Enable option for subword regularization in more tokenizers. (#11417)

* improve slow class tok usage at xlm rob

* add subword regularization for barthez

* improve barthez tok. test

* fix tokenizer tests

* add subword regularization for camembert

* add subword regularization for deberta v2 tokenizer

* add more doc to deberta v2 tokenizer

* add subword regularization for speech to text tok.

* fix sp_model_kwargs type in speech 2 text tok.

* add subword regularization for M2M100 tok.

* add more concrete type hints

* fix tests for m2m100 and s2t tok.

* add missing Any import

* fix syntax error in m2m100 tok.

* fix unpickle of m2m100 and s2t tok.

* fix test of m2m100 and s2t tok.

* improve unpickle of deberta v2 tok.

* add test for pickle of barthez & camembert

* fix pickle of barthez & camembert

* add test for deberta v2 tok. pickle

* fix m2m100 tok. pickle

* fix s2t tok. pickle

* add subword regularization to albert tok.

* refactor subword reg. test into TokenizerTesterMixin

improve albert tok. test

remove sample argument form albert tok.

check subword reg. using TokenizerTesterMixin

improve tok. tests

improve xlm roberta tok. tests

improve xlm roberta tok. tests

* add subword regularization for big bird t.

* improve xlm roberta tok. test

* add subword regularization for mbart50 tok.

* add subword regularization for pegasus tok.

* add subword regularization for reformer tok.

* add subword regularization for T5 tok.

* fix t5 tok. test formatting

* add subword regularization for xlm_proph. tok.

* add subword regularization for xlnet tok.

* add subword regularization for gert_gen tok.

* add typing to tokenizers

* add typing to xlm rob. tok

* add subword regularization for marian tok.

* add reverse tok. test

* fix marian tok test

* fix marian tok test

* fix casing in tok. tests

* fix style of tok. common test

* fix deberta v2 tok test

* add type annotations to tok. tests

* add type annotations to tok. __init__

* add typing to kokenizer

* add type annotations to tok. __init__

* don't specify the default when it's None

* fix barthez tok. doc

* move sentencepiece tok. tests to TokenizerTesterMixin

* fix unused imports

* fix albert tok. test

* add comment to sentencepiece test options

* fix Any import at big bird tok.

* fix Any import at xlm prophetnet tok.

* empty commit to trigger CI

src/transformers/models/albert/tokenization_albert.py

@@ -18,7 +18,7 @@
 import os
 import unicodedata
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -102,6 +102,20 @@ class AlbertTokenizer(PreTrainedTokenizer):
         mask_token (:obj:`str`, `optional`, defaults to :obj:`"[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.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Attributes:
         sp_model (:obj:`SentencePieceProcessor`):
@@ -125,11 +139,14 @@ class AlbertTokenizer(PreTrainedTokenizer):
         pad_token="<pad>",
         cls_token="[CLS]",
         mask_token="[MASK]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             do_lower_case=do_lower_case,
             remove_space=remove_space,
@@ -141,6 +158,7 @@ class AlbertTokenizer(PreTrainedTokenizer):
             pad_token=pad_token,
             cls_token=cls_token,
             mask_token=mask_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
@@ -149,7 +167,7 @@ class AlbertTokenizer(PreTrainedTokenizer):
         self.keep_accents = keep_accents
         self.vocab_file = vocab_file
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
     @property
@@ -168,7 +186,12 @@ class AlbertTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
     def preprocess_text(self, inputs):
@@ -186,14 +209,10 @@ class AlbertTokenizer(PreTrainedTokenizer):
 
         return outputs
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Tokenize a string."""
         text = self.preprocess_text(text)
-
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
+        pieces = self.sp_model.encode(text, out_type=str)
         new_pieces = []
         for piece in pieces:
             if len(piece) > 1 and piece[-1] == str(",") and piece[-2].isdigit():

