Adding a PretrainedTransformerTokenizer (#3145)

* Adding a PretrainedTransformerTokenizer

* pylint

* doc

Author: matt-gardner

allennlp/data/tokenizers/__init__.py

@@ -5,5 +5,6 @@
 
 from allennlp.data.tokenizers.tokenizer import Token, Tokenizer
 from allennlp.data.tokenizers.word_tokenizer import WordTokenizer
+from allennlp.data.tokenizers.pretrained_transformer_tokenizer import PretrainedTransformerTokenizer
 from allennlp.data.tokenizers.character_tokenizer import CharacterTokenizer
 from allennlp.data.tokenizers.sentence_splitter import SentenceSplitter

allennlp/data/tokenizers/character_tokenizer.py

@@ -46,10 +46,6 @@ def __init__(self,
         self._start_tokens.reverse()
         self._end_tokens = end_tokens or []
 
-    @overrides
-    def batch_tokenize(self, texts: List[str]) -> List[List[Token]]:
-        return [self.tokenize(text) for text in texts]
-
     @overrides
     def tokenize(self, text: str) -> List[Token]:
         if self._lowercase_characters:

allennlp/data/tokenizers/pretrained_transformer_tokenizer.py

@@ -0,0 +1,64 @@
+import logging
+from typing import List, Tuple
+
+from overrides import overrides
+from pytorch_transformers.tokenization_auto import AutoTokenizer
+
+from allennlp.data.tokenizers.token import Token
+from allennlp.data.tokenizers.tokenizer import Tokenizer
+
+logger = logging.getLogger(__name__)
+
+
+@Tokenizer.register("pretrained_transformer")
+class PretrainedTransformerTokenizer(Tokenizer):
+    """
+    A ``PretrainedTransformerTokenizer`` uses a model from HuggingFace's
+    ``pytorch_transformers`` library to tokenize some input text.  This often means wordpieces
+    (where ``'AllenNLP is awesome'`` might get split into ``['Allen', '##NL', '##P', 'is',
+    'awesome']``), but it could also use byte-pair encoding, or some other tokenization, depending
+    on the pretrained model that you're using.
+
+    We take a model name as an input parameter, which we will pass to
+    ``AutoTokenizer.from_pretrained``.
+
+    Parameters
+    ----------
+    model_name : ``str``
+        The name of the pretrained wordpiece tokenizer to use.
+    start_tokens : ``List[str]``, optional
+        If given, these tokens will be added to the beginning of every string we tokenize.  We try
+        to be a little bit smart about defaults here - e.g., if your model name contains ``bert``,
+        we by default add ``[CLS]`` at the beginning and ``[SEP]`` at the end.
+    end_tokens : ``List[str]``, optional
+        If given, these tokens will be added to the end of every string we tokenize.
+    """
+    def __init__(self,
+                 model_name: str,
+                 do_lowercase: bool,
+                 start_tokens: List[str] = None,
+                 end_tokens: List[str] = None) -> None:
+        if model_name.endswith("-cased") and do_lowercase:
+            logger.warning("Your pretrained model appears to be cased, "
+                           "but your tokenizer is lowercasing tokens.")
+        elif model_name.endswith("-uncased") and not do_lowercase:
+            logger.warning("Your pretrained model appears to be uncased, "
+                           "but your tokenizer is not lowercasing tokens.")
+        self._tokenizer = AutoTokenizer.from_pretrained(model_name, do_lower_case=do_lowercase)
+        default_start_tokens, default_end_tokens = _guess_start_and_end_token_defaults(model_name)
+        self._start_tokens = start_tokens if start_tokens is not None else default_start_tokens
+        self._end_tokens = end_tokens if end_tokens is not None else default_end_tokens
+
+    @overrides
+    def tokenize(self, text: str) -> List[Token]:
+        # TODO(mattg): track character offsets.  Might be too challenging to do it here, given that
+        # pytorch-transformers is dealing with the whitespace...
+        token_strings = self._start_tokens + self._tokenizer.tokenize(text) + self._end_tokens
+        return [Token(t) for t in token_strings]
+
+
+def _guess_start_and_end_token_defaults(model_name: str) -> Tuple[List[str], List[str]]:
+    if 'bert' in model_name:
+        return (['[CLS]'], ['[SEP]'])
+    else:
+        return ([], [])

