fix issue when saving Bert model

[CRGONCALVES]

I have an issue when saving the Bert Model with the code below :

import joblib
_ = joblib.dump(nn_model,"./data/nn_model.pickle",compress=True)

I fix this issue by adding the code : if "embedding_matrix"] in dict_attr: in the file train.py (see below) :

def __getstate__(self):
        """Method called before serialization for a specific treatment to save
        model weight and structure instead of standard serialization."""
        dict_attr = dict(self.__dict__)
        if "model" in dict_attr:
            del dict_attr["model"]
            if "embedding_matrix"] in dict_attr:
                del dict_attr["embedding_matrix"]
            del dict_attr["pretrained_embedding"]
        return dict_attr

Test Configuration:

• OS: windows 10
• Python version: 3.8.8
• Melusine version: 2.3.2

[CRGONCALVES]

import joblib _ = joblib.dump(CamemBert_model,"./data/CamemBert_model.pickle",compress=True)

[hugo-quantmetry]

Hello Carlos, Thank you for your contribution! I added some comments:

• a minor one to fix indentation in train.py
• a larger issue concerning files specific to you (or your company) that have been pushed to the open-source package

[TFA-MAIF]

Fixed PR has been merge.