Pytorch提取神经网络层结构、层参数及自定义初始化_pytorch 信息抽取模型

1. 提取层结构

对于一个给定的模型，如果不想模型中所有的层结构，只希望能够提取网络中的某一层或者某几层，应该如何实现呢？

首先定义一下神经网络结构：

import torch
import torch.nn as nn


class SimpleCnn(nn.Module):
    def __init__(self):
        super(SimpleCnn, self).__init__()

        layer1 = nn.Sequential()
        layer1.add_module("conv1", nn.Conv2d(3, 32, 3, 1, padding=1))
        layer1.add_module("relu1", nn.ReLU(True))
        layer1.add_module("pool1", nn.MaxPool2d(2, 2))
        self.layer1 = layer1

        layer2 = nn.Sequential()
        layer2.add_module("conv2", nn.Conv2d(32, 64, 3, 1, padding=1))
        layer2.add_module("relu2", nn.ReLU(True))
        layer2.add_module("pool2", nn.MaxPool2d(2, 2))
        self.layer2 = layer2

        layer3 = nn.Sequential()
        layer3.add_module("conv3", nn.Conv2d(64, 128, 3, 1, padding=1))
        layer3.add_module("relu3", nn.ReLU(True))
        layer3.add_module("pool3", nn.MaxPool2d(2, 2))
        self.layer3 = layer3

        layer4 = nn.Sequential()
        layer4.add_module("fc1", nn.Linear(2048, 512))
        layer4.add_module("fc_relu1", nn.ReLU(True))
        layer4.add_module("fc2", nn.Linear(512, 64))
        layer4.add_module("fc_relu2", nn.ReLU(True))
        layer4.add_module("fc3", nn.Linear(64, 10))
        self.layer4 = layer4

    def forward(self, x):
        conv1 = self.layer1(x)
        conv2 = self.layer2(conv1)
        conv3 = self.layer3(conv2)
        fc_input = conv3.view(conv3.shape[0], -1)
        fc_out = self.layer4(fc_input)

        return fc_out
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42

再来看一下nn.Module的几个重要属性

1.1 children()

children()：返回下一级模块的迭代器。比如在上述定义的神经网络模型中，它只会返回self.layer1、self.layer2、self.layer3以及self.layer4以上的迭代器，不会返回它们内部的东西。

1. 代码示例1：model.children()返回的是一个生成器

model = SimpleCnn()
print(model.children())
# 输出：<generator object Module.children at 0x7f84932e2050>
1
2
3

2. 代码示例2：对model.children()进行迭代，提取神经网络下一级模块

model = SimpleCnn()
for layer in model.children():
    print(layer)
1
2
3

Sequential(
  (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu1): ReLU(inplace=True)
  (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
Sequential(
  (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu2): ReLU(inplace=True)
  (pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
Sequential(
  (conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu3): ReLU(inplace=True)
  (pool3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
Sequential(
  (fc1): Linear(in_features=2048, out_features=512, bias=True)
  (fc_relu1): ReLU(inplace=True)
  (fc2): Linear(in_features=512, out_features=64, bias=True)
  (fc_relu2): ReLU(inplace=True)
  (fc3): Linear(in_features=64, out_features=10, bias=True)
)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

3. 代码示例3：对下一级模块继续进行迭代，得到下下级模块。

model = SimpleCnn()
for layer in model.children():
    for sublayer in layer.children():
        print(sublayer)
1
2
3
4

Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace=True)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace=True)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace=True)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
Linear(in_features=2048, out_features=512, bias=True)
ReLU(inplace=True)
Linear(in_features=512, out_features=64, bias=True)
ReLU(inplace=True)
Linear(in_features=64, out_features=10, bias=True)
1
2
3
4
5
6
7
8
9
10
11
12
13
14

4. 如果想提取神经网络的前两层，实现如下：

model = SimpleCnn()
new_model = nn.Sequential(*list(model.children())[0:2])
print(new_model)
1
2
3

Sequential(
  (0): Sequential(
    (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu1): ReLU(inplace=True)
    (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (1): Sequential(
    (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu2): ReLU(inplace=True)
    (pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
)
1
2
3
4
5
6
7
8
9
10
11
12

1.2 modules()

modules会返回模型中的所有模块的迭代器，这样就有一个好处，即它能访问到最内层，比如self.layer1.conv1这个模块。输出结果：返回神经网络的所有模块，包括各种不同级的模块。注意与1.1中代码示例1和代码示例2的区别。

