Difference in saved and loaded state dict

vision
1

Hi,

I was just wondering, there is a small difference whenever we save state dictionary of a model and reload it back. Can you please point if I am doing some mistake or if this is an insignificant issue?

Regards,
P.S: I have pasted sample code below

from torchvision import models
import torch

model = models.resnet50(pretrained=True)
model.eval()

stateDict = model.state_dict()
torch.save(model.state_dict(), “test.pth”)
reloadDict = torch.load(“test.pth”)

diff = stateDict.items() - reloadDict.items()

2

If I’m not mistaken, subtracting dict.items() would try to remove duplicated entries as seen here:

d1 = {'a': 1, 'b': 2, 'c': 10}
d2 = {'a': 3, 'b': 4}
print(d1.items() - d2.items())
> {('a', 1), ('b', 2), ('c', 10)}

d1 = {'a': 1, 'b': 2, 'c': 10}
d2 = {'a': 3, 'b': 4, 'c': 10}
print(d1.items() - d2.items())
> {('a', 1), ('b', 2)}

so compare the values via:

model = models.resnet50(pretrained=True)
model.eval()

stateDict = model.state_dict()
torch.save(model.state_dict(), 'test.pth')
reloadDict = torch.load('test.pth')

diff = stateDict.items() - reloadDict.items()
len(diff)

for key in stateDict:
    paramA = stateDict[key]
    paramB = reloadDict[key]
    print(key, (paramA - paramB).abs().max())

which will show a zero difference.

1 Like
3

Thanks a bunch, I made that silly assumption, will be careful next time in asking question!
Thanks a heap!