Loading pretrained model in Python 3.8 fails about robustness

[vietvo89]

Hello @andrewilyas and @rsokl

I got the same issue when loading model l2 0.25 and 1.0. But loading model l2 0.5 and l_inf_8 has no problem. I followed your recommendation to downgrade dill 0.3.4 to 0.3.2, but it did not work. I am using python 3.8. Please do me a favor!

Thanks

[andrewilyas]

Hi! Can you please indicate what version of dill is in each environment?

[rsokl]

Sure! Both versions are using

dill                      0.3.2                    pypi_0    pypi

[andrewilyas]

Sorry forgot to follow up on this! Can you also post your torch version on the 3.7 environment? Is it the same?

[rsokl]

It is the same - 1.6.0

Both environments have the same library versions. Only the Python version is different

[andrewilyas]

Ok looks like this is a known issue with dill (uqfoundation/dill#357), but it was patched in the latest version. If you have both 3.7 and 3.8 on your machine, you can fix this yourself by doing:

1. Update dill to 0.3.2 in both versions
2. In Python 3.7, execute:

import torch, dill
x = torch.load("cifar_l2_1_0.pt", pickle_module=dill)
torch.save(x, "cifar_l2_1_0.pt", pickle_module=dill)

3. In Python 3.8, you should now be able to load the checkpoint using your attached code.

Meanwhile, we will do this for all the links we have above so that the public versions are also cross-compatible.

[rsokl]

Great, I will give that a shot. Thank you for your help!

Closing this assuming that the above route will work without a hitch :)