src/transformers/models/barthez/tokenization_barthez.py

@@ -17,7 +17,7 @@
 
 import os
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -89,6 +89,20 @@ class BarthezTokenizer(PreTrainedTokenizer):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Attributes:
         sp_model (:obj:`SentencePieceProcessor`):
@@ -110,11 +124,14 @@ class BarthezTokenizer(PreTrainedTokenizer):
         unk_token="<unk>",
         pad_token="<pad>",
         mask_token="<mask>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             bos_token=bos_token,
             eos_token=eos_token,
@@ -123,11 +140,12 @@ class BarthezTokenizer(PreTrainedTokenizer):
             cls_token=cls_token,
             pad_token=pad_token,
             mask_token=mask_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
         self.vocab_file = vocab_file
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(str(vocab_file))
 
         self.fairseq_tokens_to_ids = {"<s>": 0, "<pad>": 1, "</s>": 2, "<unk>": 3}
@@ -219,8 +237,8 @@ class BarthezTokenizer(PreTrainedTokenizer):
         vocab.update(self.added_tokens_encoder)
         return vocab
 
-    def _tokenize(self, text):
-        return self.sp_model.EncodeAsPieces(text)
+    def _tokenize(self, text: str) -> List[str]:
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""
@@ -243,7 +261,12 @@ class BarthezTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
     def convert_tokens_to_string(self, tokens):

src/transformers/models/bert_generation/tokenization_bert_generation.py

@@ -17,7 +17,7 @@
 
 import os
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -58,6 +58,20 @@ class BertGenerationTokenizer(PreTrainedTokenizer):
             token instead.
         pad_token (:obj:`str`, `optional`, defaults to :obj:`"<pad>"`):
             The token used for padding, for example when batching sequences of different lengths.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -74,8 +88,11 @@ class BertGenerationTokenizer(PreTrainedTokenizer):
         unk_token="<unk>",
         pad_token="<pad>",
         sep_token="<::::>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         # Add extra_ids to the special token list
         super().__init__(
             bos_token=bos_token,
@@ -83,12 +100,13 @@ class BertGenerationTokenizer(PreTrainedTokenizer):
             unk_token=unk_token,
             pad_token=pad_token,
             sep_token=sep_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
         self.vocab_file = vocab_file
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
     @property
@@ -107,16 +125,17 @@ class BertGenerationTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Take as input a string and return a list of strings (tokens) for words/sub-words"""
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
-        return pieces
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""

src/transformers/models/big_bird/tokenization_big_bird.py

@@ -17,7 +17,7 @@
 
 import os
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -74,7 +74,20 @@ class BigBirdTokenizer(PreTrainedTokenizer):
         mask_token (:obj:`str`, `optional`, defaults to :obj:`"[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.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
 
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -93,8 +106,9 @@ class BigBirdTokenizer(PreTrainedTokenizer):
         sep_token="[SEP]",
         mask_token="[MASK]",
         cls_token="[CLS]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
         eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
         unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
@@ -105,6 +119,8 @@ class BigBirdTokenizer(PreTrainedTokenizer):
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             bos_token=bos_token,
             eos_token=eos_token,
@@ -113,12 +129,13 @@ class BigBirdTokenizer(PreTrainedTokenizer):
             sep_token=sep_token,
             mask_token=mask_token,
             cls_token=cls_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
         self.vocab_file = vocab_file
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
     @property
@@ -137,16 +154,17 @@ class BigBirdTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Take as input a string and return a list of strings (tokens) for words/sub-words"""
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
-        return pieces
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""

src/transformers/models/camembert/tokenization_camembert.py

@@ -17,7 +17,7 @@
 
 import os
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -85,6 +85,20 @@ class CamembertTokenizer(PreTrainedTokenizer):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Attributes:
         sp_model (:obj:`SentencePieceProcessor`):
@@ -107,11 +121,14 @@ class CamembertTokenizer(PreTrainedTokenizer):
         pad_token="<pad>",
         mask_token="<mask>",
         additional_special_tokens=["<s>NOTUSED", "</s>NOTUSED"],
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             bos_token=bos_token,
             eos_token=eos_token,
@@ -121,9 +138,10 @@ class CamembertTokenizer(PreTrainedTokenizer):
             pad_token=pad_token,
             mask_token=mask_token,
             additional_special_tokens=additional_special_tokens,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(str(vocab_file))
         self.vocab_file = vocab_file
         # HACK: These tokens were added by fairseq but don't seem to be actually used when duplicated in the actual
@@ -218,8 +236,8 @@ class CamembertTokenizer(PreTrainedTokenizer):
         vocab.update(self.added_tokens_encoder)
         return vocab
 
