Enabling Tapex in table question answering pipeline.
#16663
Conversation
| self.aggregate = bool(getattr(self.model.config, "aggregation_labels", None)) and bool( | ||
| getattr(self.model.config, "num_aggregation_labels", None) | ||
| ) | ||
| self.type = "tapas" if hasattr(self.model.config, "aggregation_labels") else None |
There was a problem hiding this comment.
Wondering if we can leverage model.config.is_encoder_decoder here (which is True for TAPEX but False for TAPAS)
There was a problem hiding this comment.
Usually I would agree that reusing more "general" concepts is better.
The thing I want to avoid is treating not encoder-decoder as tapas. currently the code for Tapas is super highly specific, so I'd rather isolate tapas than isolate tapex if you want.
Ultimately both are OK tbh.
| if truncation is None: | ||
| if self.type == "tapas": | ||
| truncation = "drop_rows_to_fit" | ||
| else: | ||
| truncation = "do_not_truncate" |
There was a problem hiding this comment.
TAPEX also supports the "drop_rows_to_fit" truncation strategy.
There was a problem hiding this comment.
The tokenizer said it didn't and gave me the regular list.
Is the tokenizer that gets resolved a BartTokenizer maybe ? (I have seen BartTokenizer in some random error log, I discarded it, but maybe ti's a True bug.
There was a problem hiding this comment.
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("microsoft/tapex-large-finetuned-wtq")
tokenizer.encode(table, truncation="drop_rows_to_fit")ValueError: drop_rows_to_fit is not a valid TruncationStrategy, please select one of ['only_first', 'only_second', 'longest_first', 'do_not_truncate']
There was a problem hiding this comment.
Thanks for reporting, investigating.
There was a problem hiding this comment.
Update: you can continue with this, TAPEX doesn't support drop_rows_to_fit in a similar manner as TAPAS does.
|
The documentation is not available anymore as the PR was closed or merged. |
|
Gentle ping @NielsRogge if we want to go live next week. |
…6663) * Enabling `Tapex` in table question answering pipeline. * Questions are independant for Tapex, making the test respect that. * Missing extra space.
4 participants
What does this PR do?
Tapexmodels in the pipelineself.typeflag, sinceTapas(the previously used models)have some non trivial defaults which do not apply to
Tapex.(Ideally thosetypeare not model specific like
tapasbutencodervsencoder-decoderwhenapplicable.
model yet: https://huggingface.co/hf-internal-testing
@NielsRogge
@LysandreJik
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.