For DPR models, transformers overwrites the model_type parameter specified in a model's config when saving the model:
https://github.com/huggingface/transformers/blob/64e78564a519cda2b4408803e2781c604e1e3bdd/src/transformers/configuration_utils.py#L626
For DPR models with a non-BERT tokenizer, e.g., CamembertTokenizer as in haystack issue #1046 this prevents loading the correct tokenizer when the model is loaded again. For example, the model_type camembert is specified in the config when loading the model from the model hub, but the model_type dpr is saved.
To overcome this problem, we set the model_type of transformers.DPRConfig on-the-fly to the model_type as specified in the model's config file.
Timoeller
commented
3 years ago
For DPR models, transformers overwrites the model_type parameter specified in a model's config when saving the model: https://github.com/huggingface/transformers/blob/64e78564a519cda2b4408803e2781c604e1e3bdd/src/transformers/configuration_utils.py#L626 For DPR models with a non-BERT tokenizer, e.g., CamembertTokenizer as in haystack issue #1046 this prevents loading the correct tokenizer when the model is loaded again. For example, the model_type camembert is specified in the config when loading the model from the model hub, but the model_type dpr is saved. To overcome this problem, we set the model_type of transformers.DPRConfig on-the-fly to the model_type as specified in the model's config file.
This PR corresponds to haystack PR #1060