pytorch-googleNet

googleNet网络结构

输入网络: 由4个分支网络构成

第一分支: 由1x1的卷积构成

第二分支: 由1x1的卷积，3x3的卷积构成

第三分支: 由1x1的卷积, 5x5的卷积构成

第四分支: 由3x3的最大值池化, 1x1的卷积构成

import torch
from torch import nn
from torch.nn import functional as F
 
class BasicConv2d(nn.Module):
    def __init__(self, in_channels, out_channels, **kwargs):
        super(BasicConv2d, self).__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs) # 构造卷积层
        self.bn = nn.BatchNorm2d(out_channels, eps=0.001) # 构造标准化
 
    def forward(self, x):
        x = self.conv(x) # 进行卷积操作
        x = self.bn(x) # 进行标准化操作
        x = F.relu(x) # 进行激活层操作
 
        return x
 
class Inception(nn.Module):
    def __init__(self, in_channels, pool_features):
        super(Inception, self).__init__()
        self.branch1x1 = BasicConv2d(in_channels, 64, kernel_size=1) # 1x1的卷积操作
 
        self.branch5x5_1 = BasicConv2d(in_channels, 48, kernel_size=1) # 进行卷积操作
        self.branch5x5_2 = BasicConv2d(48, 64, kernel_size=5, padding=2) 
 
        self.branch3x3db1_1 = BasicConv2d(in_channels, 64, kernel_size=1)
        self.branch3x3db1_2 = BasicConv2d(64, 96, kernel_size=3)
        self.branch3x3db1_3 = BasicConv2d(96, 96, kernel_size=3)
 
        self.branch_pool = BasicConv2d(in_channels, pool_features, kernel_size=1)
 
    def forward(self, x):
        branch1x1 = self.branch1x1(x)
 
        branch5x5 = self.branch5x5_1(x)
        branch5x5 = self.branch5x5_2(branch5x5)
 
        branch3x3db1_1 = self.branch3x3db1_1(x)
        branch3x3db1_2 = self.branch3x3db1_2(branch3x3db1_1)
        branch3x3db1_3 = self.branch3x3db1_3(branch3x3db1_2)
 
        branch_pool = F.max_pool2d(x, kernel_size=3, stride=1, padding=1)
        branch_pool = self.branch_pool(branch_pool)
        # 进行卷积的叠加操作
        outputs = [branch1x1, branch5x5, branch3x3db1_3, branch_pool]
        outputs = torch.cat(outputs, dim=1)
 
        return outputs