self.modules() 和 self.children()的区别

代码:

import torch
import torch.nn as nn

input_nc = 3
hide1_nc = 6
hide2_nc = 6
output_nc = 3

class Net(nn.Module):
    def __init__(self, input_nc, hide1_nc, hide2_nc, output_nc):
        super().__init__()

        acti = nn.ReLU(inplace=True)
        layer_add = nn.Sequential(nn.Linear(hide1_nc, hide2_nc),
                                  nn.Linear(hide1_nc, hide2_nc))
        self.layer1 = nn.Sequential(
            nn.Linear(input_nc, hide1_nc),
            acti,
            layer_add,
            acti
        )

        self.layer2 = nn.Sequential(
            nn.Linear(hide1_nc, hide2_nc),
            nn.ReLU(inplace=True))

        self.layer3 = nn.Linear(hide2_nc, output_nc)

        print('self.children()测试')
        for i, j in enumerate(self.children()):
            print("第{}次循环\n{}".format(i, j))
            print('*' * 20)

        print('self.modules()测试')
        for i, j in enumerate(self.modules()):
            print("第{}次循环\n{}".format(i, j))
            print('*' * 20)

    def forward(self, x):
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)

        return x

model = Net(input_nc, hide1_nc, hide2_nc, output_nc)

网络结构:

运行结果:

self.children()测试
第0次循环
Sequential(
  (0): Linear(in_features=3, out_features=6, bias=True)
  (1): ReLU(inplace)
  (2): Sequential(
    (0): Linear(in_features=6, out_features=6, bias=True)
    (1): Linear(in_features=6, out_features=6, bias=True)
  )
  (3): ReLU(inplace)
)
********************
第1次循环
Sequential(
  (0): Linear(in_features=6, out_features=6, bias=True)
  (1): ReLU(inplace)
)
********************
第2次循环
Linear(in_features=6, out_features=3, bias=True)
********************
self.modules()测试
第0次循环
Net(
  (layer1): Sequential(
    (0): Linear(in_features=3, out_features=6, bias=True)
    (1): ReLU(inplace)
    (2): Sequential(
      (0): Linear(in_features=6, out_features=6, bias=True)
      (1): Linear(in_features=6, out_features=6, bias=True)
    )
    (3): ReLU(inplace)
  )
  (layer2): Sequential(
    (0): Linear(in_features=6, out_features=6, bias=True)
    (1): ReLU(inplace)
  )
  (layer3): Linear(in_features=6, out_features=3, bias=True)
)
********************
第1次循环
Sequential(
  (0): Linear(in_features=3, out_features=6, bias=True)
  (1): ReLU(inplace)
  (2): Sequential(
    (0): Linear(in_features=6, out_features=6, bias=True)
    (1): Linear(in_features=6, out_features=6, bias=True)
  )
  (3): ReLU(inplace)
)
********************
第2次循环
Linear(in_features=3, out_features=6, bias=True)
********************
第3次循环
ReLU(inplace)
********************
第4次循环
Sequential(
  (0): Linear(in_features=6, out_features=6, bias=True)
  (1): Linear(in_features=6, out_features=6, bias=True)
)
********************
第5次循环
Linear(in_features=6, out_features=6, bias=True)
********************
第6次循环
Linear(in_features=6, out_features=6, bias=True)
********************
第7次循环
Sequential(
  (0): Linear(in_features=6, out_features=6, bias=True)
  (1): ReLU(inplace)
)
********************
第8次循环
Linear(in_features=6, out_features=6, bias=True)
********************
第9次循环
ReLU(inplace)
********************
第10次循环
Linear(in_features=6, out_features=3, bias=True)
********************

由运行结果可知：self.children()只包括网络模块的第一代儿子模块；self.modules()包含网络模块的自己本身和所有后代模块。

• self.children()存储网络结构的子层模块，即网络结构的第二列．
• self.modules()采用深度优先遍历的方式，存储了net的所有模块．值得注意的是self.modules()运行结果的第7次循环并不是nn.ReLU, 因为layer1中的acti = nn.ReLU(inplace=True)是重复的模块，有单独的命名，是类nn.ReLU()的一个固定实例,对于这种重复模块self.modules()中只返回一次; 而从self.modules()运行结果的第5和6次循环可以看出, 认为这两个nn.Linear()不是重复的模块，估计是看成类nn.Linear()不同的实例了

 

参考博客:

https://blog.csdn.net/dss_dssssd/article/details/83958518

https://blog.csdn.net/LXX516/article/details/79016980