PyTorch搭建神经网络模型的4种方法

PyTorch有多种方法搭建神经网络，下面识别手写数字为例，介绍4种搭建神经网络的方法。

方法一：torch.nn.Sequential()

torch.nn.Sequential类是torch.nn中的一种序列容器，参数会按照我们定义好的序列自动传递下去。

import torch.nn as nn
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Sequential(           # input shape (1, 28, 28)
            nn.Conv2d(1, 16, 5, 1, 2),        # output shape (16, 28, 28)
            nn.ReLU(),
            nn.MaxPool2d(2),                  # output shape (16, 14, 14)
        )
        self.conv2 = nn.Sequential(
            nn.Conv2d(16, 32, 5, 1, 2),       # output shape (32, 14, 14)
            nn.ReLU(),
            nn.MaxPool2d(2),                  # output shape (32, 7, 7)
        )
        self.linear = nn.Linear(32*7*7, 10)
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size(0), -1)
        output = self.linear(x)
        return output
 
net = Net()
print(net)

运行结果：

注意：这样做有一个问题，每一个层是没有名称，默认的是以0、1、2、3来命名，从上面的运行结果也可以看出。

方法二：torch.nn.Sequential() 搭配 collections.OrderDict()

import torch.nn as nn
from collections import OrderedDict   # OrderedDict是字典的子类，可以记住元素的添加顺序
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Sequential(OrderedDict([
            ('conv1', nn.Conv2d(1, 16, 5, 1, 2)),
            ('ReLU1', nn.ReLU()),
            ('pool1', nn.MaxPool2d(2)),
        ]))
        self.conv2 = nn.Sequential(OrderedDict([
            ('conv2', nn.Conv2d(16, 32, 5, 1, 2)),
            ('ReLU2', nn.ReLU()),
            ('pool2', nn.MaxPool2d(2)),
        ]))
        self.linear = nn.Linear(32*7*7, 10)
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size(0), -1)
        output = self.linear(x)
        return output
 
net = Net()
print(net)

运行结果：

从上面的结果中可以看出，这个时候每一个层都有了自己的名称，但是此时需要注意，我们并不能够通过名称直接获取层，依然只能通过索引index，即net.conv1[1] 是正确的，net.conv1['ReLU1']是错误的。这是因为torch.nn.Sequential()只支持index访问。

方法三：torch.nn.Sequential() 搭配 add_module()

import torch.nn as nn
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Sequential()
        self.conv1.add_module('conv1', nn.Conv2d(1, 16, 5, 1, 2))
        self.conv1.add_module('ReLU1', nn.ReLU())
        self.conv1.add_module('pool1', nn.MaxPool2d(2))
 
        self.conv2 = nn.Sequential()
        self.conv2.add_module('conv2', nn.Conv2d(16, 32, 5, 1, 2))
        self.conv2.add_module('ReLU2', nn.ReLU())
        self.conv2.add_module('pool2', nn.MaxPool2d(2))
 
        self.linear = nn.Linear(32*7*7, 10)
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = x.view(x.size(0), -1)
        output = self.linear(x)
        return output
 
net = Net()
print(net)

运行结果：

方法四

import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 16, 5, 1, 2)
        self.conv2 = nn.Conv2d(16, 32, 5, 1, 2)
        self.linear = nn.Linear(32*7*7, 10)
 
    def forward(self, x):
        x = F.max_pool2d(F.relu(self.conv1(x)), 2)
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        output = self.linear(x)
        return output
 
net = Net()
print(net)