-    def _tokenize(self, text):
-        return self.sp_model.EncodeAsPieces(text)
+    def _tokenize(self, text: str) -> List[str]:
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""
@@ -243,7 +261,12 @@ class CamembertTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
     def convert_tokens_to_string(self, tokens):

src/transformers/models/deberta_v2/tokenization_deberta_v2.py

@@ -16,7 +16,7 @@
 
 import os
 import unicodedata
-from typing import Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as sp
 import six
@@ -75,6 +75,20 @@ class DebertaV2Tokenizer(PreTrainedTokenizer):
         mask_token (:obj:`str`, `optional`, defaults to :obj:`"[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.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -92,8 +106,11 @@ class DebertaV2Tokenizer(PreTrainedTokenizer):
         pad_token="[PAD]",
         cls_token="[CLS]",
         mask_token="[MASK]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             do_lower_case=do_lower_case,
             unk_token=unk_token,
@@ -102,6 +119,7 @@ class DebertaV2Tokenizer(PreTrainedTokenizer):
             cls_token=cls_token,
             mask_token=mask_token,
             split_by_punct=split_by_punct,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
@@ -112,7 +130,7 @@ class DebertaV2Tokenizer(PreTrainedTokenizer):
             )
         self.do_lower_case = do_lower_case
         self.split_by_punct = split_by_punct
-        self._tokenizer = SPMTokenizer(vocab_file, split_by_punct=split_by_punct)
+        self._tokenizer = SPMTokenizer(vocab_file, split_by_punct=split_by_punct, sp_model_kwargs=self.sp_model_kwargs)
 
     @property
     def vocab_size(self):
@@ -127,7 +145,7 @@ class DebertaV2Tokenizer(PreTrainedTokenizer):
         vocab.update(self.get_added_vocab())
         return vocab
 
-    def _tokenize(self, text):
+    def _tokenize(self, text: str) -> List[str]:
         """Take as input a string and return a list of strings (tokens) for words/sub-words"""
         if self.do_lower_case:
             text = text.lower()
@@ -234,10 +252,34 @@ class DebertaV2Tokenizer(PreTrainedTokenizer):
 
 
 class SPMTokenizer:
-    def __init__(self, vocab_file, split_by_punct=False):
+    r"""
+    Constructs a tokenizer based on `SentencePiece <https://github.com/google/sentencepiece>`__.
+
+    Args:
+        vocab_file (:obj:`str`):
+            `SentencePiece <https://github.com/google/sentencepiece>`__ file (generally has a `.spm` extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+    """
+
+    def __init__(self, vocab_file, split_by_punct=False, sp_model_kwargs: Optional[Dict[str, Any]] = None):
         self.split_by_punct = split_by_punct
         self.vocab_file = vocab_file
-        spm = sp.SentencePieceProcessor()
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        spm = sp.SentencePieceProcessor(**self.sp_model_kwargs)
         assert os.path.exists(vocab_file)
         spm.load(vocab_file)
         bpe_vocab_size = spm.GetPieceSize()
@@ -261,7 +303,12 @@ class SPMTokenizer:
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.spm = sp.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.spm = sp.SentencePieceProcessor(**self.sp_model_kwargs)
         self.spm.Load(self.vocab_file)
 
     def tokenize(self, text):
@@ -344,10 +391,10 @@ class SPMTokenizer:
         text = convert_to_unicode(text)
         if self.split_by_punct:
             words = self._run_split_on_punc(text)
-            pieces = [self.spm.encode_as_pieces(w) for w in words]
+            pieces = [self.spm.encode(w, out_type=str) for w in words]
             return [p for w in pieces for p in w]
         else:
-            return self.spm.encode_as_pieces(text)
+            return self.spm.encode(text, out_type=str)
 
     def split_to_words(self, text):
         pieces = self._encode_as_pieces(text)

src/transformers/models/m2m_100/tokenization_m2m_100.py

@@ -16,7 +16,7 @@ import json
 from contextlib import contextmanager
 from pathlib import Path
 from shutil import copyfile
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import sentencepiece
 
