Pytorch中卷积层转向全连接层时，全连接层输入维度的确定

一、代码法：

改变的AlexNet网络结构：

输入维度为$（N，1,100，100）$，$N$为输入的样本数。

import torch.nn as nn
import torch.nn.functional as F
import torch
from torchsummary import summary

class AlexNet(nn.Module):
    def __init__(self):
        super(AlexNet, self).__init__()
        self.conv1 = nn.Conv2d(1, 96, 11, stride=2)
        self.conv2 = nn.Conv2d(96, 256, 5, stride=1, padding=2)
        self.conv3 = nn.Conv2d(256, 384, 3, stride=1, padding=1)
        self.conv4 = nn.Conv2d(384, 384, 3, stride=1, padding=1)
        self.conv5 = nn.Conv2d(384, 256, 3, stride=1, padding=1)
        self.fc1 = nn.Linear(256*6*6, 4096)
        self.fc2 = nn.Linear(4096, 4096)
        self.fc3 = nn.Linear(4096, 1000)

    def forward(self, x):
        out = F.relu(self.conv1(x))
        out = F.max_pool2d(out, 2)
        out = F.relu(self.conv2(out))
        out = F.max_pool2d(out, 2)
        out = F.relu(self.conv3(out))
        out = F.relu(self.conv4(out))
        out = F.relu(self.conv5(out))
        out = F.max_pool2d(out, 2)
        out = out.view(out.size(0), -1)
        out = F.relu(self.fc1(out))
        out = F.relu(self.fc2(out))
        out = self.fc3(out)
        return out

以 AlexNet 网络结构为例，如上代码。在AlexNet的基础上，将第一层的卷积层的输入通道变成1。代码法相对直观，随机产生一个维度的数据作为调试，我们选择的维度为$（1,100,100）$，代码如下：

net = AlexNet()
image = torch.randn(2, 1, 100, 100)
out = F.relu(net.conv1(image))
out = F.max_pool2d(out, 2)
out = F.relu(net.conv2(out))
out = F.max_pool2d(out, 2)
out = F.relu(net.conv3(out))
out = F.relu(net.conv4(out))
out = F.relu(net.conv5(out))
out = F.max_pool2d(out, 2)
out.size()

得到输出为：

torch.Size([2, 256, 5, 5])

卷积层的输出维度为：（2, 256, 5, 5）。
全连接层的输入维度为卷积层的输出维度为：$256\times 5\times 5$ 。则AlexNet中下面的这行代码：

self.fc1 = nn.Linear(256*6*6, 4096)

需要改为：

self.fc1 = nn.Linear(256*5*5,4096)

全连接层输入的维度为每个图片总的像素点个数。

AlexNet

import torch.nn as nn
import torch.nn.functional as F
import torch
from torchsummary import summary

class AlexNet(nn.Module):
    def __init__(self):
        super(AlexNet, self).__init__()
        self.conv1 = nn.Conv2d(3, 96, 11, stride=4)
        self.conv2 = nn.Conv2d(96, 256, 5, stride=1, padding=2)
        self.conv3 = nn.Conv2d(256, 384, 3, stride=1, padding=1)
        self.conv4 = nn.Conv2d(384, 384, 3, stride=1, padding=1)
        self.conv5 = nn.Conv2d(384, 256, 3, stride=1, padding=1)
        self.fc1 = nn.Linear(256*6*6, 4096)
        self.fc2 = nn.Linear(4096, 4096)
        self.fc3 = nn.Linear(4096, 1000)

    def forward(self, x):
        out = F.relu(self.conv1(x))
        out = F.max_pool2d(out, 2)
        out = F.relu(self.conv2(out))
        out = F.max_pool2d(out, 2)
        out = F.relu(self.conv3(out))
        out = F.relu(self.conv4(out))
        out = F.relu(self.conv5(out))
        out = F.max_pool2d(out, 2)
        out = out.view(out.size(0), -1)
        out = F.relu(self.fc1(out))
        out = F.relu(self.fc2(out))
        out = self.fc3(out)
        return out

调试代码为：

net = AlexNet()
image = torch.randn(2, 3, 227, 227)
out = F.relu(net.conv1(image))
out = F.max_pool2d(out, 2)
out = F.relu(net.conv2(out))
out = F.max_pool2d(out, 2)
out = F.relu(net.conv3(out))
out = F.relu(net.conv4(out))
out = F.relu(net.conv5(out))
out = F.max_pool2d(out, 2)
out.size()

输出为：

torch.Size([2, 256, 6, 6])

全连接层的输入维度为：$256\times 6\times 6$。

二、公式法：

torch.nn.Conv2d(in_channels,
               out_channels, 
               kernel_size, 
               stride=1, 
               padding=0, 
               dilation=1, 
               groups=1, 
               bias=True, 
               padding_mode='zeros', 
               device=None, 
               dtype=None)

参数如下：

• n_channels (int) ：输入图片的维度；
• out_channels (int) ：卷积层的输出维度；
• kernel_size (int or tuple) ：卷积核的大小；
• stride (int or tuple, optional) ：卷积核在原图上扫描时，需要跳跃的格数，默认跳一格；
• padding (int, tuple or str, optional) ：输入四个边的填充值，默认为0 ，不填充；
• padding_mode (string, optional) ：填充的方式有’zeros’, ‘reflect’, ‘replicate’ or ‘circular’. 默认为’zeros’，填充0值；
• dilation (int or tuple, optional) ：卷积核元素间的间隔，默认为1；
• groups (int, optional) ：输入通道和输出通道间的联系类别数，默认为1；
• bias (bool, optional) ：如果为真，添加一个学习的偏置到输出上，默认为真。

卷积层的输入和输出关系，如下：$H_{out}$和$W_{out}$为卷积层输出图片的高和宽。