Merge branch 'master' into fix-xlnet-squad2.0

transformers/tests/utils.py

+import os
+import unittest
+import tempfile
+
+from distutils.util import strtobool
+
+from transformers.file_utils import _tf_available, _torch_available
+
+
+CACHE_DIR = os.path.join(tempfile.gettempdir(), "transformers_test")
+
+SMALL_MODEL_IDENTIFIER = "julien-c/bert-xsmall-dummy"
+
+
+def parse_flag_from_env(key, default=False):
+    try:
+        value = os.environ[key]
+    except KeyError:
+        # KEY isn't set, default to `default`.
+        _value = default
+    else:
+        # KEY is set, convert it to True or False.
+        try:
+            _value = strtobool(value)
+        except ValueError:
+            # More values are supported, but let's keep the message simple.
+            raise ValueError("If set, {} must be yes or no.".format(key))
+    return _value
+
+_run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
+_run_custom_tokenizers = parse_flag_from_env("RUN_CUSTOM_TOKENIZERS", default=False)
+
+
+def slow(test_case):
+    """
+    Decorator marking a test as slow.
+
+    Slow tests are skipped by default. Set the RUN_SLOW environment variable
+    to a truthy value to run them.
+
+    """
+    if not _run_slow_tests:
+        test_case = unittest.skip("test is slow")(test_case)
+    return test_case
+
+
+def custom_tokenizers(test_case):
+    """
+    Decorator marking a test for a custom tokenizer.
+
+    Custom tokenizers require additional dependencies, and are skipped
+    by default. Set the RUN_CUSTOM_TOKENIZERS environment variable
+    to a truthy value to run them.
+    """
+    if not _run_custom_tokenizers:
+        test_case = unittest.skip("test of custom tokenizers")(test_case)
+    return test_case
+
+
+def require_torch(test_case):
+    """
+    Decorator marking a test that requires PyTorch.
+
+    These tests are skipped when PyTorch isn't installed.
+
+    """
+    if not _torch_available:
+        test_case = unittest.skip("test requires PyTorch")(test_case)
+    return test_case
+
+
+def require_tf(test_case):
+    """
+    Decorator marking a test that requires TensorFlow.
+
+    These tests are skipped when TensorFlow isn't installed.
+
+    """
+    if not _tf_available:
+        test_case = unittest.skip("test requires TensorFlow")(test_case)
+    return test_case
+
+
+if _torch_available:
+    # Set the USE_CUDA environment variable to select a GPU.
+    torch_device = "cuda" if parse_flag_from_env("USE_CUDA") else "cpu"
+else:
+    torch_device = None

transformers/tokenization_albert.py

+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain 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.
+""" Tokenization classes for ALBERT model."""
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+from .tokenization_utils import PreTrainedTokenizer
+import logging
+import unicodedata
+import six
+import os
+from shutil import copyfile
+
+logger = logging.getLogger(__name__)
+VOCAB_FILES_NAMES = {'vocab_file': 'spiece.model'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-spiece.model",
+        'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-spiece.model",
+        'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-spiece.model",
+        'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-spiece.model",
+        'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-spiece.model",
+        'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-spiece.model",
+        'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-spiece.model",
+        'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-spiece.model",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'albert-base-v1': 512,
+    'albert-large-v1': 512,
+    'albert-xlarge-v1': 512,
+    'albert-xxlarge-v1': 512,
+    'albert-base-v2': 512,
+    'albert-large-v2': 512,
+    'albert-xlarge-v2': 512,
+    'albert-xxlarge-v2': 512,
+}
+
+SPIECE_UNDERLINE = u'▁'
+
+class AlbertTokenizer(PreTrainedTokenizer):
+    """
+        SentencePiece based tokenizer. Peculiarities:
+
+            - requires `SentencePiece <https://github.com/google/sentencepiece>`_
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file,
+                 do_lower_case=True, remove_space=True, keep_accents=False,
+                 bos_token="[CLS]", eos_token="[SEP]", unk_token="<unk>", sep_token="[SEP]",
+                 pad_token="<pad>", cls_token="[CLS]", mask_token="[MASK]", **kwargs):
+        super(AlbertTokenizer, self).__init__(bos_token=bos_token, eos_token=eos_token,
+                                             unk_token=unk_token, sep_token=sep_token,
+                                             pad_token=pad_token, cls_token=cls_token,
+                                             mask_token=mask_token, **kwargs)
+
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning("You need to install SentencePiece to use AlbertTokenizer: https://github.com/google/sentencepiece"
+                           "pip install sentencepiece")
+
+        self.do_lower_case = do_lower_case
+        self.remove_space = remove_space
+        self.keep_accents = keep_accents
+        self.vocab_file = vocab_file
+
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(vocab_file)
+
+    @property
+    def vocab_size(self):
+        return len(self.sp_model)
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning("You need to install SentencePiece to use AlbertTokenizer: https://github.com/google/sentencepiece"
+                           "pip install sentencepiece")
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(self.vocab_file)
+
+    def preprocess_text(self, inputs):
+        if self.remove_space:
+            outputs = ' '.join(inputs.strip().split())
+        else:
+            outputs = inputs
+        outputs = outputs.replace("``", '"').replace("''", '"')
+
+        if six.PY2 and isinstance(outputs, str):
+            outputs = outputs.decode('utf-8')
+
+        if not self.keep_accents:
+            outputs = unicodedata.normalize('NFKD', outputs)
+            outputs = ''.join([c for c in outputs if not unicodedata.combining(c)])
+        if self.do_lower_case:
+            outputs = outputs.lower()
+
+        return outputs
+
+    def _tokenize(self, text, return_unicode=True, sample=False):
+        """ Tokenize a string.
+            return_unicode is used only for py2
+        """
+        text = self.preprocess_text(text)
+        # note(zhiliny): in some systems, sentencepiece only accepts str for py2
+        if six.PY2 and isinstance(text, unicode):
+            text = text.encode('utf-8')
+
+        if not sample:
+            pieces = self.sp_model.EncodeAsPieces(text)
+        else:
+            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
+        new_pieces = []
+        for piece in pieces:
+            if len(piece) > 1 and piece[-1] == str(',') and piece[-2].isdigit():
+                cur_pieces = self.sp_model.EncodeAsPieces(
+                    piece[:-1].replace(SPIECE_UNDERLINE, ''))
+                if piece[0] != SPIECE_UNDERLINE and cur_pieces[0][0] == SPIECE_UNDERLINE:
+                    if len(cur_pieces[0]) == 1:
+                        cur_pieces = cur_pieces[1:]
+                    else:
+                        cur_pieces[0] = cur_pieces[0][1:]
+                cur_pieces.append(piece[-1])
+                new_pieces.extend(cur_pieces)
+            else:
+                new_pieces.append(piece)
+
+        # note(zhiliny): convert back to unicode for py2
+        if six.PY2 and return_unicode:
+            ret_pieces = []
+            for piece in new_pieces:
+                if isinstance(piece, str):
+                    piece = piece.decode('utf-8')
+                ret_pieces.append(piece)
+            new_pieces = ret_pieces
+
+        return new_pieces
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        return self.sp_model.PieceToId(token)
+
+    def _convert_id_to_token(self, index, return_unicode=True):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        if six.PY2 and return_unicode and isinstance(token, str):
+            token = token.decode('utf-8')
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (strings for sub-words) in a single string."""
+        out_string = ''.join(tokens).replace(SPIECE_UNDERLINE, ' ').strip()
+        return out_string
+
+    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.
+        An ALBERT sequence has the following format:
+            single sequence: [CLS] X [SEP]
+            pair of sequences: [CLS] A [SEP] B [SEP]
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return cls + token_ids_0 + sep
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` or ``encode_plus`` methods.
+
+        Args:
+            token_ids_0: list of ids (must not contain special tokens)
+            token_ids_1: Optional list of ids (must not contain special tokens), necessary when fetching sequence ids
+                for sequence pairs
+            already_has_special_tokens: (default False) Set to True if the token list is already formated with
+                special tokens for the model
+
+        Returns:
+            A list of integers in the range [0, 1]: 0 for a special token, 1 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError("You should not supply a second sequence if the provided sequence of "
+                                 "ids is already formated with special tokens for the model.")
+            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        An ALBERT 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 1
+        | first sequence    | second sequence
+
+        if token_ids_1 is None, only returns the first portion of the mask (0'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):
+        """ Save the sentencepiece vocabulary (copy original file) and special tokens file
+            to a directory.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        out_vocab_file = os.path.join(save_directory, VOCAB_FILES_NAMES['vocab_file'])
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)

transformers/tokenization_auto.py

@@ -19,6 +19,7 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 import logging
 
 from .tokenization_bert import BertTokenizer
+from .tokenization_bert_japanese import BertJapaneseTokenizer
 from .tokenization_openai import OpenAIGPTTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer
 from .tokenization_ctrl import CTRLTokenizer
@@ -27,6 +28,10 @@ from .tokenization_xlnet import XLNetTokenizer
 from .tokenization_xlm import XLMTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer
+from .tokenization_camembert import CamembertTokenizer
+from .tokenization_albert import AlbertTokenizer
+from .tokenization_t5 import T5Tokenizer
+from .tokenization_xlm_roberta import XLMRobertaTokenizer
 
 logger = logging.getLogger(__name__)
 
@@ -41,15 +46,19 @@ class AutoTokenizer(object):
 
