Missing key(s) in state_dict about transformersum

[HHousen]

@MorenoLaQuatra I've pushed a change that removes the pinned version. Old models can now be loaded by specifying the --no_strict argument.

[HHousen]

@MorenoLaQuatra Did you install the requirements from the environment.yml file with conda env create --file environment.yml? Running that should have installed transformers==3.0.2. If it didn't you can try manually installing version 3.0.2 with pip install -U transformers==3.0.2, which should fix the issue.

[MorenoLaQuatra]

Thank you for the effort in fixing this.

[MorenoLaQuatra]

Yes I did. Actually I get an error while executing conda env create --file environment.yml. The environment was created but some libraries were missing (I installed them separately).

Installing transformers==3.0.2 solved the issue, do you have any plan to be version-independent?

[HHousen]

I do plan to make the project version-independent. In fact, TransformerSum will work with the latest version of transformers, it just cannot load models trained with versions <3.0.2 on versions >3.0.2. I'm currently not sure how to accomplish this. I might have to rebuild all of the pre-trained model files. I'm looking into it.

[HHousen]

I might be able to load models using the .from_pretrained() function from transformers and then resave them.