Skip to content

Latest commit



370 lines (291 loc) · 11.1 KB

File metadata and controls

370 lines (291 loc) · 11.1 KB


pyonmttok is the Python wrapper for OpenNMT/Tokenizer, a fast and customizable text tokenization library with BPE and SentencePiece support.


pip install pyonmttok


  • OS: Linux, macOS, Windows
  • Python version: >= 3.6
  • pip version: >= 19.3

Table of contents

  1. Tokenization
  2. Subword learning
  3. Vocabulary
  4. Token API
  5. Utilities



>>> import pyonmtok
>>> tokenizer = pyonmttok.Tokenizer("aggressive", joiner_annotate=True)
>>> tokens = tokenizer("Hello World!")
>>> tokens
['Hello', 'World', '■!']
>>> tokenizer.detokenize(tokens)
'Hello World!'



tokenizer = pyonmttok.Tokenizer(
    mode: str,
    lang: Optional[str] = None,
    bpe_model_path: Optional[str] = None,
    bpe_dropout: float = 0,
    vocabulary: Optional[List[str]] = None,
    vocabulary_path: Optional[str] = None,
    vocabulary_threshold: int = 0,
    sp_model_path: Optional[str] = None,
    sp_nbest_size: int = 0,
    sp_alpha: float = 0.1,
    joiner: str = "■",
    joiner_annotate: bool = False,
    joiner_new: bool = False,
    support_prior_joiners: bool = False,
    spacer_annotate: bool = False,
    spacer_new: bool = False,
    case_feature: bool = False,
    case_markup: bool = False,
    soft_case_regions: bool = False,
    no_substitution: bool = False,
    with_separators: bool = False,
    allow_isolated_marks: bool = False,
    preserve_placeholders: bool = False,
    preserve_segmented_tokens: bool = False,
    segment_case: bool = False,
    segment_numbers: bool = False,
    segment_alphabet_change: bool = False,
    segment_alphabet: Optional[List[str]] = None,

# SentencePiece-compatible tokenizer.
tokenizer = pyonmttok.SentencePieceTokenizer(
    model_path: str,
    vocabulary_path: Optional[str] = None,
    vocabulary_threshold: int = 0,
    nbest_size: int = 0,
    alpha: float = 0.1,

# Copy constructor.
tokenizer = pyonmttok.Tokenizer(tokenizer: pyonmttok.Tokenizer)

# Return the tokenization options (excluding options related to subword).

See the documentation for a description of each tokenization option.


# Tokenize a text.
# When training=False, subword regularization such as BPE dropout is disabled.
tokenizer.__call__(text: str, training: bool = True) -> List[str]

# Tokenize a text and return optional features.
# When as_token_objects=True, the method returns Token objects (see below).
    text: str,
    as_token_objects: bool = False,
    training: bool = True,
) -> Union[Tuple[List[str], Optional[List[List[str]]]], List[pyonmttok.Token]]

# Tokenize a batch of text.
    batch_text: List[str],
    as_token_objects: bool = False,
    training: bool = True,
) -> Union[Tuple[List[List[str]], List[Optional[List[List[str]]]]], List[List[pyonmttok.Token]]]

# Tokenize a file.
    input_path: str,
    output_path: str,
    num_threads: int = 1,
    verbose: bool = False,
    training: bool = True,
    tokens_delimiter: str = " ",


# The detokenize method converts a list of tokens back to a string.
    tokens: List[str],
    features: Optional[List[List[str]]] = None,
) -> str
tokenizer.detokenize(tokens: List[pyonmttok.Token]) -> str

# The detokenize_with_ranges method also returns a dictionary mapping a token
# index to a range in the detokenized text.
# Set merge_ranges=True to merge consecutive ranges, e.g. subwords of the same
# token in case of subword tokenization.
# Set unicode_ranges=True to return ranges over Unicode characters instead of bytes.
    tokens: Union[List[str], List[pyonmttok.Token]],
    merge_ranges: bool = False,
    unicode_ranges: bool = False,
) -> Tuple[str, Dict[int, Tuple[int, int]]]

# Detokenize a file.
    input_path: str,
    output_path: str,
    tokens_delimiter: str = " ",

Subword learning


The Python wrapper supports BPE and SentencePiece subword learning through a common interface:

1. Create the subword learner with the tokenization you want to apply, e.g.:

# BPE is trained and applied on the tokenization output before joiner (or spacer) annotations.
tokenizer = pyonmttok.Tokenizer("aggressive", joiner_annotate=True, segment_numbers=True)
learner = pyonmttok.BPELearner(tokenizer=tokenizer, symbols=32000)

# SentencePiece can learn from raw sentences so a tokenizer in not required.
learner = pyonmttok.SentencePieceLearner(vocab_size=32000, character_coverage=0.98)

2. Feed some raw data:

# Feed detokenized sentences:
learner.ingest("Hello world!")
learner.ingest("How are you?")

# or detokenized text files:

3. Start the learning process:

tokenizer = learner.learn("/data/model-32k")

The returned tokenizer instance can be used to apply subword tokenization on new data.