         The tokenizer class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Tokenizer (T5 model)
             - contains `distilbert`: DistilBertTokenizer (DistilBert model)
+            - contains `albert`: AlbertTokenizer (ALBERT model)
+            - contains `camembert`: CamembertTokenizer (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaTokenizer (XLM-RoBERTa model)
             - contains `roberta`: RobertaTokenizer (RoBERTa model)
             - contains `bert`: BertTokenizer (Bert model)
             - contains `openai-gpt`: OpenAIGPTTokenizer (OpenAI GPT model)
             - contains `gpt2`: GPT2Tokenizer (OpenAI GPT-2 model)
-            - contains `ctrl`: CTRLTokenizer (Salesforce CTRL model)
             - contains `transfo-xl`: TransfoXLTokenizer (Transformer-XL model)
             - contains `xlnet`: XLNetTokenizer (XLNet model)
             - contains `xlm`: XLMTokenizer (XLM model)
+            - contains `ctrl`: CTRLTokenizer (Salesforce CTRL model)
 
         This class cannot be instantiated using `__init__()` (throw an error).
     """
@@ -64,20 +73,26 @@ class AutoTokenizer(object):
 
         The tokenizer class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Tokenizer (T5 model)
             - contains `distilbert`: DistilBertTokenizer (DistilBert model)
-            - contains `roberta`: RobertaTokenizer (XLM model)
+            - contains `albert`: AlbertTokenizer (ALBERT model)
+            - contains `camembert`: CamembertTokenizer (CamemBERT model)
+            - contains `xlm-roberta`: XLMRobertaTokenizer (XLM-RoBERTa model)
+            - contains `roberta`: RobertaTokenizer (RoBERTa model)
+            - contains `bert-base-japanese`: BertJapaneseTokenizer (Bert model)
             - contains `bert`: BertTokenizer (Bert model)
             - contains `openai-gpt`: OpenAIGPTTokenizer (OpenAI GPT model)
             - contains `gpt2`: GPT2Tokenizer (OpenAI GPT-2 model)
-            - contains `ctrl`: CTRLTokenizer (Salesforce CTRL model)
             - contains `transfo-xl`: TransfoXLTokenizer (Transformer-XL model)
             - contains `xlnet`: XLNetTokenizer (XLNet model)
             - contains `xlm`: XLMTokenizer (XLM model)
+            - contains `ctrl`: CTRLTokenizer (Salesforce CTRL model)
 
         Params:
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a predefined tokenizer to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a predefined tokenizer that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing vocabulary files required by the tokenizer, for instance saved using the :func:`~transformers.PreTrainedTokenizer.save_pretrained` method, e.g.: ``./my_model_directory/``.
                 - (not applicable to all derived classes) a path or url to a single saved vocabulary file if and only if the tokenizer only requires a single vocabulary file (e.g. Bert, XLNet), e.g.: ``./my_model_directory/vocab.txt``.
 
@@ -87,6 +102,9 @@ class AutoTokenizer(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the vocabulary files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
             proxies: (`optional`) dict, default None:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
@@ -97,14 +115,30 @@ class AutoTokenizer(object):
 
         Examples::
 
-            tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')    # Download vocabulary from S3 and cache.
-            tokenizer = AutoTokenizer.from_pretrained('./test/bert_saved_model/')  # E.g. tokenizer was saved using `save_pretrained('./test/saved_model/')`
+            # Download vocabulary from S3 and cache.
+            tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
+
+            # Download vocabulary from S3 (user-uploaded) and cache.
+            tokenizer = AutoTokenizer.from_pretrained('dbmdz/bert-base-german-cased')
+
+            # If vocabulary files are in a directory (e.g. tokenizer was saved using `save_pretrained('./test/saved_model/')`)
+            tokenizer = AutoTokenizer.from_pretrained('./test/bert_saved_model/')
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5Tokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return AlbertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'camembert' in pretrained_model_name_or_path:
+            return CamembertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'xlm-roberta' in pretrained_model_name_or_path:
+            return XLMRobertaTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return RobertaTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'bert-base-japanese' in pretrained_model_name_or_path:
+            return BertJapaneseTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'openai-gpt' in pretrained_model_name_or_path:
@@ -121,4 +155,4 @@ class AutoTokenizer(object):
             return CTRLTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
                          "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
-                         "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
+                         "'xlm-roberta', 'xlm', 'roberta', 'distilbert,' 'camembert', 'ctrl', 'albert'".format(pretrained_model_name_or_path))

transformers/tokenization_bert.py

@@ -46,6 +46,8 @@ PRETRAINED_VOCAB_FILES_MAP = {
         'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-vocab.txt",
         'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-vocab.txt",
         'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-vocab.txt",
+        'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/vocab.txt",
+        'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/vocab.txt",
     }
 }
 
@@ -65,6 +67,8 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     'bert-base-cased-finetuned-mrpc': 512,
     'bert-base-german-dbmdz-cased': 512,
     'bert-base-german-dbmdz-uncased': 512,
+    'bert-base-finnish-cased-v1': 512,
+    'bert-base-finnish-uncased-v1': 512,
 }
 
 PRETRAINED_INIT_CONFIGURATION = {
@@ -83,6 +87,8 @@ PRETRAINED_INIT_CONFIGURATION = {
     'bert-base-cased-finetuned-mrpc': {'do_lower_case': False},
     'bert-base-german-dbmdz-cased': {'do_lower_case': False},
     'bert-base-german-dbmdz-uncased': {'do_lower_case': True},
+    'bert-base-finnish-cased-v1': {'do_lower_case': False},
+    'bert-base-finnish-uncased-v1': {'do_lower_case': True},
 }
 
 
@@ -113,12 +119,12 @@ class BertTokenizer(PreTrainedTokenizer):
 
     Args:
         vocab_file: Path to a one-wordpiece-per-line vocabulary file
-        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
         do_basic_tokenize: Whether to do basic tokenization before wordpiece.
         max_len: An artificial maximum length to truncate tokenized sequences to; Effective maximum length is always the
             minimum of this value (if specified) and the underlying BERT model's sequence length.
         never_split: List of tokens which will never be split during tokenization. Only has an effect when
-            do_wordpiece_only=False
+            do_basic_tokenize=True
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