@@ -86,6 +86,20 @@ class M2M100Tokenizer(PreTrainedTokenizer):
             token instead.
         pad_token (:obj:`str`, `optional`, defaults to :obj:`"<pad>"`):
             The token used for padding, for example when batching sequences of different lengths.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Examples::
 
@@ -118,8 +132,11 @@ class M2M100Tokenizer(PreTrainedTokenizer):
         sep_token="</s>",
         pad_token="<pad>",
         unk_token="<unk>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs,
-    ):
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             src_lang=src_lang,
             tgt_lang=tgt_lang,
@@ -128,6 +145,7 @@ class M2M100Tokenizer(PreTrainedTokenizer):
             sep_token=sep_token,
             unk_token=unk_token,
             pad_token=pad_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
@@ -135,7 +153,7 @@ class M2M100Tokenizer(PreTrainedTokenizer):
         self.encoder = load_json(vocab_file)
         self.decoder = {v: k for k, v in self.encoder.items()}
         self.spm_file = spm_file
-        self.sp_model = load_spm(spm_file)
+        self.sp_model = load_spm(spm_file, self.sp_model_kwargs)
 
         self.encoder_size = len(self.encoder)
 
@@ -169,7 +187,7 @@ class M2M100Tokenizer(PreTrainedTokenizer):
         self.set_src_lang_special_tokens(self._src_lang)
 
     def _tokenize(self, text: str) -> List[str]:
-        return self.sp_model.EncodeAsPieces(text)
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         if token in self.lang_token_to_id:
@@ -256,7 +274,12 @@ class M2M100Tokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d: Dict) -> None:
         self.__dict__ = d
-        self.sp_model = load_spm(self.spm_file)
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = load_spm(self.spm_file, self.sp_model_kwargs)
 
     def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
         save_dir = Path(save_directory)
@@ -330,8 +353,8 @@ class M2M100Tokenizer(PreTrainedTokenizer):
         return self.lang_token_to_id[lang_token]
 
 
-def load_spm(path: str) -> sentencepiece.SentencePieceProcessor:
-    spm = sentencepiece.SentencePieceProcessor()
+def load_spm(path: str, sp_model_kwargs: Dict[str, Any]) -> sentencepiece.SentencePieceProcessor:
+    spm = sentencepiece.SentencePieceProcessor(**sp_model_kwargs)
     spm.Load(str(path))
     return spm
 

src/transformers/models/marian/tokenization_marian.py

@@ -18,7 +18,7 @@ import warnings
 from contextlib import contextmanager
 from pathlib import Path
 from shutil import copyfile
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import sentencepiece
 
@@ -82,6 +82,20 @@ class MarianTokenizer(PreTrainedTokenizer):
             The maximum sentence length the model accepts.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<eop>", "<eod>"]`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Examples::
 
@@ -115,8 +129,11 @@ class MarianTokenizer(PreTrainedTokenizer):
         eos_token="</s>",
         pad_token="<pad>",
         model_max_length=512,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             # bos_token=bos_token,  unused. Start decoding with config.decoder_start_token_id
             source_lang=source_lang,
@@ -125,6 +142,7 @@ class MarianTokenizer(PreTrainedTokenizer):
             eos_token=eos_token,
             pad_token=pad_token,
             model_max_length=model_max_length,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
         assert Path(source_spm).exists(), f"cannot find spm source {source_spm}"
@@ -140,8 +158,8 @@ class MarianTokenizer(PreTrainedTokenizer):
         self.spm_files = [source_spm, target_spm]
 
         # load SentencePiece model for pre-processing
-        self.spm_source = load_spm(source_spm)
-        self.spm_target = load_spm(target_spm)
+        self.spm_source = load_spm(source_spm, self.sp_model_kwargs)
+        self.spm_target = load_spm(target_spm, self.sp_model_kwargs)
         self.current_spm = self.spm_source
 