allennlp/data/tokenizers/tokenizer.py

@@ -27,8 +27,11 @@ def batch_tokenize(self, texts: List[str]) -> List[List[Token]]:
         """
         Batches together tokenization of several texts, in case that is faster for particular
         tokenizers.
+
+        By default we just do this without batching.  Override this in your tokenizer if you have a
+        good way of doing batched computation.
         """
-        raise NotImplementedError
+        return [self.tokenize(text) for text in texts]
 
     def tokenize(self, text: str) -> List[Token]:
         """

allennlp/tests/data/tokenizers/pretrained_transformer_tokenizer_test.py

@@ -0,0 +1,12 @@
+# pylint: disable=no-self-use,invalid-name
+
+from allennlp.common.testing import AllenNlpTestCase
+from allennlp.data.tokenizers import PretrainedTransformerTokenizer
+
+class TestPretrainedTransformerTokenizer(AllenNlpTestCase):
+    def test_splits_into_wordpieces(self):
+        tokenizer = PretrainedTransformerTokenizer('bert-base-cased', do_lowercase=False)
+        sentence = "A, [MASK] AllenNLP sentence."
+        tokens = [t.text for t in tokenizer.tokenize(sentence)]
+        expected_tokens = ["[CLS]", "A", ",", "[MASK]", "Allen", "##NL", "##P", "sentence", ".", "[SEP]"]
+        assert tokens == expected_tokens

doc/api/allennlp.data.tokenizers.rst

@@ -14,6 +14,7 @@ allennlp.data.tokenizers
 * :ref:`Tokenizer<tokenizer>`
 * :ref:`WordTokenizer<word-tokenizer>`
 * :ref:`CharacterTokenizer<character-tokenizer>`
+* :ref:`PretrainedTransformerTokenizer<pretrained-transformer-tokenizer>`
 * :ref:`WordFilter<word-filter>`
 * :ref:`WordSplitter<word-splitter>`
 * :ref:`WordStemmer<word-stemmer>`
@@ -36,6 +37,12 @@ allennlp.data.tokenizers
    :undoc-members:
    :show-inheritance:
 
+.. _pretrained-transformer-tokenizer:
+.. automodule:: allennlp.data.tokenizers.pretrained_transformer_tokenizer
+   :members:
+   :undoc-members:
+   :show-inheritance:
+
 .. _word-filter:
 .. automodule:: allennlp.data.tokenizers.word_filter
    :members:
@@ -58,4 +65,4 @@ allennlp.data.tokenizers
 .. automodule:: allennlp.data.tokenizers.sentence_splitter
    :members:
    :undoc-members:
-   :show-inheritance:
+   :show-inheritance:

requirements.txt

@@ -61,6 +61,7 @@ word2number>=1.1
 
 # To use the BERT model
 pytorch-pretrained-bert>=0.6.0
+git+git://github.com/huggingface/pytorch-transformers.git@a7b4cfe9194bf93c7044a42c9f1281260ce6279e
 
 # For caching processed data
 jsonpickle

setup.py

@@ -131,6 +131,7 @@
           'sqlparse>=0.2.4',
           'word2number>=1.1',
           'pytorch-pretrained-bert>=0.6.0',
+          'pytorch-transformers @ https://api.github.com/repos/huggingface/pytorch-transformers/tarball/a7b4cfe9194bf93c7044a42c9f1281260ce6279e',
           'jsonpickle',
       ],
       entry_points={