transformers/tokenization_bert_japanese.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team 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.
+"""Tokenization classes."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import collections
+import logging
+import os
+import six
+import unicodedata
+from io import open
+
+from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer, load_vocab
+from .tokenization_utils import PreTrainedTokenizer
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {'vocab_file': 'vocab.txt'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-vocab.txt",
+        'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-vocab.txt",
+        'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-vocab.txt",
+        'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-vocab.txt"
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'bert-base-japanese': 512,
+    'bert-base-japanese-whole-word-masking': 512,
+    'bert-base-japanese-char': 512,
+    'bert-base-japanese-char-whole-word-masking': 512
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    'bert-base-japanese': {
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'wordpiece'
+    },
+    'bert-base-japanese-whole-word-masking':{
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'wordpiece'
+    },
+    'bert-base-japanese-char': {
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'character'
+    },
+    'bert-base-japanese-char-whole-word-masking': {
+        'do_lower_case': False,
+        'word_tokenizer_type': 'mecab',
+        'subword_tokenizer_type': 'character'
+    }
+}
+
+
+class BertJapaneseTokenizer(BertTokenizer):
+    """BERT tokenizer for Japanese text"""
+
+    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
+
+    def __init__(self, vocab_file, do_lower_case=False,
+                 do_word_tokenize=True, do_subword_tokenize=True,
+                 word_tokenizer_type='basic', subword_tokenizer_type='wordpiece',
+                 never_split=None, unk_token='[UNK]', sep_token='[SEP]',
+                 pad_token='[PAD]', cls_token='[CLS]', mask_token='[MASK]', **kwargs):
+        """Constructs a MecabBertTokenizer.
+
+        Args:
+            **vocab_file**: Path to a one-wordpiece-per-line vocabulary file.
+            **do_lower_case**: (`optional`) boolean (default True)
+                Whether to lower case the input.
+                Only has an effect when do_basic_tokenize=True.
+            **do_word_tokenize**: (`optional`) boolean (default True)
+                Whether to do word tokenization.
+            **do_subword_tokenize**: (`optional`) boolean (default True)
+                Whether to do subword tokenization.
+            **word_tokenizer_type**: (`optional`) string (default "basic")
+                Type of word tokenizer.
+            **subword_tokenizer_type**: (`optional`) string (default "wordpiece")
+                Type of subword tokenizer.
+        """
+        super(BertTokenizer, self).__init__(unk_token=unk_token, sep_token=sep_token,
+                                            pad_token=pad_token, cls_token=cls_token,
+                                            mask_token=mask_token, **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file))
+        self.vocab = load_vocab(vocab_file)
+        self.ids_to_tokens = collections.OrderedDict(
+            [(ids, tok) for tok, ids in self.vocab.items()])
+
+        self.do_word_tokenize = do_word_tokenize
+        if do_word_tokenize:
+            if word_tokenizer_type == 'basic':
+                self.word_tokenizer = BasicTokenizer(do_lower_case=do_lower_case,
+                                                     never_split=never_split,
+                                                     tokenize_chinese_chars=False)
+            elif word_tokenizer_type == 'mecab':
+                self.word_tokenizer = MecabTokenizer(do_lower_case=do_lower_case,
+                                                     never_split=never_split)
+            else:
+                raise ValueError(
+                    "Invalid word_tokenizer_type '{}' is specified.".format(word_tokenizer_type))
+
+        self.do_subword_tokenize = do_subword_tokenize
+        if do_subword_tokenize:
+            if subword_tokenizer_type == 'wordpiece':
+                self.subword_tokenizer = WordpieceTokenizer(vocab=self.vocab,
+                                                            unk_token=self.unk_token)
+            elif subword_tokenizer_type == 'character':
+                self.subword_tokenizer = CharacterTokenizer(vocab=self.vocab,
+                                                            unk_token=self.unk_token)
+            else:
+                raise ValueError(
+                    "Invalid subword_tokenizer_type '{}' is specified.".format(subword_tokenizer_type))
+
+
+    def _tokenize(self, text):
+        if self.do_word_tokenize:
+            tokens = self.word_tokenizer.tokenize(text,
+                                                  never_split=self.all_special_tokens)
+        else:
+            tokens = [text]
+
+        if self.do_subword_tokenize:
+            split_tokens = [sub_token for token in tokens
+                            for sub_token in self.subword_tokenizer.tokenize(token)]
+        else:
+            split_tokens = tokens
+
+        return split_tokens
+
+
+class MecabTokenizer(object):
+    """Runs basic tokenization with MeCab morphological parser."""
+
+    def __init__(self, do_lower_case=False, never_split=None, normalize_text=True):
+        """Constructs a MecabTokenizer.
+
+        Args:
+            **do_lower_case**: (`optional`) boolean (default True)
+                Whether to lower case the input.
+            **never_split**: (`optional`) list of str
+                Kept for backward compatibility purposes.
+                Now implemented directly at the base class level (see :func:`PreTrainedTokenizer.tokenize`)
+                List of token not to split.
+            **normalize_text**: (`optional`) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+        """
+        self.do_lower_case = do_lower_case
+        self.never_split = never_split if never_split is not None else []
+        self.normalize_text = normalize_text
+
+        import MeCab
+        self.mecab = MeCab.Tagger()
+
+    def tokenize(self, text, never_split=None, **kwargs):
+        """Tokenizes a piece of text."""
+        if self.normalize_text:
+            text = unicodedata.normalize('NFKC', text)
+
+        never_split = self.never_split + (never_split if never_split is not None else [])
+        tokens = []
+
+        if six.PY2:
+            mecab_output = self.mecab.parse(text.encode('utf-8')).decode('utf-8')
+        else:
+            mecab_output = self.mecab.parse(text)
+
+        cursor = 0
+        for line in mecab_output.split('\n'):
+            if line == 'EOS':
+                break
+
+            token, _ = line.split('\t')
+            token_start = text.index(token, cursor)
+            token_end = token_start + len(token)
+            if self.do_lower_case and token not in never_split:
+                token = token.lower()
+
+            tokens.append(token)
+            cursor = token_end
+
+        return tokens
+
+
+class CharacterTokenizer(object):
+    """Runs Character tokenziation."""
+
+    def __init__(self, vocab, unk_token, normalize_text=True):
+        """Constructs a CharacterTokenizer.
+
+        Args:
+            **vocab**:
+                Vocabulary object.
+            **unk_token**: str
+                A special symbol for out-of-vocabulary token.
+            **normalize_text**: (`optional`) boolean (default True)
+                Whether to apply unicode normalization to text before tokenization.
+        """
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.normalize_text = normalize_text
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into characters.
+
+        For example:
+            input = "apple"
+            output = ["a", "p", "p", "l", "e"]
+        Args:
+            text: A single token or whitespace separated tokens.
+                This should have already been passed through `BasicTokenizer`.
+        Returns:
+            A list of characters.
+        """
+        if self.normalize_text:
+            text = unicodedata.normalize('NFKC', text)
+
+        output_tokens = []
+        for i, char in enumerate(text):
+            if char not in self.vocab:
+                output_tokens.append(self.unk_token)
+                continue
+
+            output_tokens.append(char)
+
+        return output_tokens

transformers/tokenization_camembert.py

+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University 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
+""" Tokenization classes for Camembert model."""
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+import os
+from shutil import copyfile
+
+import sentencepiece as spm
+from transformers.tokenization_utils import PreTrainedTokenizer
+from .tokenization_xlnet import SPIECE_UNDERLINE
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {'vocab_file': 'sentencepiece.bpe.model'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+    'camembert-base': "https://s3.amazonaws.com/models.huggingface.co/bert/camembert-base-sentencepiece.bpe.model",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'camembert-base': None,
+}
+
+class CamembertTokenizer(PreTrainedTokenizer):
+    """
+        Adapted from RobertaTokenizer and XLNetTokenizer
+        SentencePiece based tokenizer. Peculiarities:
+
+            - requires `SentencePiece <https://github.com/google/sentencepiece>`_
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, bos_token="<s>", eos_token="</s>", sep_token="</s>",
+                 cls_token="<s>", unk_token="<unk>", pad_token='<pad>', mask_token='<mask>',
+                 additional_special_tokens=['<s>NOTUSED', '</s>NOTUSED'], **kwargs):
+        super(CamembertTokenizer, self).__init__(max_len=512, bos_token=bos_token, eos_token=eos_token, unk_token=unk_token,
+                                                 sep_token=sep_token, cls_token=cls_token, pad_token=pad_token,
+                                                 mask_token=mask_token, additional_special_tokens=additional_special_tokens,
+                                                 **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 4  # take into account special tokens
+        self.sp_model = spm.SentencePieceProcessor()
+        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
+        # sentencepiece vocabulary (this is the case for <s> and </s>
+        self.fairseq_tokens_to_ids = {'<s>NOTUSED': 0, '<pad>': 1, '</s>NOTUSED': 2, '<unk>': 3}
+        self.fairseq_offset = len(self.fairseq_tokens_to_ids)
+        self.fairseq_tokens_to_ids['<mask>'] = len(self.sp_model) + len(self.fairseq_tokens_to_ids)
+        self.fairseq_ids_to_tokens = {v: k for k, v in self.fairseq_tokens_to_ids.items()}
+
+    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 RoBERTa sequence has the following format:
+            single sequence: <s> X </s>
+            pair of sequences: <s> A </s></s> B </s>
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` or ``encode_plus`` methods.
+
+        Args:
+            token_ids_0: list of ids (must not contain special tokens)
+            token_ids_1: Optional list of ids (must not contain special tokens), necessary when fetching sequence ids
+                for sequence pairs
+            already_has_special_tokens: (default False) Set to True if the token list is already formated with
+                special tokens for the model
+
+        Returns:
+            A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError("You should not supply a second sequence if the provided sequence of "
+                                 "ids is already formated with special tokens for the model.")
+            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        A RoBERTa sequence pair mask has the following format:
+        0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence
+
+        if token_ids_1 is None, only returns the first portion of the mask (0'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 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    @property
+    def vocab_size(self):
+        return len(self.fairseq_tokens_to_ids) + len(self.sp_model)
+
+    def _tokenize(self, text):
+        return self.sp_model.EncodeAsPieces(text)
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        if token in self.fairseq_tokens_to_ids:
+            return self.fairseq_tokens_to_ids[token]
+        elif self.sp_model.PieceToId(token) == 0:
+            # Convert sentence piece unk token to fairseq unk token index
+            return self.unk_token_id
+        return self.fairseq_offset + self.sp_model.PieceToId(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        if index in self.fairseq_ids_to_tokens:
+            return self.fairseq_ids_to_tokens[index]
+        return self.sp_model.IdToPiece(index - self.fairseq_offset)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (strings for sub-words) in a single string."""
+        out_string = ''.join(tokens).replace(SPIECE_UNDERLINE, ' ').strip()
+        return out_string
+
+    def save_vocabulary(self, save_directory):
+        """ Save the sentencepiece vocabulary (copy original file) and special tokens file
+            to a directory.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        out_vocab_file = os.path.join(save_directory, VOCAB_FILES_NAMES['vocab_file'])
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)

transformers/tokenization_ctrl.py

@@ -133,9 +133,11 @@ class CTRLTokenizer(PreTrainedTokenizer):
         self.max_len_single_sentence = self.max_len # no default special tokens - you can update this value if you add special tokens
         self.max_len_sentences_pair = self.max_len # no default special tokens - you can update this value if you add special tokens
 
-        self.encoder = json.load(open(vocab_file, encoding="utf-8"))
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
         self.decoder = {v:k for k,v in self.encoder.items()}
-        merges = open(merges_file, encoding='utf-8').read().split('\n')[1:-1]
+        with open(merges_file, encoding='utf-8') as merges_handle:
+            merges = merges_handle.read().split('\n')[1:-1]
         merges = [tuple(merge.split()) for merge in merges]
         self.bpe_ranks = dict(zip(merges, range(len(merges))))
         self.cache = {}
@@ -192,9 +194,9 @@ class CTRLTokenizer(PreTrainedTokenizer):
         """
         split_tokens = []
 
-        text = text.split(' ')
+        words = re.findall(r'\S+\n?', text)
 
-        for token in text:
+        for token in words:
             split_tokens.extend([t for t in self.bpe(token).split(' ')])
         return split_tokens
 

transformers/tokenization_distilbert.py

@@ -33,12 +33,16 @@ PRETRAINED_VOCAB_FILES_MAP = {
     {
         'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
         'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt",
+        'distilbert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-german-cased-vocab.txt",
+        'distilbert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-vocab.txt",
     }
 }
 
 PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     'distilbert-base-uncased': 512,
     'distilbert-base-uncased-distilled-squad': 512,
+    'distilbert-base-german-cased': 512,
+    'distilbert-base-multilingual-cased': 512,
 }
 
 
@@ -49,12 +53,12 @@ class DistilBertTokenizer(BertTokenizer):
 
     Args:
         vocab_file: Path to a one-wordpiece-per-line vocabulary file
-        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
         do_basic_tokenize: Whether to do basic tokenization before wordpiece.
         max_len: An artificial maximum length to truncate tokenized sequences to; Effective maximum length is always the
             minimum of this value (if specified) and the underlying BERT model's sequence length.
         never_split: List of tokens which will never be split during tokenization. Only has an effect when
-            do_wordpiece_only=False
+            do_basic_tokenize=True
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

transformers/tokenization_gpt2.py

@@ -72,7 +72,7 @@ def bytes_to_unicode():
     """
     Returns list of utf-8 byte and a mapping to unicode strings.
     We specifically avoids mapping to whitespace/control characters the bpe code barfs on.
-    
+
     The reversible bpe codes work on unicode strings.
     This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
     When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
@@ -107,10 +107,10 @@ class GPT2Tokenizer(PreTrainedTokenizer):
     """
     GPT-2 BPE tokenizer. Peculiarities:
         - Byte-level Byte-Pair-Encoding
-        - Requires a space to start the input string => the encoding methods should be called with the
+        - Requires a space to start the input string => the encoding and tokenize methods should be called with the
           ``add_prefix_space`` flag set to ``True``.
-          Otherwise, this tokenizer ``encode`` and ``decode`` method will not conserve
-          the absence of a space at the beginning of a string: `tokenizer.decode(tokenizer.encode("Hello")) = " Hello"`
+          Otherwise, this tokenizer's ``encode``, ``decode``, and ``tokenize`` methods will not conserve
+          the spaces at the beginning of a string: `tokenizer.decode(tokenizer.encode(" Hello")) = "Hello"`
     """
     vocab_files_names = VOCAB_FILES_NAMES
     pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
@@ -122,13 +122,15 @@ class GPT2Tokenizer(PreTrainedTokenizer):
         self.max_len_single_sentence = self.max_len # no default special tokens - you can update this value if you add special tokens
         self.max_len_sentences_pair = self.max_len # no default special tokens - you can update this value if you add special tokens
 
-        self.encoder = json.load(open(vocab_file, encoding="utf-8"))
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
         self.decoder = {v: k for k, v in self.encoder.items()}
         self.errors = errors  # how to handle errors in decoding
         self.byte_encoder = bytes_to_unicode()
         self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
-        bpe_data = open(merges_file, encoding='utf-8').read().split('\n')[1:-1]
-        bpe_merges = [tuple(merge.split()) for merge in bpe_data]
+        with open(merges_file, encoding='utf-8') as merges_handle:
+            bpe_merges = merges_handle.read().split('\n')[1:-1]
+        bpe_merges = [tuple(merge.split()) for merge in bpe_merges]
         self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
         self.cache = {}
 
@@ -184,7 +186,7 @@ class GPT2Tokenizer(PreTrainedTokenizer):
         """ Tokenize a string.
             Args:
                 - add_prefix_space (boolean, default False):
-                    Begin the sentence with at least one space toto get invariance to word order in GPT-2 (and RoBERTa) tokenizers.
+                    Begin the sentence with at least one space to get invariance to word order in GPT-2 (and RoBERTa) tokenizers.
         """
         if add_prefix_space:
             text = ' ' + text
@@ -234,4 +236,4 @@ class GPT2Tokenizer(PreTrainedTokenizer):
                 writer.write(' '.join(bpe_tokens) + u'\n')
                 index += 1
 
-        return vocab_file, merge_file
+        return vocab_file, merge_file
\ No newline at end of file

transformers/tokenization_openai.py

@@ -101,9 +101,11 @@ class OpenAIGPTTokenizer(PreTrainedTokenizer):
             self.nlp = BasicTokenizer(do_lower_case=True)
             self.fix_text = None
 
-        self.encoder = json.load(open(vocab_file, encoding="utf-8"))
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
         self.decoder = {v:k for k,v in self.encoder.items()}
-        merges = open(merges_file, encoding='utf-8').read().split('\n')[1:-1]
+        with open(merges_file, encoding='utf-8') as merges_handle:
+            merges = merges_handle.read().split('\n')[1:-1]
         merges = [tuple(merge.split()) for merge in merges]
         self.bpe_ranks = dict(zip(merges, range(len(merges))))
         self.cache = {}

transformers/tokenization_t5.py

+# coding=utf-8
+# Copyright 2018 T5 Authors and 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.
+""" Tokenization class for model T5."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import logging
+import os
+import re
+import six
+from shutil import copyfile
+
+from .tokenization_utils import PreTrainedTokenizer
+
+logger = logging.getLogger(__name__)
+
+SPIECE_UNDERLINE = u'▁'
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to file names for serializing Tokenizer instances
+####################################################
+VOCAB_FILES_NAMES = {'vocab_file': 'spiece.model'}
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to pretrained vocabulary URL for all the model shortcut names.
+####################################################
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+    }
+}
+
+####################################################
+# Mapping from model shortcut names to max length of inputs
+####################################################
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    't5-small': 512,
+    't5-base': 512,
+    't5-large': 512,
+    't5-3b': 512,
+    't5-11b': 512,
+}
+
+class T5Tokenizer(PreTrainedTokenizer):
+    """
+        SentencePiece based tokenizer. Peculiarities:
+
+            - requires `SentencePiece <https://github.com/google/sentencepiece>`_
+            - `extra_ids` add a number of extra ids added to the end of the vocabulary for use as sentinels.
+                These tokens are accessible as `<extra_id_{%d}>` where `{%d}` is a number between 0 and extra_ids-1.
+                Extra tokens are indexed from the end of the vocabulary up to beginnning (<extra_id_0> is the last token in the vocabulary)
+                (like in T5 preprocessing
+                see: https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117)
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, eos_token="</s>", unk_token="<unk>",
+                 pad_token="<pad>", extra_ids=100, additional_special_tokens=None, **kwargs):
+        # Add extra_ids to the special token list
+        if extra_ids > 0:
+            if additional_special_tokens is None:
+                additional_special_tokens = []
+            additional_special_tokens.extend([u"<extra_id_{}>".format(i) for i in range(extra_ids)])
+
+        super(T5Tokenizer, self).__init__(eos_token=eos_token, unk_token=unk_token,
+                                          pad_token=pad_token, additional_special_tokens=additional_special_tokens,
+                                          **kwargs)
+
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning("You need to install SentencePiece to use T5Tokenizer:"
+                           "https://github.com/google/sentencepiece"
+                           "pip install sentencepiece")
+
+        self.vocab_file = vocab_file
+        self._extra_ids = extra_ids
+
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(vocab_file)
+
+    @property
+    def vocab_size(self):
+        return self.sp_model.get_piece_size() + self._extra_ids
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning("You need to install SentencePiece to use XLNetTokenizer: https://github.com/google/sentencepiece"
+                           "pip install sentencepiece")
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(self.vocab_file)
+
+    def _tokenize(self, text, return_unicode=True, sample=False):
+        """ 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)
+
+        # convert back to unicode for py2
+        if six.PY2 and return_unicode:
+            ret_pieces = []
+            for piece in pieces:
+                if isinstance(piece, str):
+                    piece = piece.decode('utf-8')
+                ret_pieces.append(piece)
+            pieces = ret_pieces
+
+        return pieces
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        if token.startswith(u"<extra_id_"):
+            l = re.match(r'<extra_id_(\d+)>', token)
+            num = int(l.group(1))
+            return self.vocab_size - num - 1
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index, return_unicode=True):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        if index < self.sp_model.get_piece_size():
+            token = self.sp_model.IdToPiece(index)
+        else:
+            token = u"<extra_id_{}>".format(self.vocab_size - 1 - index)
+        if six.PY2 and return_unicode and isinstance(token, str):
+            token = token.decode('utf-8')
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """ Converts a sequence of tokens (string) in a single string. """
+        out_string = self.sp_model.decode_pieces(tokens)
+        return out_string
+
+    def save_vocabulary(self, save_directory):
+        """ Save the sentencepiece vocabulary (copy original file) and special tokens file
+            to a directory.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        out_vocab_file = os.path.join(save_directory, VOCAB_FILES_NAMES['vocab_file'])
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)

transformers/tokenization_utils.py

@@ -22,9 +22,10 @@ import json
 import six
 import copy
 import itertools
+import re
 from io import open
 
-from .file_utils import cached_path, is_tf_available, is_torch_available
+from .file_utils import cached_path, is_remote_url, hf_bucket_url, is_tf_available, is_torch_available
 
 if is_tf_available():
     import tensorflow as tf
@@ -77,6 +78,8 @@ class PreTrainedTokenizer(object):
                                  "pad_token", "cls_token", "mask_token",
                                  "additional_special_tokens"]
 
+    padding_side = "right"
+
     @property
     def bos_token(self):
         """ Beginning of sentence token (string). Log an error if used while not having been set. """
@@ -190,6 +193,11 @@ class PreTrainedTokenizer(object):
         """ Id of the padding token in the vocabulary. Log an error if used while not having been set. """
         return self.convert_tokens_to_ids(self.pad_token)
 