         # Multilingual target side: default to using first supported language code.
@@ -172,7 +190,7 @@ class MarianTokenizer(PreTrainedTokenizer):
 
     def _tokenize(self, text: str) -> List[str]:
         code, text = self.remove_language_code(text)
-        pieces = self.current_spm.EncodeAsPieces(text)
+        pieces = self.current_spm.encode(text, out_type=str)
         return code + pieces
 
     def _convert_id_to_token(self, index: int) -> str:
@@ -283,7 +301,12 @@ class MarianTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d: Dict) -> None:
         self.__dict__ = d
-        self.spm_source, self.spm_target = (load_spm(f) for f in self.spm_files)
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.spm_source, self.spm_target = (load_spm(f, self.sp_model_kwargs) for f in self.spm_files)
         self.current_spm = self.spm_source
         self._setup_normalizer()
 
@@ -308,8 +331,8 @@ class MarianTokenizer(PreTrainedTokenizer):
             return self._special_token_mask(token_ids_0 + token_ids_1) + [1]
 
 
-def load_spm(path: str) -> sentencepiece.SentencePieceProcessor:
-    spm = sentencepiece.SentencePieceProcessor()
+def load_spm(path: str, sp_model_kwargs: Dict[str, Any]) -> sentencepiece.SentencePieceProcessor:
+    spm = sentencepiece.SentencePieceProcessor(**sp_model_kwargs)
     spm.Load(path)
     return spm
 

src/transformers/models/mbart/tokenization_mbart50.py

@@ -16,7 +16,7 @@
 import os
 from contextlib import contextmanager
 from shutil import copyfile
-from typing import Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -76,6 +76,20 @@ class MBart50Tokenizer(PreTrainedTokenizer):
         mask_token (:obj:`str`, `optional`, defaults to :obj:`"<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.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Examples::
 
@@ -108,11 +122,14 @@ class MBart50Tokenizer(PreTrainedTokenizer):
         unk_token="<unk>",
         pad_token="<pad>",
         mask_token="<mask>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             src_lang=src_lang,
             tgt_lang=tgt_lang,
@@ -122,10 +139,11 @@ class MBart50Tokenizer(PreTrainedTokenizer):
             cls_token=cls_token,
             pad_token=pad_token,
             mask_token=mask_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(str(vocab_file))
         self.vocab_file = vocab_file
 
@@ -177,7 +195,12 @@ class MBart50Tokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d: Dict) -> None:
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
     def get_vocab(self) -> Dict:
@@ -186,7 +209,7 @@ class MBart50Tokenizer(PreTrainedTokenizer):
         return vocab
 
     def _tokenize(self, text: str) -> List[str]:
-        return self.sp_model.EncodeAsPieces(text)
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token: str) -> int:
         """Converts a token (str) in an id using the vocab."""

src/transformers/models/pegasus/tokenization_pegasus.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 import os
 from shutil import copyfile
-from typing import Dict, List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -77,6 +77,20 @@ class PegasusTokenizer(PreTrainedTokenizer):
             tokenizer
             <https://github.com/google-research/pegasus/blob/939830367bcf411193d2b5eca2f2f90f3f9260ca/pegasus/ops/pretrain_parsing_ops.cc#L66>`__
             that uses the tokens 2 - 104 only for pretraining
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
     """
     vocab_files_names = VOCAB_FILES_NAMES
 
@@ -95,10 +109,10 @@ class PegasusTokenizer(PreTrainedTokenizer):
         mask_token_sent="<mask_1>",
         additional_special_tokens=None,
         offset=103,  # entries 2 - 104 are only used for pretraining
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         self.offset = offset
-
         if additional_special_tokens is not None:
             assert isinstance(
                 additional_special_tokens, list
@@ -123,6 +137,8 @@ class PegasusTokenizer(PreTrainedTokenizer):
             additional_special_tokens = [mask_token_sent] if mask_token_sent is not None else []
             additional_special_tokens += [f"<unk_{i}>" for i in range(2, self.offset)]
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             eos_token=eos_token,
             unk_token=unk_token,
@@ -131,11 +147,12 @@ class PegasusTokenizer(PreTrainedTokenizer):
             mask_token_sent=mask_token_sent,
             offset=offset,
             additional_special_tokens=additional_special_tokens,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
         self.mask_token_sent = mask_token_sent
         self.vocab_file = vocab_file
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
         # add special tokens to encoder dict
@@ -175,16 +192,17 @@ class PegasusTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Take as input a string and return a list of strings (tokens) for words/sub-words"""
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
-        return pieces
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token: str) -> int:
         """Converts a token (str) to an id using the vocab."""

