关于.state_dict()和.load_state_dict()分别是浅拷贝和深拷贝的问题

参考链接：https://blog.csdn.net/qq_43827595/article/details/115324118

https://blog.csdn.net/qq_41596915/article/details/122005538

关于深拷贝和浅拷贝的参考文章：

https://blog.csdn.net/weixin_50829653/article/details/127675849?spm=1001.2101.3001.6650.2&utm_medium=distribute.pc_relevant.none-task-blog-2%7Edefault%7EBlogCommendFromBaidu%7ERate-2-127675849-blog-127951831.235%5Ev43%5Epc_blog_bottom_relevance_base1&depth_1-utm_source=distribute.pc_relevant.none-task-blog-2%7Edefault%7EBlogCommendFromBaidu%7ERate-2-127675849-blog-127951831.235%5Ev43%5Epc_blog_bottom_relevance_base1&utm_relevant_index=3

https://blog.csdn.net/weixin_44133119/article/details/123307291

• model.state_dict()实际上是浅拷贝，如果令param=model.state_dict()，那么当你修改param，相应地也会修改model的参数。model这个对象实际上是指向各个参数矩阵的，而浅拷贝只会拷贝最外层的这些“指针”。
• model.load_state_dict(xxx) 是深拷贝

原理图：

​

model.state_dict()["conv1.weight"]指向tensor1

1. model.state_dict()["conv1.weight"].copy_(...)原地改变了tensor1的值，原model也受到影响
2. 相当于改变了箭头的指向，原model没有受到影响
3. model.state_dict()["conv1.weight"][0]=torch.tensor(0)相当于改变了第二层参数箭头的指向，原model受到影响

class Tudui(nn.Module):
    def __init__(self):
        super(Tudui, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=6, kernel_size=3, stride=1, padding=0)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(in_channels=6, out_channels=18, kernel_size=3, stride=1, padding=0)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        x = self.relu(x)
        return x

1、model.state_dict()["conv1.weight"].copy_(...)

a = Tudui()
m = a.state_dict()
for k, v in m.items():
    m[k].copy_(torch.zeros_like(v))
    break
f = a.state_dict()

​

可见原模型参数被改变了

2、model.state_dict()["conv1.weight"]=torch.tensor(0)

a = Tudui()
m = a.state_dict()
for k, v in m.items():
    m[k] = torch.tensor(0)
    break
f = a.state_dict()

​

可见原模型参数没有被改变

3、model.state_dict()["conv1.weight"][0]=torch.tensor(0)
 

a = Tudui()
m = a.state_dict()
for k, v in m.items():
    m[k][0] = torch.tensor(0)
    break
f = a.state_dict()

​

可见原模型参数被改变了