+    @property
+    def pad_token_type_id(self):
+        """ Id of the padding token type in the vocabulary."""
+        return self._pad_token_type_id
+
     @property
     def cls_token_id(self):
         """ Id of the classification token in the vocabulary. E.g. to extract a summary of an input sequence leveraging self-attention along the full depth of the model. Log an error if used while not having been set. """
@@ -213,10 +221,14 @@ class PreTrainedTokenizer(object):
         self._pad_token = None
         self._cls_token = None
         self._mask_token = None
+        self._pad_token_type_id = 0
         self._additional_special_tokens = []
 
         self.max_len = max_len if max_len is not None else int(1e12)
 
+        # Padding side is right by default and over-riden in subclasses. If specified in the kwargs, it is changed.
+        self.padding_side = kwargs.pop('padding_side', self.padding_side)
+        
         # Added tokens
         self.added_tokens_encoder = {}
         self.added_tokens_decoder = {}
@@ -243,6 +255,7 @@ class PreTrainedTokenizer(object):
             pretrained_model_name_or_path: either:
 
                 - a string with the `shortcut name` of a predefined tokenizer to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a predefined tokenizer that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing vocabulary files required by the tokenizer, for instance saved using the :func:`~transformers.PreTrainedTokenizer.save_pretrained` method, e.g.: ``./my_model_directory/``.
                 - (not applicable to all derived classes) a path or url to a single saved vocabulary file if and only if the tokenizer only requires a single vocabulary file (e.g. Bert, XLNet), e.g.: ``./my_model_directory/vocab.txt``.
 
@@ -252,6 +265,9 @@ class PreTrainedTokenizer(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the vocabulary files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
             proxies: (`optional`) dict, default None:
                 A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
                 The proxies are used on each request.
@@ -267,6 +283,9 @@ class PreTrainedTokenizer(object):
             # Download vocabulary from S3 and cache.
             tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
 
+            # Download vocabulary from S3 (user-uploaded) and cache.
+            tokenizer = BertTokenizer.from_pretrained('dbmdz/bert-base-german-cased')
+
             # If vocabulary files are in a directory (e.g. tokenizer was saved using `save_pretrained('./test/saved_model/')`)
             tokenizer = BertTokenizer.from_pretrained('./test/saved_model/')
 
@@ -287,6 +306,7 @@ class PreTrainedTokenizer(object):
     def _from_pretrained(cls, pretrained_model_name_or_path, *init_inputs, **kwargs):
         cache_dir = kwargs.pop('cache_dir', None)
         force_download = kwargs.pop('force_download', False)
+        resume_download = kwargs.pop('resume_download', False)
         proxies = kwargs.pop('proxies', None)
 
         s3_models = list(cls.max_model_input_sizes.keys())
@@ -311,12 +331,15 @@ class PreTrainedTokenizer(object):
                 if os.path.isdir(pretrained_model_name_or_path):
                     # If a directory is provided we look for the standard filenames
                     full_file_name = os.path.join(pretrained_model_name_or_path, file_name)
-                else:
+                    if not os.path.exists(full_file_name):
+                        logger.info("Didn't find file {}. We won't load it.".format(full_file_name))
+                        full_file_name = None
+                elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
                     # If a path to a file is provided we use it (will only work for non-BPE tokenizer using a single vocabulary file)
                     full_file_name = pretrained_model_name_or_path
-                if not os.path.exists(full_file_name):
-                    logger.info("Didn't find file {}. We won't load it.".format(full_file_name))
-                    full_file_name = None
+                else:
+                    full_file_name = hf_bucket_url(pretrained_model_name_or_path, postfix=file_name)
+                
                 vocab_files[file_id] = full_file_name
 
             # Look for the additional tokens files
@@ -343,7 +366,7 @@ class PreTrainedTokenizer(object):
                     "We assumed '{}' was a path or url to a directory containing vocabulary files "
                     "named {} but couldn't find such vocabulary files at this path or url.".format(
                         pretrained_model_name_or_path, ', '.join(s3_models),
-                        pretrained_model_name_or_path, 
+                        pretrained_model_name_or_path,
                         list(cls.vocab_files_names.values())))
 
         # Get files from url, cache, or disk depending on the case
@@ -353,7 +376,7 @@ class PreTrainedTokenizer(object):
                 if file_path is None:
                     resolved_vocab_files[file_id] = None
                 else:
-                    resolved_vocab_files[file_id] = cached_path(file_path, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+                    resolved_vocab_files[file_id] = cached_path(file_path, cache_dir=cache_dir, force_download=force_download, proxies=proxies, resume_download=resume_download)
         except EnvironmentError:
             if pretrained_model_name_or_path in s3_models:
                 msg = "Couldn't reach server at '{}' to download vocabulary files."
@@ -378,7 +401,8 @@ class PreTrainedTokenizer(object):
         # Did we saved some inputs and kwargs to reload ?
         tokenizer_config_file = resolved_vocab_files.pop('tokenizer_config_file', None)
         if tokenizer_config_file is not None:
-            init_kwargs = json.load(open(tokenizer_config_file, encoding="utf-8"))
+            with open(tokenizer_config_file, encoding="utf-8") as tokenizer_config_handle:
+                init_kwargs = json.load(tokenizer_config_handle)
             saved_init_inputs = init_kwargs.pop('init_inputs', ())
             if not init_inputs:
                 init_inputs = saved_init_inputs
@@ -403,13 +427,18 @@ class PreTrainedTokenizer(object):
             if args_name not in init_kwargs:
                 init_kwargs[args_name] = file_path
         if special_tokens_map_file is not None:
-            special_tokens_map = json.load(open(special_tokens_map_file, encoding="utf-8"))
+            with open(special_tokens_map_file, encoding="utf-8") as special_tokens_map_handle:
+                special_tokens_map = json.load(special_tokens_map_handle)
             for key, value in special_tokens_map.items():
                 if key not in init_kwargs:
                     init_kwargs[key] = value
 
         # Instantiate tokenizer.
-        tokenizer = cls(*init_inputs, **init_kwargs)
+        try:
+            tokenizer = cls(*init_inputs, **init_kwargs)
+        except OSError:
+            OSError("Unable to load vocabulary from file. "
+                    "Please check that the provided vocabulary is accessible and not corrupted.")
 
         # Save inputs and kwargs for saving and re-loading with ``save_pretrained``
         tokenizer.init_inputs = init_inputs
@@ -417,7 +446,8 @@ class PreTrainedTokenizer(object):
 
         # Add supplementary tokens.
         if added_tokens_file is not None:
-            added_tok_encoder = json.load(open(added_tokens_file, encoding="utf-8"))
+            with open(added_tokens_file, encoding="utf-8") as added_tokens_handle:
+                added_tok_encoder = json.load(added_tokens_handle)
             added_tok_decoder = {v:k for k, v in added_tok_encoder.items()}
             tokenizer.added_tokens_encoder.update(added_tok_encoder)
             tokenizer.added_tokens_decoder.update(added_tok_decoder)
@@ -513,6 +543,8 @@ class PreTrainedTokenizer(object):
         to_add_tokens = []
         for token in new_tokens:
             assert isinstance(token, str) or (six.PY2 and isinstance(token, unicode))
+            if self.init_kwargs.get('do_lower_case', False) and token not in self.all_special_tokens:
+                token = token.lower()
             if token != self.unk_token and \
                     self.convert_tokens_to_ids(token) == self.convert_tokens_to_ids(self.unk_token) and \
                     token not in to_add_tokens:
@@ -605,7 +637,25 @@ class PreTrainedTokenizer(object):
             vocabularies (BPE/SentencePieces/WordPieces).
 
             Take care of added tokens.
+
+            text: The sequence to be encoded.
+            **kwargs: passed to the child `self.tokenize()` method
         """
+        all_special_tokens = self.all_special_tokens
+
+        def lowercase_text(t):
+            # convert non-special tokens to lowercase
+            escaped_special_toks = [re.escape(s_tok) for s_tok in all_special_tokens]
+            pattern = r'(' + r'|'.join(escaped_special_toks) + r')|' + \
+                      r'(.+?)'
+            return re.sub(
+                pattern,
+                lambda m: m.groups()[0] or m.groups()[1].lower(),
+                t)
+
+        if self.init_kwargs.get('do_lower_case', False):
+            text = lowercase_text(text)
+
         def split_on_token(tok, text):
             result = []
             split_text = text.split(tok)
@@ -625,7 +675,7 @@ class PreTrainedTokenizer(object):
             return result
 
         def split_on_tokens(tok_list, text):
-            if not text:
+            if not text.strip():
                 return []
             if not tok_list:
                 return self._tokenize(text, **kwargs)
@@ -636,17 +686,17 @@ class PreTrainedTokenizer(object):
                 tokenized_text = []
                 for sub_text in text_list:
                     if sub_text not in self.added_tokens_encoder \
-                            and sub_text not in self.all_special_tokens:
+                            and sub_text not in all_special_tokens:
                         tokenized_text += split_on_token(tok, sub_text)
                     else:
                         tokenized_text += [sub_text]
                 text_list = tokenized_text
 
             return list(itertools.chain.from_iterable((self._tokenize(token, **kwargs) if token not \
-                    in self.added_tokens_encoder and token not in self.all_special_tokens \
+                    in self.added_tokens_encoder and token not in all_special_tokens \
                     else [token] for token in tokenized_text)))
 
