⚠️⚠️[T5Tokenize] Fix T5 family tokenizers⚠️⚠️

[ArthurZucker]

What does this PR do?

Fixes the T5Tokenizer (not the fast one yet). (at the same time adresses part of #11531)
When converting UMT5 I created a reproduction snippet for any t5x model form the original repo. I realized that a very very small variation in the input completely changes the output for non-finetuned models. The issue lies with the way we process <extra_id_xx>.

Example:

# t5-base tokenizer
>>> tokenizer.encode("<extra_id_0>. Hello", add_special_tokens = False)
[32099, 3, 5, 8774] # ['<extra_id_0>', ' ▁', '.', '▁Hello']
# seqio.SentencePieceVocabulary(vocab_path, extra_ids = 300)
>>> processor.encode("<extra_id_0>. Hello")
[32099, 5, 8774] # ['<extra_id_0>', '.', '▁Hello']

#after fix: 
>>> tokenizer.encode("<extra_id_0>. Hello", add_special_tokens = False)
[32099, 5, 8774] # ['<extra_id_0>', '.', '▁Hello']

The reason is that t5x wrapps arround sentencepiece, and adds the extra id to the vocab, but they are not saved that way.
We don't add them to the vocab, so when we tokenize, we split on special tokens, thus the sentencepiece model only sees:

>>> tokenizer.sp_model.encode(". Hello")
[273, 274, 9] 

While the original model never sees a . (or a lot of other characters) alone, and thus we add an extra space...

This is a bug fix with regards to training, it is breaking in the sense that is should remove the space.

TODO:

• Extra checks should be added to make sure this does not add anything else (like stripping a . This for example would break: tokenizer.encode(". Hello") as it remove the prefix space that is normally added.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ArthurZucker]

Actually switch t5 tests have to be updated!
This means I have to check if the models were trained with this extra token (if they used HF tokenizer) or not.

• tests.models.instructblip.test_modeling_instructblip.InstructBlipModelIntegrationTest testMethod=test_inference_flant5_xl failing on main too so not related.....

• tests.models.mt5.test_modeling_flax_mt5.MT5IntegrationTest also fails on main...

• tests/models/t5/test_tokenization_t5.py the issue comes from the convert_slow modification. Need to investigate
  - [ ] tests/models/t5/test_tokenization_t5.py:399 T5TokenizationTest.test_get_sentinel_token_ids_for_fasttokenizer
  - [ ] tests/test_tokenization_common.py:3425 T5TokenizationTest.test_save_pretrained
  - [ ] tests/models/t5/test_tokenization_t5.py:271 T5TokenizationTest.test_special_tokens_initialization

[ArthurZucker]

This can also be made non "breakable" with a flag. Up to debate since it is a bug fix.

[sgugger]

Thanks for the fix! Let's roll with it since it's a bug fix and if people complain about the breaking change we will see if we add a flag to enable the buggy behavior.

[ArthurZucker]

Edit: just to make sure, I did more testing and unfortunately , there is one bug:

>>>tokenizer.tokenize("Hello <extra_id_0>")
['_', '_Hello', '<extra_id_0>']

instead of

>>>tokenizer.tokenize("Hello <extra_id_0>")
['_Hello', '<extra_id_0>']

This is because we have to prepend a _ instead of a space. (text = SPIECE_UNDERLINE + text. Not a single test caught this when runing pytest tests -k t5 which is interesting.
Fixing asap and adding tests. This is becoming very complex 😓

[pointonjoel]

I'm getting this legacy behaviour warning come up when simply loading a T5 tokenizer - it appears even before using the tokenizer. Is there an updated way to load the tokenizer? The warning appears when running the following lines of code:

from transformers import AutoTokenizer
tokeniser = AutoTokenizer.from_pretrained("google/mt5-small")

The error is:
You are using the legacy behaviour of the <class 'transformers.models.t5.tokenization_t5.T5Tokenizer'>. This means that tokens that come after special tokens will not be properly handled. We recommend you to read the related pull request available at #24565
/usr/local/lib/python3.10/dist-packages/transformers/convert_slow_tokenizer.py:470: UserWarning: The sentencepiece tokenizer that you are converting to a fast tokenizer uses the byte fallback option which is not implemented in the fast tokenizers. In practice this means that the fast version of the tokenizer can produce unknown tokens whereas the sentencepiece version would have converted these unknown tokens into a sequence of byte tokens matching the original piece of text.
warnings.warn(

[ArthurZucker]

Yep, just set legacy=False. The goal of the warning is for you to decide wether or not you thing the legacy behaviour is alright with you or not.