Add FastTokenizer to REALM (#15211)

* Remove BertTokenizer abstraction

* Add FastTokenizer to REALM

* Fix config archive map

* Fix copies

* Update realm.mdx

* Apply suggestions from code review

docs/source/index.mdx

@@ -246,7 +246,7 @@ Flax), PyTorch, and/or TensorFlow.
 |         ProphetNet          |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           QDQBert           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             RAG             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
-|            Realm            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            Realm            |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |          Reformer           |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
 |           RemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |          RetriBERT          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |

docs/source/model_doc/realm.mdx

@@ -49,6 +49,11 @@ This model was contributed by [qqaatw](https://huggingface.co/qqaatw). The origi
     - save_vocabulary
     - batch_encode_candidates
 
+## RealmTokenizerFast
+
+[[autodoc]] RealmTokenizerFast
+    - batch_encode_candidates
+
 ## RealmRetriever
 
 [[autodoc]] RealmRetriever

src/transformers/__init__.py

@@ -419,6 +419,7 @@ else:
 # tokenizers-backed objects
 if is_tokenizers_available():
     # Fast tokenizers
+    _import_structure["models.realm"].append("RealmTokenizerFast")
     _import_structure["models.fnet"].append("FNetTokenizerFast")
     _import_structure["models.roformer"].append("RoFormerTokenizerFast")
     _import_structure["models.clip"].append("CLIPTokenizerFast")
@@ -2542,6 +2543,7 @@ if TYPE_CHECKING:
         from .models.mt5 import MT5TokenizerFast
         from .models.openai import OpenAIGPTTokenizerFast
         from .models.pegasus import PegasusTokenizerFast
+        from .models.realm import RealmTokenizerFast
         from .models.reformer import ReformerTokenizerFast
         from .models.rembert import RemBertTokenizerFast
         from .models.retribert import RetriBertTokenizerFast

src/transformers/convert_slow_tokenizer.py

@@ -942,6 +942,7 @@ SLOW_TO_FAST_CONVERTERS = {
     "MobileBertTokenizer": BertConverter,
     "OpenAIGPTTokenizer": OpenAIGPTConverter,
     "PegasusTokenizer": PegasusConverter,
+    "RealmTokenizer": BertConverter,
     "ReformerTokenizer": ReformerConverter,
     "RemBertTokenizer": RemBertConverter,
     "RetriBertTokenizer": BertConverter,

src/transformers/models/realm/__init__.py

@@ -25,6 +25,8 @@ _import_structure = {
     "tokenization_realm": ["RealmTokenizer"],
 }
 
+if is_tokenizers_available():
+    _import_structure["tokenization_realm_fast"] = ["RealmTokenizerFast"]
 
 if is_torch_available():
     _import_structure["modeling_realm"] = [
@@ -44,6 +46,9 @@ if TYPE_CHECKING:
     from .configuration_realm import REALM_PRETRAINED_CONFIG_ARCHIVE_MAP, RealmConfig
     from .tokenization_realm import RealmTokenizer
 
+    if is_tokenizers_available():
+        from .tokenization_realm import RealmTokenizerFast
+
     if is_torch_available():
         from .modeling_realm import (
             REALM_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/models/realm/configuration_realm.py

@@ -21,14 +21,14 @@ from ...utils import logging
 logger = logging.get_logger(__name__)
 
 REALM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
-    "realm-cc-news-pretrained-embedder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-embedder/resolve/main/config.json",
-    "realm-cc-news-pretrained-encoder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-encoder/resolve/main/config.json",
-    "realm-cc-news-pretrained-scorer": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-scorer/resolve/main/config.json",
-    "realm-cc-news-pretrained-openqa": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-openqa/aresolve/main/config.json",
-    "realm-orqa-nq-openqa": "https://huggingface.co/qqaatw/realm-orqa-nq-openqa/resolve/main/config.json",
-    "realm-orqa-nq-reader": "https://huggingface.co/qqaatw/realm-orqa-nq-reader/resolve/main/config.json",
-    "realm-orqa-wq-openqa": "https://huggingface.co/qqaatw/realm-orqa-wq-openqa/resolve/main/config.json",
-    "realm-orqa-wq-reader": "https://huggingface.co/qqaatw/realm-orqa-wq-reader/resolve/main/config.json",
+    "qqaatw/realm-cc-news-pretrained-embedder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-embedder/resolve/main/config.json",
+    "qqaatw/realm-cc-news-pretrained-encoder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-encoder/resolve/main/config.json",
+    "qqaatw/realm-cc-news-pretrained-scorer": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-scorer/resolve/main/config.json",
+    "qqaatw/realm-cc-news-pretrained-openqa": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-openqa/aresolve/main/config.json",
+    "qqaatw/realm-orqa-nq-openqa": "https://huggingface.co/qqaatw/realm-orqa-nq-openqa/resolve/main/config.json",
+    "qqaatw/realm-orqa-nq-reader": "https://huggingface.co/qqaatw/realm-orqa-nq-reader/resolve/main/config.json",
+    "qqaatw/realm-orqa-wq-openqa": "https://huggingface.co/qqaatw/realm-orqa-wq-openqa/resolve/main/config.json",
+    "qqaatw/realm-orqa-wq-reader": "https://huggingface.co/qqaatw/realm-orqa-wq-reader/resolve/main/config.json",
     # See all REALM models at https://huggingface.co/models?filter=realm
 }
 

