Is it possible to load a haystack-generated model in transformers

[Avs-safety]

Hi,

I've fine-tuned a BERT model on my own SQuAD-like data using haystack, which results in several files (e.g. language_model.bin, prediction_head_0.bin, etc). Does anyone know if there is a way to load the model via a transformers pipeline? It doesn't work when I try unless I rename some of the files (e.g. language_model.bin to model.bin), however I assume this then ignores the prediction head...

for example loading via:

from transformers import pipeline

question_answerer = pipeline("question-answering", model=r"file_path")

Appreciate any help!

[MichelBartels]

There already exists a save_to_remote method in the FARMReader class which uploads a model to the HF hub.
If you want to save it locally, you would manually need to save both the model and the tokenizer into the same folder:

transformer_model = reader.inferencer.model.convert_to_transformers()[0]
reader.inferencer.processor.tokenizer.save_pretrained(your_save_path)
transformer_model.save_pretrained(your_save_path)

[Avs-safety]

That's great, thank you!