-        added_tokens = list(self.added_tokens_encoder.keys()) + self.all_special_tokens
+        added_tokens = list(self.added_tokens_encoder.keys()) + all_special_tokens
         tokenized_text = split_on_tokens(added_tokens, text)
         return tokenized_text
 
@@ -692,6 +742,7 @@ class PreTrainedTokenizer(object):
                max_length=None,
                stride=0,
                truncation_strategy='longest_first',
+               pad_to_max_length=False,
                return_tensors=None,
                **kwargs):
         """
@@ -718,6 +769,12 @@ class PreTrainedTokenizer(object):
                 - 'only_first': Only truncate the first sequence
                 - 'only_second': Only truncate the second sequence
                 - 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
+            pad_to_max_length: if set to True, the returned sequences will be padded according to the model's padding side and
+                padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
+                The tokenizer padding sides are handled by the following strings:
+                - 'left': pads on the left of the sequences
+                - 'right': pads on the right of the sequences   
+                Defaults to False: no padding.
             return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
                 or PyTorch torch.Tensor instead of a list of python integers.
             **kwargs: passed to the `self.tokenize()` method
@@ -728,6 +785,7 @@ class PreTrainedTokenizer(object):
                                           add_special_tokens=add_special_tokens,
                                           stride=stride,
                                           truncation_strategy=truncation_strategy,
+                                          pad_to_max_length=pad_to_max_length,
                                           return_tensors=return_tensors,
                                           **kwargs)
 
@@ -740,7 +798,12 @@ class PreTrainedTokenizer(object):
                     max_length=None,
                     stride=0,
                     truncation_strategy='longest_first',
+                    pad_to_max_length=False,
                     return_tensors=None,
+                    return_token_type_ids=True,
+                    return_attention_mask=True,
+                    return_overflowing_tokens=False,
+                    return_special_tokens_mask=False,
                     **kwargs):
         """
         Returns a dictionary containing the encoded sequence or sequence pair and additional informations:
@@ -765,9 +828,40 @@ class PreTrainedTokenizer(object):
                 - 'only_first': Only truncate the first sequence
                 - 'only_second': Only truncate the second sequence
                 - 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
+            pad_to_max_length: if set to True, the returned sequences will be padded according to the model's padding side and
+                padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
+                The tokenizer padding sides are handled by the following strings:
+                - 'left': pads on the left of the sequences
+                - 'right': pads on the right of the sequences   
+                Defaults to False: no padding.
             return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
                 or PyTorch torch.Tensor instead of a list of python integers.
+            return_token_type_ids: (optional) Set to False to avoid returning token_type_ids (default True).
+            return_attention_mask: (optional) Set to False to avoir returning attention mask (default True)
+            return_overflowing_tokens: (optional) Set to True to return overflowing token information (default False).
+            return_special_tokens_mask: (optional) Set to True to return special tokens mask information (default False).
             **kwargs: passed to the `self.tokenize()` method
+
+        Return:
+            A Dictionary of shape::
+
+                {
+                    input_ids: list[int],
+                    token_type_ids: list[int] if return_token_type_ids is True (default)
+                    attention_mask: list[int] if return_attention_mask is True (default)
+                    overflowing_tokens: list[int] if a ``max_length`` is specified and return_overflowing_tokens is True
+                    num_truncated_tokens: int if a ``max_length`` is specified and return_overflowing_tokens is True
+                    special_tokens_mask: list[int] if ``add_special_tokens`` if set to ``True`` and return_special_tokens_mask is True
+                }
+
+            With the fields:
+                ``input_ids``: list of token ids to be fed to a model
+                ``token_type_ids``: list of token type ids to be fed to a model
+                ``attention_mask``: list of indices specifying which tokens should be attended to by the model
+                ``overflowing_tokens``: list of overflowing tokens if a max length is specified.
+                ``num_truncated_tokens``: number of overflowing tokens a ``max_length`` is specified
+                ``special_tokens_mask``: if adding special tokens, this is a list of [0, 1], with 0 specifying special added
+                tokens and 1 specifying sequence tokens.
         """
 
         def get_input_ids(text):
@@ -786,13 +880,110 @@ class PreTrainedTokenizer(object):
         return self.prepare_for_model(first_ids,
                                       pair_ids=second_ids,
                                       max_length=max_length,
+                                      pad_to_max_length=pad_to_max_length,
                                       add_special_tokens=add_special_tokens,
                                       stride=stride,
                                       truncation_strategy=truncation_strategy,
-                                      return_tensors=return_tensors)
+                                      return_tensors=return_tensors,
+                                      return_attention_mask=return_attention_mask,
+                                      return_token_type_ids=return_token_type_ids,
+                                      return_overflowing_tokens=return_overflowing_tokens,
+                                      return_special_tokens_mask=return_special_tokens_mask)
+
+    def batch_encode_plus(self,
+                          batch_text_or_text_pairs=None,
+                          add_special_tokens=False,
+                          max_length=None,
+                          stride=0,
+                          truncation_strategy='longest_first',
+                          return_tensors=None,
+                          return_input_lengths=False,
+                          return_attention_masks=False,
+                          **kwargs):
+        """
+        Returns a dictionary containing the encoded sequence or sequence pair and additional information:
+        the mask for sequence classification and the overflowing elements if a ``max_length`` is specified.
+
+        Args:
+            batch_text_or_text_pairs: Batch of sequences or pair of sequences to be encoded.
+                This can be a list of string/string-sequences/int-sequences or a list of pair of
+                string/string-sequences/int-sequence (see details in encode_plus)
+            add_special_tokens: if set to ``True``, the sequences will be encoded with the special tokens relative
+                to their model.
+            max_length: if set to a number, will limit the total sequence returned so that it has a maximum length.
+                If there are overflowing tokens, those will be added to the returned dictionary`
+            stride: if set to a number along with max_length, the overflowing tokens returned will contain some tokens
+                from the main sequence returned. The value of this argument defines the number of additional tokens.
+            truncation_strategy: string selected in the following options:
+                - 'longest_first' (default) Iteratively reduce the inputs sequence until the input is under max_length
+                    starting from the longest one at each token (when there is a pair of input sequences)
+                - 'only_first': Only truncate the first sequence
+                - 'only_second': Only truncate the second sequence
+                - 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
+            return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
+                or PyTorch torch.Tensor instead of a list of python integers.
+            **kwargs: passed to the `self.tokenize()` method
+        """
+        batch_outputs = {}
+        for ids_or_pair_ids in batch_text_or_text_pairs:
+            if isinstance(ids_or_pair_ids, (list, tuple)):
+                assert len(ids_or_pair_ids) == 2
+                ids, pair_ids = ids_or_pair_ids
+            else:
+                ids, pair_ids = ids_or_pair_ids, None
+            outputs = self.encode_plus(ids, pair_ids, add_special_tokens=add_special_tokens, max_length=max_length,
+                                       stride=stride, truncation_strategy=truncation_strategy, return_tensors=None)
+
+            # Append the non-padded length to the output
+            if return_input_lengths:
+                outputs['input_len'] = len(outputs['input_ids'])
+
+            for key, value in outputs.items():
+                if key not in batch_outputs:
+                    batch_outputs[key] = []
+                batch_outputs[key].append(value)
+
+        # Compute longest sequence size
+        max_seq_len = max(map(len, batch_outputs['input_ids']))
+
+        if return_attention_masks:
+            # Allow the model to not give any special attention to padded input
+            batch_outputs['attention_mask'] = [[0] * len(v) for v in batch_outputs['input_ids']]
+
+        if return_tensors is not None:
+
+            # Do the tensor conversion in batch
+            for key, value in batch_outputs.items():
+
+                padded_value = value
+                if key != 'input_len':
+                    # Padding handle
+                    padded_value = [v + [self.pad_token_id if key == 'input_ids' else 1] * (max_seq_len - len(v)) for v in padded_value]
+
+                if return_tensors == 'tf' and is_tf_available():
+                    batch_outputs[key] = tf.constant(padded_value)
+                elif return_tensors == 'pt' and is_torch_available():
+                    batch_outputs[key] = torch.tensor(padded_value)
+                elif return_tensors is not None:
+                    logger.warning("Unable to convert output to tensors format {}, PyTorch or TensorFlow is not available.".format(return_tensors))
+
+        # encoder_attention_mask requires 1 for real token, 0 for padding, just invert value
+        if return_attention_masks:
+            if is_tf_available():
+                batch_outputs['attention_mask'] = tf.abs(batch_outputs['attention_mask'] - 1)
+            else:
+                batch_outputs['attention_mask'] = torch.abs(batch_outputs['attention_mask'] - 1)
+
+        return batch_outputs
 
     def prepare_for_model(self, ids, pair_ids=None, max_length=None, add_special_tokens=True, stride=0,
-                          truncation_strategy='longest_first', return_tensors=None):
+                          truncation_strategy='longest_first',
+                          pad_to_max_length=False,
+                          return_tensors=None,
+                          return_token_type_ids=True,
+                          return_attention_mask=True,
+                          return_overflowing_tokens=False,
+                          return_special_tokens_mask=False):
         """
         Prepares a sequence of input id, or a pair of sequences of inputs ids so that it can be used by the model.
         It adds special tokens, truncates
@@ -815,23 +1006,36 @@ class PreTrainedTokenizer(object):
                 - 'only_first': Only truncate the first sequence
                 - 'only_second': Only truncate the second sequence
                 - 'do_not_truncate': Does not truncate (raise an error if the input sequence is longer than max_length)
+            pad_to_max_length: if set to True, the returned sequences will be padded according to the model's padding side and
+                padding index, up to their max length. If no max length is specified, the padding is done up to the model's max length.
+                The tokenizer padding sides are handled by the following strings:
+                - 'left': pads on the left of the sequences
+                - 'right': pads on the right of the sequences   
+                Defaults to False: no padding.
             return_tensors: (optional) can be set to 'tf' or 'pt' to return respectively TensorFlow tf.constant
                 or PyTorch torch.Tensor instead of a list of python integers.
+            return_token_type_ids: (optional) Set to False to avoid returning token_type_ids (default True).
+            return_attention_mask: (optional) Set to False to avoid returning attention mask (default True)
+            return_overflowing_tokens: (optional) Set to True to return overflowing token information (default False).
+            return_special_tokens_mask: (optional) Set to True to return special tokens mask information (default False).
 
