查看PyTorch网络模型结构，提取参数

nn.Module里面有两个特别重要的关于参数的属性，分别是named_parameters()和parameters()。named_parameters()是给出网络层的名字和参数的迭代器，parameters()会给出一个网络的全部参数的迭代器。

• 举例说明：

import torch
import torch.nn as nn
from torch.autograd import Variable


class Model(nn.Module):
    def __init__(self, c_h, c_out):
        super(Model, self).__init__()
        self.conv2d = nn.Conv2d(c_h, c_out, kernel_size=(2, 3))
        self.conv1d = nn.Conv1d(c_h, c_out, kernel_size=1)
        self.linear = nn.Linear(c_h, c_out)

        # 设置可学习参数
        self.weight_1 = nn.Parameter(torch.FloatTensor(1), requires_grad=True)

        # init
        self.weight_1.data.fill_(0.25)

        self.value = torch.tensor(1)
        self.test = Variable(self.value)


model = Model(3, 4)

"""
nn.Module里面有两个特别重要的关于参数的属性，分别是named_parameters()和parameters()。
named_parameters()是给出网络层的名字和参数的迭代器，parameters()会给出一个网络的全部参数的迭代器。
"""

def print_network(model, name):
    """Print out the network information."""
    num_params = 0
    print(model)
    print(name)
    for p in model.parameters():
        num_params += p.numel()  # numel()获取tensor中一共包含多少个元素
    print("The number of parameters: {}".format(num_params))

    for p in model.named_parameters():
        print(len(p))
        print(p)


if __name__ == '__main__':
    print_network(model, 'm')

• 结果

Model(
  (conv2d): Conv2d(3, 4, kernel_size=(2, 3), stride=(1, 1))
  (conv1d): Conv1d(3, 4, kernel_size=(1,), stride=(1,))
  (linear): Linear(in_features=3, out_features=4, bias=True)
)
m
The number of parameters: 109
2
('weight_1', Parameter containing:
tensor([0.2500], requires_grad=True))
2
('conv2d.weight', Parameter containing:
tensor([[[[ 0.1952,  0.1276,  0.1480],
          [ 0.0232,  0.0853,  0.0471]],

         [[ 0.2219, -0.1046,  0.0738],
          [-0.0163, -0.0100,  0.1927]],

         [[ 0.1709,  0.0461,  0.0137],
          [ 0.1762,  0.0511,  0.0319]]],


        [[[-0.2333,  0.0426, -0.0946],
          [ 0.0997,  0.0225,  0.0358]],

         [[-0.1744,  0.0281, -0.1157],
          [ 0.0461,  0.1030, -0.1360]],

         [[ 0.2107,  0.1877,  0.1386],
          [-0.0401, -0.0251, -0.0446]]],


        [[[ 0.1975, -0.0323,  0.0199],
          [ 0.0722, -0.1193, -0.2274]],

         [[ 0.2100,  0.1317,  0.2273],
          [ 0.0426, -0.0987, -0.2054]],

         [[ 0.1995, -0.1346,  0.0621],
          [ 0.0949,  0.0104, -0.0909]]],


        [[[-0.1033, -0.0297, -0.0506],
          [ 0.1811, -0.1832,  0.1673]],

         [[ 0.2206, -0.0414,  0.0783],
          [ 0.0911, -0.0315, -0.1597]],

         [[-0.0766,  0.2039,  0.1840],
          [-0.0629, -0.0137, -0.2211]]]], requires_grad=True))
2
('conv2d.bias', Parameter containing:
tensor([0.2280, 0.1339, 0.0501, 0.0747], requires_grad=True))
2
('conv1d.weight', Parameter containing:
tensor([[[ 0.3422],
         [-0.3975],
         [ 0.3537]],

        [[-0.5534],
         [ 0.5642],
         [-0.1036]],

        [[ 0.2595],
         [-0.5371],
         [ 0.3225]],

        [[ 0.0061],
         [ 0.3125],
         [ 0.2301]]], requires_grad=True))
2
('conv1d.bias', Parameter containing:
tensor([ 0.0465, -0.3245,  0.4590, -0.4186], requires_grad=True))
2
('linear.weight', Parameter containing:
tensor([[-0.4758,  0.1002, -0.3299],
        [-0.2612,  0.4921, -0.2546],
        [ 0.2926, -0.1463,  0.0831],
        [ 0.2100, -0.0217,  0.1397]], requires_grad=True))
2
('linear.bias', Parameter containing:
tensor([-0.1661,  0.5512, -0.0258,  0.3975], requires_grad=True))