolution 224x224x64, output size of 224x224x64; after two largest convolution of a pool of 2x2 layer, the step 2, so that after the maximum cell layer, the output becomes the size 112x112x64. If the use of padding, kernel_size = time 1, padding = 0; kernel_size = time 3, padding = 1; when kernel_size = 5, padding = 2, and so on. That is, only convolution kernel size is odd, we just padding. Illustrated as follows:

the structure of a convolutional network and the second section the first section is very similar to the convolution of the network. Two convolution kernel size is 3x3, but the number is 128 * 128, s = 1, p = 1. The largest pool of the same layer of the largest pools layer and the first paragraph of convolution. I.e. 2x2, s = 2. As shown below:

a third segment and the first two different convolutional network it is the convolution becomes three layers. Convolution kernel size of each layer is still convolution 3x3, but each layer to increase the number of output channels convolution 256, s = 1, p = 1. The largest pool of the same layer and the first two paragraphs. As shown below:

consistent fourth paragraph convolutional network and a third segment convolutional network. Only the output channel number of each convolution layer 512 becomes. As shown below:

The last section of the convolutional network is also a three layer convolution kernel size is 3x3 convolution plus a maximum cell layer, but no increase in the number of output channels convolution layer, which will remain at 512. As shown below:

Therefore, section 5 after convolution, the output size of 7x7x512, total 2 + 2 + 3 + 3 + 3 = 13 layers. After that, the whole connection.
The first fully connected 4096 units.
The second fully connected 4096 units.
The third unit 1000 has fully connected. After using softmax ().
Plus three fully connected layers, a total of 13 + 3 = 16 layers.