         Return:
             A Dictionary of shape::
 
                 {
                     input_ids: list[int],
-                    overflowing_tokens: list[int] if a ``max_length`` is specified, else None
-                    special_tokens_mask: list[int] if ``add_special_tokens`` if set to ``True``
+                    token_type_ids: list[int] if return_token_type_ids is True (default)
+                    overflowing_tokens: list[int] if a ``max_length`` is specified and return_overflowing_tokens is True
+                    num_truncated_tokens: int if a ``max_length`` is specified and return_overflowing_tokens is True
+                    special_tokens_mask: list[int] if ``add_special_tokens`` if set to ``True`` and return_special_tokens_mask is True
                 }
 
             With the fields:
-                ``input_ids``: list of tokens to be fed to a model
+                ``input_ids``: list of token ids to be fed to a model
+                ``token_type_ids``: list of token type ids to be fed to a model
 
                 ``overflowing_tokens``: list of overflowing tokens if a max length is specified.
-
+                ``num_truncated_tokens``: number of overflowing tokens a ``max_length`` is specified
                 ``special_tokens_mask``: if adding special tokens, this is a list of [0, 1], with 0 specifying special added
                 tokens and 1 specifying sequence tokens.
         """
@@ -840,45 +1044,99 @@ class PreTrainedTokenizer(object):
         len_pair_ids = len(pair_ids) if pair else 0
 
         encoded_inputs = {}
+
+        # Handle max sequence length
         total_len = len_ids + len_pair_ids + (self.num_added_tokens(pair=pair) if add_special_tokens else 0)
         if max_length and total_len > max_length:
             ids, pair_ids, overflowing_tokens = self.truncate_sequences(ids, pair_ids=pair_ids,
                                                                         num_tokens_to_remove=total_len-max_length,
                                                                         truncation_strategy=truncation_strategy,
                                                                         stride=stride)
-            encoded_inputs["overflowing_tokens"] = overflowing_tokens
-            encoded_inputs["num_truncated_tokens"] = total_len - max_length
+            if return_overflowing_tokens:
+                encoded_inputs["overflowing_tokens"] = overflowing_tokens
+                encoded_inputs["num_truncated_tokens"] = total_len - max_length
 
+        # Handle special_tokens
         if add_special_tokens:
             sequence = self.build_inputs_with_special_tokens(ids, pair_ids)
             token_type_ids = self.create_token_type_ids_from_sequences(ids, pair_ids)
-            encoded_inputs["special_tokens_mask"] = self.get_special_tokens_mask(ids, pair_ids)
         else:
             sequence = ids + pair_ids if pair else ids
             token_type_ids = [0] * len(ids) + ([1] * len(pair_ids) if pair else [])
 
-        if return_tensors == 'tf' and is_tf_available():
-            sequence = tf.constant([sequence])
-            token_type_ids = tf.constant([token_type_ids])
-        elif return_tensors == 'pt' and is_torch_available():
-            sequence = torch.tensor([sequence])
-            token_type_ids = torch.tensor([token_type_ids])
-        elif return_tensors is not None:
-            logger.warning("Unable to convert output to tensors format {}, PyTorch or TensorFlow is not available.".format(return_tensors))
+        if return_special_tokens_mask:
+            encoded_inputs["special_tokens_mask"] = self.get_special_tokens_mask(ids, pair_ids)
 
         encoded_inputs["input_ids"] = sequence
-        encoded_inputs["token_type_ids"] = token_type_ids
+        if return_token_type_ids:
+            encoded_inputs["token_type_ids"] = token_type_ids
 
         if max_length and len(encoded_inputs["input_ids"]) > max_length:
             encoded_inputs["input_ids"] = encoded_inputs["input_ids"][:max_length]
-            encoded_inputs["token_type_ids"] = encoded_inputs["token_type_ids"][:max_length]
-            encoded_inputs["special_tokens_mask"] = encoded_inputs["special_tokens_mask"][:max_length]
+            if return_token_type_ids:
+                encoded_inputs["token_type_ids"] = encoded_inputs["token_type_ids"][:max_length]
+            if return_special_tokens_mask:
+                encoded_inputs["special_tokens_mask"] = encoded_inputs["special_tokens_mask"][:max_length]
 
         if max_length is None and len(encoded_inputs["input_ids"]) > self.max_len:
             logger.warning("Token indices sequence length is longer than the specified maximum sequence length "
                            "for this model ({} > {}). Running this sequence through the model will result in "
                            "indexing errors".format(len(ids), self.max_len))
                            
+        needs_to_be_padded = pad_to_max_length and (
+            max_length and len(encoded_inputs["input_ids"]) < max_length
+            or 
+            max_length is None and len(encoded_inputs["input_ids"]) < self.max_len and self.max_len <= 10000
+        )
+
+        if pad_to_max_length and max_length is None and self.max_len > 10000:
+            logger.warning("Sequence can't be padded as no maximum length is specified and the model maximum length is too high.")
+
+        if needs_to_be_padded:
+            difference = (max_length if max_length is not None else self.max_len) - len(encoded_inputs["input_ids"])
+
+            if self.padding_side == 'right':
+                if return_attention_mask:
+                    encoded_inputs["attention_mask"] = [1] * len(encoded_inputs["input_ids"]) + [0] * difference
+                if return_token_type_ids:
+                    encoded_inputs["token_type_ids"] = encoded_inputs["token_type_ids"] + [self.pad_token_type_id] * difference
+                if return_special_tokens_mask:
+                    encoded_inputs["special_tokens_mask"] = encoded_inputs["special_tokens_mask"] + [1] * difference
+                encoded_inputs["input_ids"] = encoded_inputs["input_ids"] + [self.pad_token_id] * difference
+            elif self.padding_side == 'left':
+                if return_attention_mask:
+                    encoded_inputs["attention_mask"] = [0] * difference + [1] * len(encoded_inputs["input_ids"])
+                if return_token_type_ids:
+                    encoded_inputs["token_type_ids"] = [self.pad_token_type_id] * difference + encoded_inputs["token_type_ids"]
+                if return_special_tokens_mask:
+                    encoded_inputs["special_tokens_mask"] = [1] * difference + encoded_inputs["special_tokens_mask"]
+                encoded_inputs["input_ids"] = [self.pad_token_id] * difference + encoded_inputs["input_ids"]
+
+            else:
+                raise ValueError("Invalid padding strategy:" + str(self.padding_side))
+            
+        elif return_attention_mask:
+            encoded_inputs["attention_mask"] = [1] * len(encoded_inputs["input_ids"])
+
+        # Prepare inputs as tensors if asked
+        if return_tensors == 'tf' and is_tf_available():
+            encoded_inputs["input_ids"] = tf.constant([encoded_inputs["input_ids"]])
+            encoded_inputs["token_type_ids"] = tf.constant([encoded_inputs["token_type_ids"]])
+
+            if "attention_mask" in encoded_inputs:
+                encoded_inputs["attention_mask"] = tf.constant([encoded_inputs["attention_mask"]])
+
+        elif return_tensors == 'pt' and is_torch_available():
+            encoded_inputs["input_ids"] = torch.tensor([encoded_inputs["input_ids"]])
+            encoded_inputs["token_type_ids"] = torch.tensor([encoded_inputs["token_type_ids"]])
+
+            if "attention_mask" in encoded_inputs:
+                encoded_inputs["attention_mask"] = torch.tensor([encoded_inputs["attention_mask"]])
+        elif return_tensors is not None:
+            logger.warning(
+                "Unable to convert output to tensors format {}, PyTorch or TensorFlow is not available.".format(
+                    return_tensors))
+
         return encoded_inputs
 
     def truncate_sequences(self, ids, pair_ids=None, num_tokens_to_remove=0, truncation_strategy='longest_first', stride=0):
@@ -922,7 +1180,6 @@ class PreTrainedTokenizer(object):
         return (ids, pair_ids, overflowing_tokens)
 
     def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
-        logger.warning("This tokenizer does not make use of special tokens.")
         if token_ids_1 is None:
             return len(token_ids_0) * [0]
         return [0] * len(token_ids_0) + [1] * len(token_ids_1)
@@ -935,7 +1192,6 @@ class PreTrainedTokenizer(object):
             single sequence: <s> X </s>
             pair of sequences: <s> A </s></s> B </s>
         """
-        logger.warning("This tokenizer does not make use of special tokens. Input is returned with no modification.")
         if token_ids_1 is None:
             return token_ids_0
         return token_ids_0 + token_ids_1
@@ -971,6 +1227,7 @@ class PreTrainedTokenizer(object):
                 return self._convert_id_to_token(ids)
         tokens = []
         for index in ids:
+            index = int(index)
             if skip_special_tokens and index in self.all_special_ids:
                 continue
             if index in self.added_tokens_decoder:
@@ -1014,12 +1271,12 @@ class PreTrainedTokenizer(object):
                 if current_sub_text:
                     sub_texts.append(self.convert_tokens_to_string(current_sub_text))
                     current_sub_text = []
-                sub_texts.append(" " + token)
+                sub_texts.append(token)
             else:
                 current_sub_text.append(token)
         if current_sub_text:
             sub_texts.append(self.convert_tokens_to_string(current_sub_text))
-        text = ''.join(sub_texts)
+        text = ' '.join(sub_texts)
 
         if clean_up_tokenization_spaces:
             clean_text = self.clean_up_tokenization(text)

transformers/tokenization_xlm.py

@@ -12,7 +12,7 @@
 # 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.
-"""Tokenization classes for OpenAI GPT."""
+"""Tokenization classes for XLM."""
 from __future__ import (absolute_import, division, print_function,
                         unicode_literals)
 
