[Classification] VGGNet

VGGNet

• VGGNet 연구팀은 망의 깊이가 깊을 수록 model의 성능에 어떤 영향을 끼치는지 연구.
• 5x5 Conv를 한번 하는것 보다, 3x3 Conv를 두번 하는것이 망의 깊이가 깊어지고, parameter 수도 적어짐.

 

• VGGNet은 모든 Convolution filter size를 3x3으로 고정해서 사용.

 

VGGNet16 구조

 

 

Pytorch Code

import torch
import torch.nn as nn


def CBR2d(in_channels, out_channels, _kernal_size, _stride, _padding):
    return nn.Sequential(
        nn.Conv2d(in_channels, out_channels, kernel_size=_kernal_size, stride=_stride, padding=_padding),
        nn.BatchNorm2d(out_channels),
        nn.ReLU(inplace=True)
    )


class QVGGNet(nn.Module):
    def __init__(self, num_classes=2):
        super(QVGGNet, self).__init__()

        self.pooling = nn.MaxPool2d(2, 2)
        self.relu = nn.ReLU(True)
        self.dropout = nn.Dropout()

        self.conv1_1 = CBR2d(3, 64, 3, 1, 1)
        self.conv1_2 = CBR2d(64, 64, 3, 1, 1)

        self.conv2_1 = CBR2d(64, 128, 3, 1, 1)
        self.conv2_2 = CBR2d(128, 128, 3, 1, 1)

        self.conv3_1 = CBR2d(128, 256, 3, 1, 1)
        self.conv3_2 = CBR2d(256, 256, 3, 1, 1)
        self.conv3_3 = CBR2d(256, 256, 3, 1, 1)

        self.conv4_1 = CBR2d(256, 512, 3, 1, 1)
        self.conv4_2 = CBR2d(512, 512, 3, 1, 1)
        self.conv4_3 = CBR2d(512, 512, 3, 1, 1)

        self.conv5_1 = CBR2d(512, 512, 3, 1, 1)
        self.conv5_2 = CBR2d(512, 512, 3, 1, 1)
        self.conv5_3 = CBR2d(512, 512, 3, 1, 1)

        self.fc1 = nn.Linear(7*7*512, 4096)
        self.fc2 = nn.Linear(4096, 4096)
        self.fc3 = nn.Linear(4096, num_classes)

        # input image size 224x224x3
        self.feature = nn.Sequential(
            self.conv1_1,
            self.conv1_2,
            self.pooling,       # 112x112x128

            self.conv2_1,
            self.conv2_2,
            self.pooling,       # 56x56x256

            self.conv3_1,
            self.conv3_2,
            self.conv3_3,
            self.pooling,       # 28x28x512

            self.conv4_1,
            self.conv4_2,
            self.conv4_3,
            self.pooling,       # 14x14x512

            self.conv5_1,
            self.conv5_2,
            self.conv5_3,
            self.pooling,       # 7x7x512
        )

        self.classifier = nn.Sequential(
            self.fc1,
            self.relu,
            self.dropout,
            self.fc2,
            self.relu,
            self.dropout,
            self.fc3,
        )

    def forward(self, x):
        x = self.feature(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x