src/transformers/models/reformer/tokenization_reformer.py

@@ -17,7 +17,7 @@
 
 import os
 from shutil import copyfile
-from typing import Dict, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -68,6 +68,20 @@ class ReformerTokenizer(PreTrainedTokenizer):
             The token used for padding, for example when batching sequences of different lengths.
         additional_special_tokens (:obj:`List[str]`, `optional`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
@@ -75,16 +89,27 @@ class ReformerTokenizer(PreTrainedTokenizer):
     max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
     model_input_names = ["input_ids", "attention_mask"]
 
-    def __init__(self, vocab_file, eos_token="</s>", unk_token="<unk>", additional_special_tokens=[], **kwargs):
+    def __init__(
+        self,
+        vocab_file,
+        eos_token="</s>",
+        unk_token="<unk>",
+        additional_special_tokens=[],
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             eos_token=eos_token,
             unk_token=unk_token,
             additional_special_tokens=additional_special_tokens,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
         self.vocab_file = vocab_file
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
     @property
@@ -103,16 +128,17 @@ class ReformerTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Take as input a string and return a list of strings (tokens) for words/sub-words"""
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
-        return pieces
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""

src/transformers/models/speech_to_text/tokenization_speech_to_text.py

@@ -17,7 +17,7 @@
 import json
 from pathlib import Path
 from shutil import copyfile
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 import sentencepiece
 
@@ -79,6 +79,21 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
             Whether or not to lowercase the input when tokenizing.
         tgt_lang (:obj:`str`, `optional`):
             A string representing the target language.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+
         **kwargs
             Additional keyword arguments passed along to :class:`~transformers.PreTrainedTokenizer`
     """
@@ -102,8 +117,11 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
         do_lower_case=False,
         tgt_lang=None,
         lang_codes=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs,
-    ):
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             bos_token=bos_token,
             eos_token=eos_token,
@@ -113,6 +131,7 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
             do_lower_case=do_lower_case,
             tgt_lang=tgt_lang,
             lang_codes=lang_codes,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
         self.do_upper_case = do_upper_case
@@ -121,7 +140,7 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
         self.encoder = load_json(vocab_file)
         self.decoder = {v: k for k, v in self.encoder.items()}
         self.spm_file = spm_file
-        self.sp_model = load_spm(spm_file)
+        self.sp_model = load_spm(spm_file, self.sp_model_kwargs)
 
         if lang_codes is not None:
             self.lang_codes = lang_codes
@@ -155,7 +174,7 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
         self.prefix_tokens = [lang_code_id]
 
     def _tokenize(self, text: str) -> List[str]:
-        return self.sp_model.EncodeAsPieces(text)
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         return self.encoder.get(token, self.encoder[self.unk_token])
@@ -221,7 +240,12 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d: Dict) -> None:
         self.__dict__ = d
-        self.sp_model = load_spm(self.spm_file)
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = load_spm(self.spm_file, self.sp_model_kwargs)
 
     def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
         save_dir = Path(save_directory)
@@ -241,8 +265,8 @@ class Speech2TextTokenizer(PreTrainedTokenizer):
         return (str(vocab_save_path), str(spm_save_path))
 
 
-def load_spm(path: str) -> sentencepiece.SentencePieceProcessor:
-    spm = sentencepiece.SentencePieceProcessor()
+def load_spm(path: str, sp_model_kwargs: Dict[str, Any]) -> sentencepiece.SentencePieceProcessor:
+    spm = sentencepiece.SentencePieceProcessor(**sp_model_kwargs)
     spm.Load(str(path))
     return spm
 

