
import torch 
import torch.nn as nn
import numpy as np
from typing import Optional, Tuple, Any
device = ('cuda' if torch.cuda.is_available() else 'cpu')
print(device)

cuda


class InceptionModule(nn.Module):
    """
    in_channel: 初始输入模型通道数量
    out1: 第一层1x1卷积输出
    pre_in3: 3x3卷积输入 
    com_out3: 3x3卷积输出
    pre_in5: 5x5卷积输入 
    com_out5: 5x5卷积输出
    pool_out: 池化层输出
    """
    def __init__(self, in_channel, out1, pre_in3, com_out3, pre_in5, com_out5, pool_out):
        super(InceptionModule, self).__init__()
        self.con1 = nn.Conv2d(in_channel, out1, kernel_size= 1) # 1x1卷积层

        self.con3 = nn.Sequential(
            nn.Conv2d(in_channel, pre_in3, kernel_size= 1),
            nn.Conv2d(pre_in3, com_out3, kernel_size=3, padding= 1),
        ) # 1x1+3x3

        self.con5 = nn.Sequential(
            nn.Conv2d(in_channel, pre_in5, kernel_size= 1),
            nn.Conv2d(pre_in5, com_out5, kernel_size= 5, padding= 2)
        ) # 1x1+5x5

        self.pool = nn.Sequential(
            nn.MaxPool2d(kernel_size=3, stride=1, padding=1, ceil_mode=True),
            nn.Conv2d(in_channel, pool_out, kernel_size= 1, stride= 1)
        ) # max_pooling+1x1
    def forward(self, x):                          
        x1 = self.con1(x)
        x3 = self.con3(x)
        x5 = self.con5(x)
        max_x = self.pool(x)
        # print(x1.size(), x3.size(), x5.size(), max_x.size())
        return torch.concatenate([x1, x3, x5, max_x], 1)

class ConvBlock(nn.Module):
    """
    定义一个卷积网络类型，简化后续卷积操作
    """
    def __init__(self, in_channels: int, out_channels: int, **kwargs: Any) -> None:
        super(ConvBlock, self).__init__()
        self.conv = nn.Conv2d(in_channels, out_channels, bias=False, **kwargs)
        self.batch_norm = nn.BatchNorm2d(out_channels, eps= 0.001)
        self.relu = nn.ReLU(True)

    def forward(self, x):
        out = self.conv(x)
        out = self.batch_norm(out)
        out = self.relu(out)
        return out

class InceptionAux(nn.Module):
    """
    在GoogLeNet中存在分支判断函数，因此补充分支判断函数定义
    """
    def __init__(
            self, 
            input_size: int, 
            num_classes: int= 1000, 
            dropout: float=0.7) -> None:
        super(InceptionAux, self).__init__()
        self.avg_pool = nn.AvgPool2d(kernel_size=5, stride=3, ceil_mode=True)
        self.conv = ConvBlock(input_size, 128, kernel_size=1, stride=1)
        self.fc1 = nn.Linear(2048, 1024)
        self.fc2 = nn.Linear(1024, num_classes)
        self.relu = nn.ReLU(True)
        self.dropout = nn.Dropout(dropout, True)
    def forward(self, x):
        x = self.avg_pool(x)
        x = self.conv(x)
        x = torch.flatten(x, 1)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc2(x)
        return x


class GoogLeNet(nn.Module):
    """
    input_size:输入模型图片channel
    out_label:输出图片类型
    """
    def __init__(
            self, 
            input_size: int=3, 
            num_classes: int= 1000, 
            dropout: float = 0.4, 
            dropout_aux: float = 0.7,
            inception_aux: bool= True) -> None:
        super(GoogLeNet, self).__init__()
        
        self.inception_aux = inception_aux # 判断是否走分支
        self.conv1 = ConvBlock(input_size, 64, kernel_size=7, stride=2, padding=3)
        self.max_pool1 = nn.MaxPool2d(kernel_size= 3, stride=2) # 56x56x64
        self.conv2 = ConvBlock(64, 192, kernel_size=1, stride=1) # 56x56x192
        self.conv3 = ConvBlock(192, 192, kernel_size=3, stride=1, padding=1) # 56x56x192
        self.max_pool2 = nn.MaxPool2d(kernel_size=3, stride=2) # 28x28x192

        self.incept3a = InceptionModule(192, 64, 96, 128, 16, 32, 32) # 28x28x256
        self.incept3b = InceptionModule(256, 128, 128, 192, 32, 96, 64) # 28x28x480
        self.max_pool3 = nn.MaxPool2d(kernel_size=3, stride=2) # 14x14x480

        self.incept4a = InceptionModule(480, 192, 96, 208, 16, 48, 64) # 14x14x512
        # 判断
        self.incept4b = InceptionModule(512, 160, 112, 224, 24, 64, 64) # 14x14x512
        self.incept4c = InceptionModule(512, 128, 128, 256, 24, 64, 64) # 14x14x512
        self.incept4d = InceptionModule(512, 112, 144, 288, 32, 64, 64) # 14x14x528
        # 判断
        self.incept4e = InceptionModule(528, 256, 160, 320, 32, 128, 128) # 14x14x832

        self.max_pool4 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) # 7x7x832
        self.incept5a = InceptionModule(832, 256, 160, 320, 32, 128, 128) # 7x7x832
        self.incept5b = InceptionModule(832, 384, 192, 384, 48, 128, 128) # 7x7x1024

        self.avg_pool = nn.AvgPool2d(kernel_size=7, stride=1) # 1x1x1024
        self.linear = nn.Linear(1024, num_classes) # 1x1x1000
        self.dropout = nn.Dropout(dropout_aux)

        # 补充判断函数
        if inception_aux:
            self.inception_aux1 = InceptionAux(512, num_classes, dropout)
            self.inception_aux2 = InceptionAux(528, num_classes, dropout)
        else:
            self.inception_aux1 = None
            self.inception_aux2 = None

    def forward(self, x):
        x = self.conv1(x)
        x = self.max_pool1(x)
        x = self.max_pool2(self.conv3(self.conv2(x)))
        x = self.incept3a(x)
        x = self.incept3b(x)
        x = self.max_pool3(x)
        x = self.incept4a(x)

        if self.inception_aux1 is not None:
            aux1 = self.inception_aux1(x)

        x = self.incept4b(x)
        x = self.incept4c(x)
        x = self.incept4d(x)

        if self.inception_aux2 is not None:
            aux2 = self.inception_aux2(x)

        x = self.incept4e(x)
        x = self.max_pool4(x)
        x = self.incept5a(x)
        x = self.incept5b(x)
        x = self.avg_pool(x)
        x = torch.flatten(x, 1)
        x = self.dropout(x)
        x = self.linear(x)
        
        return x, aux2, aux1


model = GoogLeNet().to(device)
x = torch.randn(1, 3, 224, 224).to(device)
y = model(x)

GoogLeNet¶