HappyBERT throws an error while running the Q&A example

[KiranAkadas]

Python 3.8.5

happy_bert = HappyBERT()
question = "Who does Ernie live with?"
text = "Ernie is an orange Muppet character on the long running PBS and HBO children's television show Sesame Street. He and his roommate Bert form the comic duo Bert and Ernie, one of the program's centerpieces, with Ernie acting the role of the naïve troublemaker and Bert the world weary foil."  # Source: https://en.wikipedia.org/wiki/Ernie_(Sesame_Street)
result = happy_bert.answer_question(question,text)
print(type(result)) # prints: <class 'str'>
print(result) # prints: bert

The example given in the official documentation gives the following error when run:

TypeError                                 Traceback (most recent call last)
<ipython-input-8-618d260e51a7> in <module>()
      2 question = "Who does Ernie live with?"
      3 text = "Ernie is an orange Muppet character on the long running PBS and HBO children's television show Sesame Street. He and his roommate Bert form the comic duo Bert and Ernie, one of the program's centerpieces, with Ernie acting the role of the naïve troublemaker and Bert the world weary foil."  # Source: https://en.wikipedia.org/wiki/Ernie_(Sesame_Street)
----> 4 result = happy_bert.answer_question(question,text)
      5 print(type(result)) # prints: <class 'str'>
      6 print(result) # prints: bert

/usr/local/lib/python3.6/dist-packages/happytransformer/happy_bert.py in answer_question(self, question, text)
    140                                                token_type_ids=segment_tensor)
    141         all_tokens = self.tokenizer.convert_ids_to_tokens(input_ids)
--> 142         answer_list = all_tokens[torch.argmax(start_scores):
    143                                  torch.argmax(end_scores)+1]
    144         answer = self.tokenizer.convert_tokens_to_string(answer_list)

TypeError: argmax(): argument 'input' (position 1) must be Tensor, not str

[ted537]

You say Python 3.8.5 but your path is at /usr/local/lib/python3.6

This is probably related to the dependencies in setup.py, as they are not versioned currently

[skymansandy]

Hey @ted537, @EricFillion. I faced this error too.

Looks like destructuring the output of the qa model into start_scores and end_scores is somehow causing them to be type string instead of tensor. Not sure if this has something to do with the python version being used?

For a temporary solution, I changed the library code a bit to make it run on my machine. Stored the qa model output in a scores and used indexing to get the scores. (scores[0], scores[1] for start and end scores).

PS: I'm on python 3.8.5.

[EricFillion]

Thank you @KiranAkadas and @skymansandy. I just created a PR to address this issue. Ted and I are both in the middle of exams, but we'll be sure to publish an update shortly.

[EricFillion]

I just published an update to PyPi that resolves this issue. I hope it helps :) @KiranAkadas @skymansandy