QA Documentation: I got error just copy and pasting documentation

[andreabac3]

Environment info

• transformers version: 4.3.1
• Platform:Manjaro Linux
• Python version: 1.5.1
• PyTorch version (GPU?): Yes
• Using GPU in script?: Yes
• Using distributed or parallel set-up in script?: No

Who can help

@sgugger

Information

I am trying the to train a QA model following the huggingface documentation, I just copied and pasted the code in my machine (and in Colab) but I was not able to proceed in the training phase because I got None value.

To reproduce

Steps to reproduce the behavior:

1. Go to the documentation: https://huggingface.co/transformers/custom_datasets.html at Squad training section
2. Copy and paste the code as you can see from my pastebin: https://pastebin.com/hZvq7Zs7
3. And you got the following error
   File "/home/andrea/PycharmProjects/qa-srl/test.py", line 78, in __getitem__ return {key: torch.tensor(val[idx]) for key, val in self.encodings.items()} RuntimeError: Could not infer dtype of NoneType
4. My naive solution was modifying the getitem method from the SquadDataset class in order to avoid to serve the val[idx] == None

[sgugger]

Pinging @joeddav on this one, since he wrote this tutorial :-)

[andreabac3]

Thank you @sgugger for the reply.
Ok I can wait for the answer from @joeddav.

Have a nice day.

[joeddav]

Figured it out. answer_end is the character position immediately after the answer, so end_position should be derived from answer_end - 1. I'm not sure why I was able to run it without this error previously (perhaps a resolved tokenizer bug?), but this should be correct.

def add_token_positions(encodings, answers):
    start_positions = []
    end_positions = []
    for i in range(len(answers)):
        start_positions.append(encodings.char_to_token(i, answers[i]['answer_start']))
        end_positions.append(encodings.char_to_token(i, answers[i]['answer_end'] - 1))
 
        # if start position is None, the answer passage has been truncated
        if start_positions[-1] is None:
            start_positions[-1] = tokenizer.model_max_length
            end_positions[-1] = tokenizer.model_max_length

    encodings.update({'start_positions': start_positions, 'end_positions': end_positions})

[LysandreJik]

Closed by #10217

[andreabac3]

Thank you @joeddav the posted code works perfectly.

[andreabac3]

Sorry for bothering you @joeddav again, I have a question related to the code posted by you here.
I am still getting None with the dataset built by myself using this code. My dataset works perfectly with the run_squad original script.
In this snipped posted by you I encounter None in the vector of end_positions and I don't know how fix it. I saw the condition in which there's a None the start_positions but what I have to do in the case the None is only in the end_positions vector?

Kind regards,
Andrea