src/transformers/models/realm/tokenization_realm_fast.py

+# coding=utf-8
+# Copyright 2022 The REALM authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Fast Tokenization classes for REALM."""
+
+import json
+from typing import List, Optional, Tuple
+
+from tokenizers import normalizers
+
+from ...file_utils import PaddingStrategy
+from ...tokenization_utils_base import BatchEncoding
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+from .tokenization_realm import RealmTokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt", "tokenizer_file": "tokenizer.json"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "qqaatw/realm-cc-news-pretrained-embedder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-embedder/resolve/main/vocab.txt",
+        "qqaatw/realm-cc-news-pretrained-encoder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-encoder/resolve/main/vocab.txt",
+        "qqaatw/realm-cc-news-pretrained-scorer": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-scorer/resolve/main/vocab.txt",
+        "qqaatw/realm-cc-news-pretrained-openqa": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-openqa/aresolve/main/vocab.txt",
+        "qqaatw/realm-orqa-nq-openqa": "https://huggingface.co/qqaatw/realm-orqa-nq-openqa/resolve/main/vocab.txt",
+        "qqaatw/realm-orqa-nq-reader": "https://huggingface.co/qqaatw/realm-orqa-nq-reader/resolve/main/vocab.txt",
+        "qqaatw/realm-orqa-wq-openqa": "https://huggingface.co/qqaatw/realm-orqa-wq-openqa/resolve/main/vocab.txt",
+        "qqaatw/realm-orqa-wq-reader": "https://huggingface.co/qqaatw/realm-orqa-wq-reader/resolve/main/vocab.txt",
+    },
+    "tokenizer_file": {
+        "qqaatw/realm-cc-news-pretrained-embedder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-embedder/resolve/main/tokenizer.jsont",
+        "qqaatw/realm-cc-news-pretrained-encoder": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-encoder/resolve/main/tokenizer.json",
+        "qqaatw/realm-cc-news-pretrained-scorer": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-scorer/resolve/main/tokenizer.json",
+        "qqaatw/realm-cc-news-pretrained-openqa": "https://huggingface.co/qqaatw/realm-cc-news-pretrained-openqa/aresolve/main/tokenizer.json",
+        "qqaatw/realm-orqa-nq-openqa": "https://huggingface.co/qqaatw/realm-orqa-nq-openqa/resolve/main/tokenizer.json",
+        "qqaatw/realm-orqa-nq-reader": "https://huggingface.co/qqaatw/realm-orqa-nq-reader/resolve/main/tokenizer.json",
+        "qqaatw/realm-orqa-wq-openqa": "https://huggingface.co/qqaatw/realm-orqa-wq-openqa/resolve/main/tokenizer.json",
+        "qqaatw/realm-orqa-wq-reader": "https://huggingface.co/qqaatw/realm-orqa-wq-reader/resolve/main/tokenizer.json",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "qqaatw/realm-cc-news-pretrained-embedder": 512,
+    "qqaatw/realm-cc-news-pretrained-encoder": 512,
+    "qqaatw/realm-cc-news-pretrained-scorer": 512,
+    "qqaatw/realm-cc-news-pretrained-openqa": 512,
+    "qqaatw/realm-orqa-nq-openqa": 512,
+    "qqaatw/realm-orqa-nq-reader": 512,
+    "qqaatw/realm-orqa-wq-openqa": 512,
+    "qqaatw/realm-orqa-wq-reader": 512,
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "qqaatw/realm-cc-news-pretrained-embedder": {"do_lower_case": True},
+    "qqaatw/realm-cc-news-pretrained-encoder": {"do_lower_case": True},
+    "qqaatw/realm-cc-news-pretrained-scorer": {"do_lower_case": True},
+    "qqaatw/realm-cc-news-pretrained-openqa": {"do_lower_case": True},
+    "qqaatw/realm-orqa-nq-openqa": {"do_lower_case": True},
+    "qqaatw/realm-orqa-nq-reader": {"do_lower_case": True},
+    "qqaatw/realm-orqa-wq-openqa": {"do_lower_case": True},
+    "qqaatw/realm-orqa-wq-reader": {"do_lower_case": True},
+}
+
+
+class RealmTokenizerFast(PreTrainedTokenizerFast):
+    r"""
+    Construct a "fast" REALM tokenizer (backed by HuggingFace's *tokenizers* library). Based on WordPiece.
+
+    [`RealmTokenizerFast`] is identical to [`BertTokenizerFast`] and runs end-to-end tokenization: punctuation
+    splitting and wordpiece.
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            File containing the vocabulary.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+        unk_token (`str`, *optional*, defaults to `"[UNK]"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        sep_token (`str`, *optional*, defaults to `"[SEP]"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        pad_token (`str`, *optional*, defaults to `"[PAD]"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (`str`, *optional*, defaults to `"[CLS]"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        mask_token (`str`, *optional*, defaults to `"[MASK]"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        clean_text (`bool`, *optional*, defaults to `True`):
+            Whether or not to clean the text before tokenization by removing any control characters and replacing all
+            whitespaces by the classic one.
+        tokenize_chinese_chars (`bool`, *optional*, defaults to `True`):
+            Whether or not to tokenize Chinese characters. This should likely be deactivated for Japanese (see [this
+            issue](https://github.com/huggingface/transformers/issues/328)).
+        strip_accents (`bool`, *optional*):
+            Whether or not to strip all accents. If this option is not specified, then it will be determined by the
+            value for `lowercase` (as in the original BERT).
