torch.nn.Module中modules和children的区别

torch.nn.Module中modules()和children()的区别

pytorch中自定义的网络结构一般都是通过继承nn.Module类来实现的。nn.Module中存在两个相似的method：torch.nn.Module.modules()和torch.nn.Modules.children()。两者有什么联系和区别呢？

import torch.nn as nn
nn.Modules.modules??

Returns an iterator over all modules in the network

import torch.nn as nn
nn.Modules.children??

Returns an iterator over immediate children modules.

• 相同点：返回网络中定义的所有module
• 不同点：modules会递归地返回网络定义的所有module，而children只会返回直接子module
• 例子：

import torch
import torch.nn as nn
# 定义一个简单的网络结构
class Net(nn.Module):
	def __init__(self):
		super(Net, self).__init__()
		self.conv1 = nn.Sequential(nn.Conv2d(3, 3, 3), nn.Conv2d(3, 3, 3))
		self.conv2 = nn.Sequential(nn.Conv2d(3, 3, 3), nn.Conv2d(3, 3, 3))
	def forward(self, x):
		x = self.conv1(x)
		x = self.conv2(x)
		return x
net = Net()
# 实例化网络结构
for index, module in enumerate(net.modules()):
	print(index)
	print(module)

0
Net(
(conv1): Sequential(
(0): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
(1): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
)
(conv2): Sequential(
(0): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
(1): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
)
)
1
Sequential(
(0): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
(1): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
)
2
Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
3
Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
4
Sequential(
(0): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
(1): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
)
5
Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
6
Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))

for index, child in enumerate(net.children()):
	print(index)
	print(child)

0
Sequential(
(0): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
(1): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
)
1
Sequential(
(0): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
(1): Conv2d(3, 3, kernel_size=(3, 3), stride=(1, 1))
)