model.copy() is broken

[bwasti]

🐛 Bug

Calling a scripted model.copy() is broken. Throws an error here:

RuntimeError: class '__torch__.torch.nn.modules.module.___torch_mangle_5.Module' already defined. (register_type at caffe2/torch/csrc/jit/script/compilation_unit.h:167)

To Reproduce

Steps to reproduce the behavior:

1. create a scripted model
2. call model.copy()

Expected behavior

Expected a copy to be made.

• PyTorch Version (e.g., 1.0): master
• OS (e.g., Linux): linus
• How you installed PyTorch (conda, pip, source): source
• Build command you used (if compiling from source): setup.py
• Python version: 3.7

cc @ezyang @gchanan @zou3519 @jerryzh168 @suo

[vadimkantorov]

Also it seems that the copy method is not documented: https://pytorch.org/docs/master/search.html?q=copy&check_keywords=yes&area=default seems to not show any method named copy()

[zou3519]

I don't think model.copy() exists or is supposed to exist on nn modules. This issue seems to be JIT-specific so I am tossing it onto the JIT queue.

[suo]

Assigning to @ZolotukhinM as I believe he implemented copy() (feel free to bounce back to me if I'm wrong about that). Can you take a look?

[ZolotukhinM]

I think @jerryzh168 fixed that in #29269

[jerryzh168]

@bwasti can you provide an example? I can't reproduce with random scripted models

[jerryzh168]

@bwasti my PR is landed, could you verify if this is resolved?

[jerryzh168]

I'll close for now, feel free to re-open if the problem persists