+        wordpieces_prefix (`str`, *optional*, defaults to `"##"`):
+            The prefix for subwords.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    slow_tokenizer_class = RealmTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        tokenizer_file=None,
+        do_lower_case=True,
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        tokenize_chinese_chars=True,
+        strip_accents=None,
+        **kwargs
+    ):
+        super().__init__(
+            vocab_file,
+            tokenizer_file=tokenizer_file,
+            do_lower_case=do_lower_case,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            tokenize_chinese_chars=tokenize_chinese_chars,
+            strip_accents=strip_accents,
+            **kwargs,
+        )
+
+        normalizer_state = json.loads(self.backend_tokenizer.normalizer.__getstate__())
+        if (
+            normalizer_state.get("lowercase", do_lower_case) != do_lower_case
+            or normalizer_state.get("strip_accents", strip_accents) != strip_accents
+            or normalizer_state.get("handle_chinese_chars", tokenize_chinese_chars) != tokenize_chinese_chars
+        ):
+            normalizer_class = getattr(normalizers, normalizer_state.pop("type"))
+            normalizer_state["lowercase"] = do_lower_case
+            normalizer_state["strip_accents"] = strip_accents
+            normalizer_state["handle_chinese_chars"] = tokenize_chinese_chars
+            self.backend_tokenizer.normalizer = normalizer_class(**normalizer_state)
+
+        self.do_lower_case = do_lower_case
+
+    def batch_encode_candidates(self, text, **kwargs):
+        r"""
+        Encode a batch of text or text pair. This method is similar to regular __call__ method but has the following
+        differences:
+
+            1. Handle additional num_candidate axis. (batch_size, num_candidates, text)
+            2. Always pad the sequences to *max_length*.
+            3. Must specify *max_length* in order to stack packs of candidates into a batch.
+
+            - single sequence: `[CLS] X [SEP]`
+            - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            text (`List[List[str]]`):
+                The batch of sequences to be encoded. Each sequence must be in this format: (batch_size,
+                num_candidates, text).
+            text_pair (`List[List[str]]`, *optional*):
+                The batch of sequences to be encoded. Each sequence must be in this format: (batch_size,
+                num_candidates, text).
+            **kwargs:
+                Keyword arguments of the __call__ method.
+
+        Returns:
+            [`BatchEncoding`]: Encoded text or text pair.
+
+        Example:
+
+        ```python
+        >>> from transformers import RealmTokenizerFast
+
+        >>> # batch_size = 2, num_candidates = 2
+        >>> text = [["Hello world!", "Nice to meet you!"], ["The cute cat.", "The adorable dog."]]
+
+        >>> tokenizer = RealmTokenizerFast.from_pretrained("qqaatw/realm-cc-news-pretrained-encoder")
+        >>> tokenized_text = tokenizer.batch_encode_candidates(text, max_length=10, return_tensors="pt")
+        ```"""
+
+        # Always using a fixed sequence length to encode in order to stack candidates into a batch.
+        kwargs["padding"] = PaddingStrategy.MAX_LENGTH
+
+        batch_text = text
+        batch_text_pair = kwargs.pop("text_pair", None)
+        return_tensors = kwargs.pop("return_tensors", None)
+
+        output_data = {
+            "input_ids": [],
+            "attention_mask": [],
+            "token_type_ids": [],
+        }
+
+        for idx, candidate_text in enumerate(batch_text):
+            if batch_text_pair is not None:
+                candidate_text_pair = batch_text_pair[idx]
+            else:
+                candidate_text_pair = None
+
+            encoded_candidates = super().__call__(candidate_text, candidate_text_pair, return_tensors=None, **kwargs)
+
+            encoded_input_ids = encoded_candidates.get("input_ids")
+            encoded_attention_mask = encoded_candidates.get("attention_mask")
+            encoded_token_type_ids = encoded_candidates.get("token_type_ids")
+
+            if encoded_input_ids is not None:
+                output_data["input_ids"].append(encoded_input_ids)
+            if encoded_attention_mask is not None:
+                output_data["attention_mask"].append(encoded_attention_mask)
+            if encoded_token_type_ids is not None:
+                output_data["token_type_ids"].append(encoded_token_type_ids)
+
+        output_data = dict((key, item) for key, item in output_data.items() if len(item) != 0)
+
+        return BatchEncoding(output_data, tensor_type=return_tensors)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A REALM sequence has the following format:
+
+        - single sequence: `[CLS] X [SEP]`
+        - pair of sequences: `[CLS] A [SEP] B [SEP]`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        output = [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+
+        if token_ids_1:
+            output += token_ids_1 + [self.sep_token_id]
+
+        return output
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. A REALM sequence
+        pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        If `token_ids_1` is `None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
+        return tuple(files)

