load model issue using pickle

[lizaitang]

When I am using python 3.8 to load the image net model , the code reports that File "/home/zaitang/anaconda3/envs/pytorch/lib/python3.9/site-packages/torch/serialization.py", line 762, in _legacy_load magic_number = pickle_module.load(f, **pickle_load_args) _pickle.UnpicklingError: invalid load key, '<'.

Could you please help to fix it?

[fra31]

Hi,

this is probably related. Please let me know if that helps to solve it.

[lizaitang]

Sorry, but still not fixed, I have tried the method.

[fra31]

Could you please try to delete also the previously downloaded checkpoint (if a file with the same name exists it doesn't get downloaded again)?

[lizaitang]

Sorry but I have also tried.

[lizaitang]

Thank you so much, I just solved it, for anyone with same issue, you need to uninstall robustbench first and reinstall with latest version. Thanks again!

[fra31]

Ok, good!

[lizaitang]

Thanks, another small issue is that I am using a imagenet dataset generated by GAN , to check the accuracy on the load_model , which preprocessing should be implemented?

[fra31]

I think one usually keeps the same preprocessing used during training. For the models in the Model Zoo we specify for each classifier the preprocessing (see here).

[lizaitang]

Thanks, which means we do not need to do preprocessing by ourselves? Just load the model?

[fra31]

For the models in the Model Zoo, the preprocessing is done automatically when loading the dataset. If you want to evaluate other models, you can either specify one of the predefined preprocessing methods here or pass a customized one with https://github.com/RobustBench/robustbench/blob/df316217f4bb929e96525044e3aca620238b6f16/robustbench/eval.py#L33 in the benchmark function.