model = SimpleCnn()
for layer in model.modules():
    print(layer)
1
2
3

SimpleCnn(
  (layer1): Sequential(
    (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu1): ReLU(inplace=True)
    (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (layer2): Sequential(
    (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu2): ReLU(inplace=True)
    (pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (layer3): Sequential(
    (conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu3): ReLU(inplace=True)
    (pool3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (layer4): Sequential(
    (fc1): Linear(in_features=2048, out_features=512, bias=True)
    (fc_relu1): ReLU(inplace=True)
    (fc2): Linear(in_features=512, out_features=64, bias=True)
    (fc_relu2): ReLU(inplace=True)
    (fc3): Linear(in_features=64, out_features=10, bias=True)
  )
)
Sequential(
  (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu1): ReLU(inplace=True)
  (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace=True)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
Sequential(
  (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu2): ReLU(inplace=True)
  (pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace=True)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
Sequential(
  (conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu3): ReLU(inplace=True)
  (pool3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
ReLU(inplace=True)
MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
Sequential(
  (fc1): Linear(in_features=2048, out_features=512, bias=True)
  (fc_relu1): ReLU(inplace=True)
  (fc2): Linear(in_features=512, out_features=64, bias=True)
  (fc_relu2): ReLU(inplace=True)
  (fc3): Linear(in_features=64, out_features=10, bias=True)
)
Linear(in_features=2048, out_features=512, bias=True)
ReLU(inplace=True)
Linear(in_features=512, out_features=64, bias=True)
ReLU(inplace=True)
Linear(in_features=64, out_features=10, bias=True)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

1.3 named_children()和named_modules()

named_children()、named_modules()分别与children()、modules()相对应，其不仅返回模块的迭代器，还会返回网络层的名字。

1. 代码示例1

model = SimpleCnn()
for layer in model.named_modules():
    print(layer)
1
2
3

('', SimpleCnn(
  (layer1): Sequential(
    (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu1): ReLU(inplace=True)
    (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (layer2): Sequential(
    (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu2): ReLU(inplace=True)
    (pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (layer3): Sequential(
    (conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
    (relu3): ReLU(inplace=True)
    (pool3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
  )
  (layer4): Sequential(
    (fc1): Linear(in_features=2048, out_features=512, bias=True)
    (fc_relu1): ReLU(inplace=True)
    (fc2): Linear(in_features=512, out_features=64, bias=True)
    (fc_relu2): ReLU(inplace=True)
    (fc3): Linear(in_features=64, out_features=10, bias=True)
  )
))
('layer1', Sequential(
  (conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu1): ReLU(inplace=True)
  (pool1): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
))
('layer1.conv1', Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))
('layer1.relu1', ReLU(inplace=True))
('layer1.pool1', MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))
('layer2', Sequential(
  (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu2): ReLU(inplace=True)
  (pool2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
))
('layer2.conv2', Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))
('layer2.relu2', ReLU(inplace=True))
('layer2.pool2', MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))
('layer3', Sequential(
  (conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (relu3): ReLU(inplace=True)
  (pool3): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
))
('layer3.conv3', Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)))
('layer3.relu3', ReLU(inplace=True))
('layer3.pool3', MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False))
('layer4', Sequential(
  (fc1): Linear(in_features=2048, out_features=512, bias=True)
  (fc_relu1): ReLU(inplace=True)
  (fc2): Linear(in_features=512, out_features=64, bias=True)
  (fc_relu2): ReLU(inplace=True)
  (fc3): Linear(in_features=64, out_features=10, bias=True)
))
('layer4.fc1', Linear(in_features=2048, out_features=512, bias=True))
('layer4.fc_relu1', ReLU(inplace=True))
('layer4.fc2', Linear(in_features=512, out_features=64, bias=True))
('layer4.fc_relu2', ReLU(inplace=True))
('layer4.fc3', Linear(in_features=64, out_features=10, bias=True))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

2. 如果希望提取模型中所有的卷积层，实现如下：

model = SimpleCnn()
new_model = nn.Sequential()
for layer in model.named_modules():
    if isinstance(layer[1], nn.Conv2d):
        conv_name = layer[0].replace('.', '_')
        new_model.add_module(conv_name, layer[1])

print(new_model)
1
2
3
4
5
6
7
8

Sequential(
  (layer1_conv1): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (layer2_conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (layer3_conv3): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
)
1
2
3
4
5

2.提取参数及自定义初始化

2.1 提取参数

nn.Module里有两个关于参数的属性：

• parameters()：给出一个网络的全部参数的迭代器
• named_parameters()：给出网络层的名字和参数的迭代器

1. 代码示例1：获取模型的参数。还可以通过para.grad获取梯度

model = SimpleCnn()
for para in model.parameters():
    print(para)
1
2
3

2. 代码示例2：获取模型网络层的名字和参数。model.named_parameters()迭代出形式为（名字，参数值）的元组。

model = SimpleCnn()
for name, para in model.named_parameters():
    print(name)
1
2
3

2.2 自定义初始化

model = SimpleCnn()
for m in model.modules():
    if isinstance(m, nn.Conv2d):
        # nn.init.normal_(m.weight.data)
        # nn.init.xavier_normal_(m.weight.data)
        nn.init.kaiming_normal_(m.weight.data)
        m.bias.data.fill_(0)
    elif isinstance(m, nn.Linear):
        m.weight.data.normal_()
1
2
3
4
5
6
7
8
9