如何将 model.state_dict() 存储在临时变量中以备后用？

Question

I tried to store the state dict of my model in a variable temporarily and wanted to restore it to my model later, but the content of this variable changed automatically as the model updated.

There is a minimal example:

import torch as t
import torch.nn as nn
from torch.optim import Adam


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.fc = nn.Linear(3, 2)

    def forward(self, x):
        return self.fc(x)


net = Net()
loss_fc = nn.MSELoss()
optimizer = Adam(net.parameters())

weights = net.state_dict()
print(weights)

x = t.rand((5, 3))
y = t.rand((5, 2))
loss = loss_fc(net(x), y)

optimizer.zero_grad()
loss.backward()
optimizer.step()

print(weights)

I thought the two outputs would be the same, but I got (outputs may change due to random initialization)

OrderedDict([('fc.weight', tensor([[-0.5557,  0.0544, -0.2277],
        [-0.0793,  0.4334, -0.1548]])), ('fc.bias', tensor([-0.2204,  0.2846]))])
OrderedDict([('fc.weight', tensor([[-0.5547,  0.0554, -0.2267],
        [-0.0783,  0.4344, -0.1538]])), ('fc.bias', tensor([-0.2194,  0.2856]))])

The content of weights changed, which is so weird.

I also tried .copy() and t.no_grad() as following, but they did not help.

with t.no_grad():
    weights = net.state_dict().copy()

Yes, I know that I can save state dict using t.save() , but I just want to figure out what happened in the previous example.

I'm using Python 3.8.5 and Pytorch 1.8.1

Thanks for any help.

Answer 1

That's how OrderedDict works. Here's a simpler example:

from collections import OrderedDict

# a mutable variable
l = [1,2,3]

# an OrderedDict with an entry pointing to that mutable variable
x = OrderedDict([("a", l)])

# if you change the list
l[1] = 20

# the change is reflected in the OrderedDict
print(x)
# >> OrderedDict([('a', [1, 20, 3])])

If you want to avoid that, you'll have to do a deepcopy rather than a shallow copy :

from copy import deepcopy
x2 = deepcopy(x)

print(x2)
# >> OrderedDict([('a', [1, 20, 3])])

# now, if you change the list
l[2] = 30

# you do not change your copy
print(x2)
# >> OrderedDict([('a', [1, 20, 3])])

# but you keep changing the original dict
print(x)
# >> OrderedDict([('a', [1, 20, 30])])

As Tensor is also mutable, the same behaviour is expected in your case. Therefore, you can use:

from copy import deepcopy

weights = deepcopy(net.state_dict())