src/transformers/models/t5/tokenization_t5.py

@@ -19,7 +19,7 @@ import os
 import re
 import warnings
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -81,6 +81,20 @@ class T5Tokenizer(PreTrainedTokenizer):
             <https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117>`__).
         additional_special_tokens (:obj:`List[str]`, `optional`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Attributes:
         sp_model (:obj:`SentencePieceProcessor`):
@@ -100,8 +114,9 @@ class T5Tokenizer(PreTrainedTokenizer):
         pad_token="<pad>",
         extra_ids=100,
         additional_special_tokens=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Add extra_ids to the special token list
         if extra_ids > 0 and additional_special_tokens is None:
             additional_special_tokens = [f"<extra_id_{i}>" for i in range(extra_ids)]
@@ -114,19 +129,22 @@ class T5Tokenizer(PreTrainedTokenizer):
                     "In this case the additional_special_tokens must include the extra_ids tokens"
                 )
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             eos_token=eos_token,
             unk_token=unk_token,
             pad_token=pad_token,
             extra_ids=extra_ids,
             additional_special_tokens=additional_special_tokens,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
         self.vocab_file = vocab_file
         self._extra_ids = extra_ids
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
     @property
@@ -231,16 +249,17 @@ class T5Tokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Take as input a string and return a list of strings (tokens) for words/sub-words"""
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
-        return pieces
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""

src/transformers/models/xlm_prophetnet/tokenization_xlm_prophetnet.py

@@ -16,7 +16,7 @@
 import collections
 import os
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 from ...tokenization_utils import PreTrainedTokenizer
 from ...utils import logging
@@ -96,6 +96,20 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Attributes:
         sp_model (:obj:`SentencePieceProcessor`):
@@ -117,8 +131,11 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
         pad_token="[PAD]",
         cls_token="[CLS]",
         mask_token="[MASK]",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             bos_token=bos_token,
             eos_token=eos_token,
@@ -127,6 +144,7 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
             pad_token=pad_token,
             cls_token=cls_token,
             mask_token=mask_token,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
@@ -139,7 +157,7 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
             )
             raise
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(str(vocab_file))
         self.vocab_file = vocab_file
 
@@ -177,7 +195,12 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
                 "pip install sentencepiece"
             )
             raise
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
     def get_special_tokens_mask(
@@ -241,8 +264,8 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
         vocab.update(self.added_tokens_encoder)
         return vocab
 
-    def _tokenize(self, text):
-        return self.sp_model.EncodeAsPieces(text)
+    def _tokenize(self, text: str) -> str:
+        return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):
         """Converts a token (str) in an id using the vocab."""

src/transformers/models/xlm_roberta/tokenization_xlm_roberta.py

@@ -17,7 +17,7 @@
 
 import os
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -94,7 +94,7 @@ class XLMRobertaTokenizer(PreTrainedTokenizer):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
-        sp_model_kwargs (:obj:`dict`, `optional`, defaults to :obj:`None`):
+        sp_model_kwargs (:obj:`dict`, `optional`):
             Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
             <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
 
@@ -129,9 +129,9 @@ class XLMRobertaTokenizer(PreTrainedTokenizer):
         unk_token="<unk>",
         pad_token="<pad>",
         mask_token="<mask>",
-        sp_model_kwargs=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
@@ -271,7 +271,7 @@ class XLMRobertaTokenizer(PreTrainedTokenizer):
         vocab.update(self.added_tokens_encoder)
         return vocab
 
-    def _tokenize(self, text):
+    def _tokenize(self, text: str) -> List[str]:
         return self.sp_model.encode(text, out_type=str)
 
     def _convert_token_to_id(self, token):

src/transformers/models/xlnet/tokenization_xlnet.py

@@ -18,7 +18,7 @@
 import os
 import unicodedata
 from shutil import copyfile
-from typing import List, Optional, Tuple
+from typing import Any, Dict, List, Optional, Tuple
 
 import sentencepiece as spm
 
@@ -99,6 +99,20 @@ class XLNetTokenizer(PreTrainedTokenizer):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<eop>", "<eod>"]`):
             Additional special tokens used by the tokenizer.
+        sp_model_kwargs (:obj:`dict`, `optional`):
+            Will be passed to the ``SentencePieceProcessor.__init__()`` method. The `Python wrapper for SentencePiece
+            <https://github.com/google/sentencepiece/tree/master/python>`__ can be used, among other things, to set:
+
+            - ``enable_sampling``: Enable subword regularization.
+            - ``nbest_size``: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - ``nbest_size = {0,1}``: No sampling is performed.
+              - ``nbest_size > 1``: samples from the nbest_size results.
+              - ``nbest_size < 0``: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - ``alpha``: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
 
