ResNet缩小版

最新推荐文章于 2022-10-01 10:52:10 发布

云溪看海也看你

最新推荐文章于 2022-10-01 10:52:10 发布

阅读量152

点赞数

文章标签： python 深度学习 pytorch

本文链接：https://blog.csdn.net/qq_37211687/article/details/125985039

版权

该博客介绍了如何实现ResNet深度学习模型，主要通过ResBlock残差单元模块来构建。模型包括初始化卷积层、多个ResBlock组成的残差层以及全局平均池化和全连接层。代码详细展示了每个层的定义和前向传播过程，适用于图像分类任务。

摘要由CSDN通过智能技术生成

import torch
import torch.nn as nn
from torch.nn.functional import log_softmax


class ResBlock(nn.Module):
    '''
    在这里实现残差的单元模块
    '''
    def __init__(self,in_channel,out_channel,stride):
        super(ResBlock, self).__init__()
        self.layer = nn.Sequential(
            nn.Conv2d(in_channel,out_channel,(3,3),stride,padding=1),
            nn.BatchNorm2d(out_channel),
            nn.ReLU(),
            nn.Conv2d(out_channel,out_channel,(3,3),(1,1),padding=1),
            nn.BatchNorm2d(out_channel)
        )
        self.shutcut = nn.Sequential()
        if stride!=1 or in_channel!=out_channel:
            self.shutcut = nn.Sequential(
                nn.Conv2d(in_channel, out_channel, (3, 3), stride,padding=1),
                nn.BatchNorm2d(out_channel),
                nn.ReLU()
            )
    def forward(self,x):
        left = self.layer(x)
        # print('left',left.size())
        right = self.shutcut(x)
        # print('right', left.size())
        out = left + right
        return out



class ResNet(nn.Module):
    def make_layer(self,resblock,in_channel,out_channel,stride,num_block):
        block = nn.Sequential()
        for i in range(num_block):
            # 仅有第一次的时候进行一次降采样
            if stride != 1 and i==0:
                in_stride = stride
            else:
                in_stride = 1
            res_block = resblock(in_channel,out_channel,in_stride)
            block.append(res_block)
            in_channel = out_channel
        return block


    def __init__(self,resblock):
        super(ResNet, self).__init__()
        self.inchannel = 32
        self.init_conv = nn.Sequential(
            nn.Conv2d(3, self.inchannel, (3, 3), (1, 1), 1),
            nn.BatchNorm2d(32),
            nn.ReLU()
        )
        self.layer1 = self.make_layer(resblock, 32, 64,  2, 2) # 14
        self.layer2 = self.make_layer(resblock, 64, 128, 2, 2) # 7
        self.layer3 = self.make_layer(resblock, 128, 256, 2, 2) # 4
        self.layer4 = self.make_layer(resblock, 256, 512, 2, 2) # 2
        self.avg = nn.AvgPool2d(2,1)
        self.linear = nn.Linear(512,10)

    def forward(self,x):
        out = self.init_conv(x)
        # print('1',out.size())
        out = self.layer1(out)
        # print('2', out.size())
        out = self.layer2(out)
        # print('3', out.size())

        out = self.layer3(out)
        # print('4', out.size())
        out = self.layer4(out)
        # print('5', out.size())
        avg = self.avg(out)
        out = avg.view(avg.size()[0],-1)
        out = self.linear(out)
        return log_softmax(out,dim=1)


def resnet():
   return ResNet(ResBlock)


resnet()