src/transformers/utils/dummy_tokenizers_objects.py

@@ -234,6 +234,13 @@ class PegasusTokenizerFast(metaclass=DummyObject):
         requires_backends(self, ["tokenizers"])
 
 
+class RealmTokenizerFast(metaclass=DummyObject):
+    _backends = ["tokenizers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tokenizers"])
+
+
 class ReformerTokenizerFast(metaclass=DummyObject):
     _backends = ["tokenizers"]
 

tests/test_tokenization_realm.py

@@ -16,6 +16,7 @@
 import os
 import unittest
 
+from transformers import RealmTokenizerFast
 from transformers.models.bert.tokenization_bert import (
     VOCAB_FILES_NAMES,
     BasicTokenizer,
@@ -34,8 +35,8 @@ from .test_tokenization_common import TokenizerTesterMixin, filter_non_english
 class RealmTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
 
     tokenizer_class = RealmTokenizer
-    rust_tokenizer_class = None
-    test_rust_tokenizer = False
+    rust_tokenizer_class = RealmTokenizerFast
+    test_rust_tokenizer = True
     space_between_special_tokens = True
     from_pretrained_filter = filter_non_english
 
@@ -301,14 +302,21 @@ class RealmTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
 
     @slow
     def test_batch_encode_candidates(self):
-        tokenizer = self.tokenizer_class.from_pretrained("bert-base-uncased")
-
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+                tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs)
                 text = [["Hello world!", "Nice to meet you!"], ["The cute cat.", "The adorable dog."]]
 
-        encoded_sentence = tokenizer.batch_encode_candidates(text, max_length=10, return_tensors="pt")
+                encoded_sentence_r = tokenizer_r.batch_encode_candidates(text, max_length=10, return_tensors="np")
+                encoded_sentence_p = tokenizer_p.batch_encode_candidates(text, max_length=10, return_tensors="np")
 
                 expected_shape = (2, 2, 10)
 
-        assert encoded_sentence["input_ids"].shape == expected_shape
-        assert encoded_sentence["attention_mask"].shape == expected_shape
-        assert encoded_sentence["token_type_ids"].shape == expected_shape
+                self.assertEqual(encoded_sentence_r["input_ids"].shape, expected_shape)
+                self.assertEqual(encoded_sentence_r["attention_mask"].shape, expected_shape)
+                self.assertEqual(encoded_sentence_r["token_type_ids"].shape, expected_shape)
+
+                self.assertEqual(encoded_sentence_p["input_ids"].shape, expected_shape)
+                self.assertEqual(encoded_sentence_p["attention_mask"].shape, expected_shape)
+                self.assertEqual(encoded_sentence_p["token_type_ids"].shape, expected_shape)