     Attributes:
         sp_model (:obj:`SentencePieceProcessor`):
@@ -124,11 +138,14 @@ class XLNetTokenizer(PreTrainedTokenizer):
         cls_token="<cls>",
         mask_token="<mask>",
         additional_special_tokens=["<eop>", "<eod>"],
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs
-    ):
+    ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
         super().__init__(
             do_lower_case=do_lower_case,
             remove_space=remove_space,
@@ -141,6 +158,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
             cls_token=cls_token,
             mask_token=mask_token,
             additional_special_tokens=additional_special_tokens,
+            sp_model_kwargs=self.sp_model_kwargs,
             **kwargs,
         )
 
@@ -151,7 +169,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
         self.keep_accents = keep_accents
         self.vocab_file = vocab_file
 
-        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(vocab_file)
 
     @property
@@ -170,7 +188,12 @@ class XLNetTokenizer(PreTrainedTokenizer):
 
     def __setstate__(self, d):
         self.__dict__ = d
-        self.sp_model = spm.SentencePieceProcessor()
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
         self.sp_model.Load(self.vocab_file)
 
     def preprocess_text(self, inputs):
@@ -188,14 +211,10 @@ class XLNetTokenizer(PreTrainedTokenizer):
 
         return outputs
 
-    def _tokenize(self, text, sample=False):
+    def _tokenize(self, text: str) -> List[str]:
         """Tokenize a string."""
         text = self.preprocess_text(text)
-
-        if not sample:
-            pieces = self.sp_model.EncodeAsPieces(text)
-        else:
-            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
+        pieces = self.sp_model.encode(text, out_type=str)
         new_pieces = []
         for piece in pieces:
             if len(piece) > 1 and piece[-1] == str(",") and piece[-2].isdigit():

tests/test_tokenization_albert.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import os
 import unittest
 
@@ -33,6 +32,8 @@ class AlbertTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     tokenizer_class = AlbertTokenizer
     rust_tokenizer_class = AlbertTokenizerFast
     test_rust_tokenizer = True
+    test_sentencepiece = True
+    test_sentencepiece_ignore_case = True
 
     def setUp(self):
         super().setUp()

tests/test_tokenization_barthez.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import unittest
 
 from transformers import BarthezTokenizer, BarthezTokenizerFast, BatchEncoding
@@ -24,12 +23,13 @@ from .test_tokenization_common import TokenizerTesterMixin
 
 @require_tokenizers
 @require_sentencepiece
-@slow
+@slow  # see https://github.com/huggingface/transformers/issues/11457
 class BarthezTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
 
     tokenizer_class = BarthezTokenizer
     rust_tokenizer_class = BarthezTokenizerFast
     test_rust_tokenizer = True
+    test_sentencepiece = True
 
     def setUp(self):
         super().setUp()

tests/test_tokenization_bert_generation.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import os
 import unittest
 
@@ -33,6 +32,7 @@ SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)), "fixture
 class BertGenerationTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
 
     tokenizer_class = BertGenerationTokenizer
+    test_sentencepiece = True
 
     def setUp(self):
         super().setUp()

tests/test_tokenization_big_bird.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import os
 import unittest
 
@@ -36,11 +35,12 @@ class BigBirdTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
     tokenizer_class = BigBirdTokenizer
     rust_tokenizer_class = BigBirdTokenizerFast
     test_rust_tokenizer = True
+    test_sentencepiece = True
 
     def setUp(self):
         super().setUp()
 
-        tokenizer = BigBirdTokenizer(SAMPLE_VOCAB, keep_accents=True)
+        tokenizer = self.tokenizer_class(SAMPLE_VOCAB, keep_accents=True)
         tokenizer.save_pretrained(self.tmpdirname)
 
     def test_rust_and_python_full_tokenizers(self):