State_dict() round trip fails

#1

I tried the following:

torch.save(net.state_dict(), f"file")
weights = torch.load('file1', map_location=device)

net1 = Net()
net1.load_state_dict(weights)

net2 = Net()
net2.load_state_dict(weights)

torch.save(net1.state_dict(), f"file1")
torch.save(net2.state_dict(), f"file2")

Then I find that file1 and file2 are different. How can it be? My goal is to find out if the weights of a net has changed since last time.

#2

Are you seeing different parameters in file1 and file2 after reloading them?
If so, could you post the model architecture so that we could reproduce this issue?

#3

I am seeing the same parameters. In fact, I figured out. torch.save appears to add some randomness to the byte representation it writes out, similar to salting.