net网络查看其参数state_dict，data，named_parameters

jjw_zyfx

已于 2022-03-04 11:58:38 修改

阅读量1.1k

点赞数 4

分类专栏： python pytorch 文章标签： python pytorch 深度学习计算机视觉机器学习

于 2022-03-04 11:14:22 首次发布

本文链接：https://blog.csdn.net/jjw_zyfx/article/details/123272065

版权

PyTorch nn.Linear 权重初始化偏置初始化自定义初始化

关键词由CSDN通过智能技术生成

python 同时被 2 个专栏收录

125 篇文章 2 订阅

订阅专栏

pytorch

69 篇文章 8 订阅

订阅专栏

代码如下：

net = nn.Linear(2, 2)
print(net.state_dict())

输出结果：

OrderedDict([
('weight', tensor([[ 0.6155, -0.4649],
                   [-0.1848, -0.0663]])), 
('bias', tensor([-0.0265, -0.4134]))])

从输出结果可以看出，在适用nn.Linear的时候会自动给weight和bias赋初值，点进Linear后可以发现

Attributes:
weight: the learnable weights of the module of shape :math: $out_features , in_features ) (\text{out\_features}, \text{in\_features})$ . The values are initialized from :
$in_features math: \mathcal{U}(-\sqrt{k}, \sqrt{k}), where math:k = \frac{1}{\text{in\_features}}$
bias: the learnable bias of the module of shape :math: $out_features ) (\text{out\_features})$ .
If :attr:bias is True, the values are initialized from
$in_features math:\mathcal{U}(-\sqrt{k}, \sqrt{k}) where math:k = \frac{1}{\text{in\_features}}$
即nn.Linear会自动给初始化

自定义初始化方法如下：


net = nn.Sequential(nn.Flatten(), nn.Linear(784, 10))
#nn.Flatten将多维张量压缩为二维（一行）进而进行线性操作、
 
def init_weights(m):
    if type(m) == nn.Linear:
        nn.init.normal_(m.weight, std=0.01)
 
net.apply(init_weights);

其他参数获取方法：

net = nn.Linear(2, 2)
print('net.state_dict()\n', net.state_dict())
print('net.bias\n', net.bias)
print('net.bias.data\n', net.bias.data)
print('net.bias.grad\n', net.bias.grad)  # 因为没有backward所以没有赋梯度值
print('net.named_parameters()\n', net.named_parameters())
print([(name, param) for name, param in net.named_parameters()])

输出结果：

net.state_dict()
 OrderedDict([('weight', tensor([[-0.6907, -0.4128],
        [-0.4212,  0.0620]])), ('bias', tensor([-0.2019, -0.3731]))])
net.bias
 Parameter containing:
tensor([-0.2019, -0.3731], requires_grad=True)
net.bias.data
 tensor([-0.2019, -0.3731])
net.bias.grad
 None
net.named_parameters()
 <generator object Module.named_parameters at 0x151b22ba0>
[('weight', Parameter containing:
tensor([[-0.6907, -0.4128],
        [-0.4212,  0.0620]], requires_grad=True)), ('bias', Parameter containing:
tensor([-0.2019, -0.3731], requires_grad=True))]