# See
# for argument documentation.
learner = pyonmttok.BPELearner(
    tokenizer: Optional[pyonmttok.Tokenizer] = None,  # Defaults to tokenization mode "space".
    symbols: int = 10000,
    min_frequency: int = 2,
    total_symbols: bool = False,

# See
# for available training options.
learner = pyonmttok.SentencePieceLearner(
    tokenizer: Optional[pyonmttok.Tokenizer] = None,  # Defaults to tokenization mode "none".
    keep_vocab: bool = False,  # Keep the generated vocabulary (model_path will act like model_prefix in spm_train)

learner.ingest(text: str)
learner.ingest_file(path: str)
learner.ingest_token(token: Union[str, pyonmttok.Token])

learner.learn(model_path: str, verbose: bool = False) -> pyonmttok.Tokenizer



tokenizer = pyonmttok.Tokenizer("aggressive", joiner_annotate=True)

with open("train.txt") as train_file:
    vocab = pyonmttok.build_vocab_from_lines(
        special_tokens=["<blank>", "<unk>", "<s>", "</s>"],

with open("vocab.txt", "w") as vocab_file:
    for token in vocab.ids_to_tokens:
        vocab_file.write("%s\n" % token)


# Special tokens are added with ids 0, 1, etc., and are never removed by a resize.
vocab = pyonmttok.Vocab(special_tokens: Optional[List[str]] = None)

# Read-only properties.
vocab.tokens_to_ids -> Dict[str, int]
vocab.ids_to_tokens -> List[str]
vocab.counters -> List[int]

# Get or set the ID returned for out-of-vocabulary tokens.
# By default, it is the ID of the token <unk> if present in the vocabulary, len(vocab) otherwise.
vocab.default_id -> int

vocab.lookup_token(token: str) -> int
vocab.lookup_index(index: int) -> str

# Calls lookup_token on a batch of tokens.
vocab.__call__(tokens: List[str]) -> List[int]

vocab.__len__() -> int                  # Implements: len(vocab)
vocab.__contains__(token: str) -> bool  # Implements: "hello" in vocab
vocab.__getitem__(token: str) -> int    # Implements: vocab["hello"]

# Add tokens to the vocabulary after tokenization.
# If a tokenizer is not set, the text is split on spaces.
vocab.add_from_text(text: str, tokenizer: Optional[pyonmttok.Tokenizer] = None) -> None
vocab.add_from_file(path: str, tokenizer: Optional[pyonmttok.Tokenizer] = None) -> None
vocab.add_token(token: str, count: int = 1) -> None

vocab.resize(maximum_size: int = 0, minimum_frequency: int = 1) -> None

# Build a vocabulary from an iterator of lines.
# If a tokenizer is not set, the lines are split on spaces.
    lines: Iterable[str],
    tokenizer: Optional[pyonmttok.Tokenizer] = None,
    maximum_size: int = 0,
    minimum_frequency: int = 1,
    special_tokens: Optional[List[str]] = None,
) -> pyonmttok.Vocab

# Build a vocabulary from an iterator of tokens.
    tokens: Iterable[str],
    maximum_size: int = 0,
    minimum_frequency: int = 1,
    special_tokens: Optional[List[str]] = None,
) -> pyonmttok.Vocab

Token API

The Token API allows to tokenize text into pyonmttok.Token objects. This API can be useful to apply some logics at the token level but still retain enough information to write the tokenization on disk or detokenize.


>>> tokenizer = pyonmttok.Tokenizer("aggressive", joiner_annotate=True)
>>> tokens = tokenizer.tokenize("Hello World!", as_token_objects=True)
>>> tokens
[Token('Hello'), Token('World'), Token('!', join_left=True)]
>>> tokens[-1].surface
>>> tokenizer.serialize_tokens(tokens)[0]
['Hello', 'World', '■!']
>>> tokens[-1].surface = '.'
>>> tokenizer.serialize_tokens(tokens)[0]
['Hello', 'World', '■.']
>>> tokenizer.detokenize(tokens)
'Hello World.'


The pyonmttok.Token class has the following attributes:

  • surface: a string, the token value
  • type: a pyonmttok.TokenType value, the type of the token
  • join_left: a boolean, whether the token should be joined to the token on the left or not
  • join_right: a boolean, whether the token should be joined to the token on the right or not
  • preserve: a boolean, whether joiners and spacers can be attached to this token or not
  • features: a list of string, the features attached to the token
  • spacer: a boolean, whether the token is prefixed by a SentencePiece spacer or not (only set when using SentencePiece)
  • casing: a pyonmttok.Casing value, the casing of the token (only set when tokenizing with case_feature or case_markup)

The pyonmttok.TokenType enumeration is used to identify tokens that were split by a subword tokenization. The enumeration has the following values:

  • TokenType.WORD

The pyonmttok.Casing enumeration is used to identify the original casing of a token that was lowercased by the case_feature or case_markup tokenization options. The enumeration has the following values:

  • Casing.LOWERCASE
  • Casing.UPPERCASE
  • Casing.MIXED
  • Casing.NONE

The Tokenizer instances provide methods to serialize or deserialize Token objects:

# Serialize Token objects to strings that can be saved on disk.
    tokens: List[pyonmttok.Token],
) -> Tuple[List[str], Optional[List[List[str]]]]

# Deserialize strings into Token objects.
    tokens: List[str],
    features: Optional[List[List[str]]] = None,
) -> List[pyonmttok.Token]



# Returns True if the string has the placeholder format.
pyonmttok.is_placeholder(token: str)

# Sets the random seed for reproducible tokenization.
pyonmttok.set_random_seed(seed: int)

# Checks if the language code is valid.
pyonmttok.is_valid_language(lang: str).