MobileNet V1模型——pytorch实现

论文传送门：https://arxiv.org/pdf/1704.04861.pdf

MobileNet V1的目的：对图片进行特征提取，依据特征进行分类。

MobileNet V1的优点：轻量。

MobileNet V1的方法：使用深度可分离卷积(Depthwise Separable Convolution)，减少模型参数量。深度可分离卷积由深度卷积(Depthwise Convolution)和逐点卷积(Pointwise Convolution)共同完成。

 MobileNet V1的结构：主要由两部分构成：

①特征提取部分：由1个标准卷积块和13个深度可分离卷积块构成；(红框部分实际实现时s=1)

②分类部分：AvgPool+FC+Softmax。

 

import torch
import torch.nn as nn


def standard_conv_block(in_channel, out_channel, strid=1):  # 定义Strandard convolutional layer with batchnorm and ReLU
    return nn.Sequential(
        nn.Conv2d(in_channel, out_channel, 3, strid, 1, bias=False),  # conv
        nn.BatchNorm2d(out_channel),  # bn
        nn.ReLU()  # relu
    )


def depthwise_separable_conv_block(in_channel, out_channel,
                                   strid=1):  # 定义Depthwise Separable convolutions with Depthwise and Pointwise layers followed by batchnorm and ReLU
    return nn.Sequential(
        nn.Conv2d(in_channel, in_channel, 3, strid, 1, groups=in_channel, bias=False),  # conv,使用与输入通道数相同组数的分组卷积实现Depthwise Convolution
        nn.BatchNorm2d(in_channel),  # bn
        nn.ReLU(),  # relu
        nn.Conv2d(in_channel, out_channel, 1, 1, 0, bias=False),  # 1x1conv,Pointwise Convolution
        nn.BatchNorm2d(out_channel),  # bn
        nn.ReLU()  # relu
    )


class MobileNetV1(nn.Module):  # 定义MobileNet结构
    def __init__(self, num_classes=1000):  # 初始化方法
        super(MobileNetV1, self).__init__()  # 继承初始化方法

        self.num_classes = num_classes  # 类别数量
        self.feature = nn.Sequential(  # 特征提取部分
            standard_conv_block(3, 32, strid=2),  # standard conv block,(n,3,224,224)-->(n,32,112,112)
            depthwise_separable_conv_block(32, 64),  # depthwise separable conv block,(n,32,112,112)-->(n,64,112,112)
            depthwise_separable_conv_block(64, 128, strid=2),  # depthwise separable conv block,(n,64,112,112)-->(n,128,56,56)
            depthwise_separable_conv_block(128, 128),  # depthwise separable conv block,(n,128,56,56)-->(n,128,56,56)
            depthwise_separable_conv_block(128, 256, strid=2),  # depthwise separable conv block,(n,128,56,56)-->(n,256,28,28)
            depthwise_separable_conv_block(256, 256),  # depthwise separable conv block,(n,256,28,28)-->(n,256,28,28)
            depthwise_separable_conv_block(256, 512, strid=2),  # depthwise separable conv block,(n,256,28,28)-->(n,512,14,14)
            depthwise_separable_conv_block(512, 512),  # depthwise separable conv block,(n,512,14,14)-->(n,512,14,14)
            depthwise_separable_conv_block(512, 512),  # depthwise separable conv block,(n,512,14,14)-->(n,512,14,14)
            depthwise_separable_conv_block(512, 512),  # depthwise separable conv block,(n,512,14,14)-->(n,512,14,14)
            depthwise_separable_conv_block(512, 512),  # depthwise separable conv block,(n,512,14,14)-->(n,512,14,14)
            depthwise_separable_conv_block(512, 512),  # depthwise separable conv block,(n,512,14,14)-->(n,512,14,14)
            depthwise_separable_conv_block(512, 1024, strid=2),  # depthwise separable conv block,(n,512,14,14)-->(n,1024,7,7)
            depthwise_separable_conv_block(1024, 1024),  # depthwise separable conv block,(n,1024,7,7)-->(n,1024,7,7)
            nn.AdaptiveAvgPool2d(1)  # avgpool,为方便后续转为特征向量,这里将avgpool放入特征提取部分,(n,1024,7,7)-->(n,1024,1,1)
        )
        self.classifier = nn.Sequential(  # 分类部分
            nn.Linear(1024, self.num_classes),  # linear,(n,1024)-->(n,num_classes)
            nn.Softmax(dim=1)  # softmax
        )

    def forward(self, x):  # 前传函数
        x = self.feature(x)  # 特征提取，获得特征层,(n,1024,1,1)
        x = torch.flatten(x, 1)  # 将三维特征层压缩至一维特征向量,(n,1024,1,1)-->(n,1024)
        return self.classifier(x)  # 分类，输出类别概率,(n,num_classes)