Deepcopy of nn.module changes type of its parameters to tensors?

#1

I noticed deepcopying a module causes its parameters() to be tensors rather than nn.Parameters.

import torch.nn
import copy

l = torch.nn.Linear(3,1)
c = copy.deepcopy(l)
print([type(p) for p in l.parameters()])
print([type(p) for p in c.parameters()])

[<class 'torch.nn.parameter.Parameter'>, <class 'torch.nn.parameter.Parameter'>]
[<class 'torch.Tensor'>, <class 'torch.Tensor'>]

Why does this happen and can this cause any problems when working with the copy of the module later on?

#2

Thanks for reporting it!
It might be this old bug again.

#3

@hdkgr what PyTorch version are you on? print(torch.__version__)

#4

I could reproduce this issue in 1.0.0.dev20181007 and 1.0.0a0+dfad8b6.

#5

I don’t have the system at hand, but I’m quite sure it’s the 0.4.1 stable build.

#6

This looks like a regression in 0.4.0 / 0.4.1, We reopened the issue and an engineer is working on issuing a fix.

1 Like