@@ -524,7 +524,7 @@ class XLMTokenizer(PreTrainedTokenizer):
 
         - argument ``special_tokens`` and function ``set_special_tokens``, can be used to add additional symbols \
         (ex: "__classify__") to a vocabulary
-        
+
         - `lang2id` attribute maps the languages supported by the model with their ids if provided (automatically set for pretrained vocabularies)
 
         - `id2lang` attributes does reverse mapping if provided (automatically set for pretrained vocabularies)
@@ -549,6 +549,10 @@ class XLMTokenizer(PreTrainedTokenizer):
                                            additional_special_tokens=additional_special_tokens,
                                            **kwargs)
 
+
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens 
+
         # cache of sm.MosesPunctNormalizer instance
         self.cache_moses_punct_normalizer = dict()
         # cache of sm.MosesTokenizer instance
@@ -564,9 +568,11 @@ class XLMTokenizer(PreTrainedTokenizer):
         self.ja_word_tokenizer = None
         self.zh_word_tokenizer = None
 
-        self.encoder = json.load(open(vocab_file, encoding="utf-8"))
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
         self.decoder = {v:k for k,v in self.encoder.items()}
-        merges = open(merges_file, encoding='utf-8').read().split('\n')[:-1]
+        with open(merges_file, encoding='utf-8') as merges_handle:
+            merges = merges_handle.read().split('\n')[:-1]
         merges = [tuple(merge.split()[:2]) for merge in merges]
         self.bpe_ranks = dict(zip(merges, range(len(merges))))
         self.cache = {}
@@ -758,9 +764,9 @@ class XLMTokenizer(PreTrainedTokenizer):
         """
         Build model inputs from a sequence or a pair of sequence for sequence classification tasks
         by concatenating and adding special tokens.
-        A RoBERTa sequence has the following format:
+        A XLM sequence has the following format:
             single sequence: <s> X </s>
-            pair of sequences: <s> A </s></s> B </s>
+            pair of sequences: <s> A </s> B </s>
         """
         if token_ids_1 is None:
             return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]

transformers/tokenization_xlm_roberta.py

+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University 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
+""" Tokenization classes for XLM-RoBERTa model."""
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+import os
+from shutil import copyfile
+
+import sentencepiece as spm
+from transformers.tokenization_utils import PreTrainedTokenizer
+from .tokenization_xlnet import SPIECE_UNDERLINE
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {'vocab_file': 'sentencepiece.bpe.model'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+    'xlm-roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-base-sentencepiece.bpe.model",
+    'xlm-roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-sentencepiece.bpe.model",
+    'xlm-roberta-large-finetuned-conll02-dutch': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-dutch-sentencepiece.bpe.model",
+    'xlm-roberta-large-finetuned-conll02-spanish': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll02-spanish-sentencepiece.bpe.model",
+    'xlm-roberta-large-finetuned-conll03-english': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-english-sentencepiece.bpe.model",
+    'xlm-roberta-large-finetuned-conll03-german': "https://s3.amazonaws.com/models.huggingface.co/bert/xlm-roberta-large-finetuned-conll03-german-sentencepiece.bpe.model",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'xlm-roberta-base': 512,
+    'xlm-roberta-large': 512,
+    'xlm-roberta-large-finetuned-conll02-dutch': 512,
+    'xlm-roberta-large-finetuned-conll02-spanish': 512,
+    'xlm-roberta-large-finetuned-conll03-english': 512,
+    'xlm-roberta-large-finetuned-conll03-german': 512,
+}
+
+class XLMRobertaTokenizer(PreTrainedTokenizer):
+    """
+        Adapted from RobertaTokenizer and XLNetTokenizer
+        SentencePiece based tokenizer. Peculiarities:
+
+            - requires `SentencePiece <https://github.com/google/sentencepiece>`_
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, bos_token="<s>", eos_token="</s>", sep_token="</s>",
+                 cls_token="<s>", unk_token="<unk>", pad_token='<pad>', mask_token='<mask>',
+                 **kwargs):
+        super(XLMRobertaTokenizer, self).__init__(bos_token=bos_token, eos_token=eos_token, unk_token=unk_token,
+                                                  sep_token=sep_token, cls_token=cls_token, pad_token=pad_token,
+                                                  mask_token=mask_token,
+                                                  **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 4  # take into account special tokens
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(str(vocab_file))
+        self.vocab_file = vocab_file
+
+        # Original fairseq vocab and spm vocab must be "aligned":
+        # Vocab    |    0    |    1    |   2    |    3    |  4  |  5  |  6  |   7   |   8   |  9
+        # -------- | ------- | ------- | ------ | ------- | --- | --- | --- | ----- | ----- | ----
+        # fairseq  | '<s>'   | '<pad>' | '</s>' | '<unk>' | ',' | '.' | '▁' | 's'   | '▁de' | '-'
+        # spm      | '<unk>' | '<s>'   | '</s>' | ','     | '.' | '▁' | 's' | '▁de' | '-'   | '▁a'
+
+        # Mimic fairseq token-to-id alignment for the first 4 token
+        self.fairseq_tokens_to_ids = {"<s>": 0, "<pad>": 1, "</s>": 2, "<unk>": 3}
+
+        # The first "real" token "," has position 4 in the original fairseq vocab and position 3 in the spm vocab
+        self.fairseq_offset = 1
+
+        self.fairseq_tokens_to_ids['<mask>'] = len(self.sp_model) + len(self.fairseq_tokens_to_ids)
+        self.fairseq_ids_to_tokens = {v: k for k, v in self.fairseq_tokens_to_ids.items()}
+
+    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 RoBERTa sequence has the following format:
+            single sequence: <s> X </s>
+            pair of sequences: <s> A </s></s> B </s>
+        """
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        """
+        Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer ``prepare_for_model`` or ``encode_plus`` methods.
+
+        Args:
+            token_ids_0: list of ids (must not contain special tokens)
+            token_ids_1: Optional list of ids (must not contain special tokens), necessary when fetching sequence ids
+                for sequence pairs
+            already_has_special_tokens: (default False) Set to True if the token list is already formated with
+                special tokens for the model
+
+        Returns:
+            A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            if token_ids_1 is not None:
+                raise ValueError("You should not supply a second sequence if the provided sequence of "
+                                 "ids is already formated with special tokens for the model.")
+            return list(map(lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0, token_ids_0))
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
+        A RoBERTa sequence pair mask has the following format:
+        0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence
+
+        if token_ids_1 is None, only returns the first portion of the mask (0'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 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    @property
+    def vocab_size(self):
+        return len(self.sp_model) + len(self.fairseq_tokens_to_ids)
+
+    def _tokenize(self, text):
+        return self.sp_model.EncodeAsPieces(text)
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        if token in self.fairseq_tokens_to_ids:
+            return self.fairseq_tokens_to_ids[token]
+        return self.sp_model.PieceToId(token) + self.fairseq_offset
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        if index in self.fairseq_ids_to_tokens:
+            return self.fairseq_ids_to_tokens[index]
+        return self.sp_model.IdToPiece(index - self.fairseq_offset)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (strings for sub-words) in a single string."""
+        out_string = ''.join(tokens).replace(SPIECE_UNDERLINE, ' ').strip()
+        return out_string
+
+    def save_vocabulary(self, save_directory):
+        """ Save the sentencepiece vocabulary (copy original file) and special tokens file
+            to a directory.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        out_vocab_file = os.path.join(save_directory, VOCAB_FILES_NAMES['vocab_file'])
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)

transformers/tokenization_xlnet.py

@@ -60,6 +60,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
     vocab_files_names = VOCAB_FILES_NAMES
     pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
     max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    padding_side = "left"
 
     def __init__(self, vocab_file,
                  do_lower_case=False, remove_space=True, keep_accents=False,
@@ -74,6 +75,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
 
         self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
         self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+        self._pad_token_type_id = 3
 
         try:
             import sentencepiece as spm
@@ -141,7 +143,7 @@ class XLNetTokenizer(PreTrainedTokenizer):
             pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
         new_pieces = []
         for piece in pieces:
-            if len(piece) > 1 and piece[-1] == ',' and piece[-2].isdigit():
+            if len(piece) > 1 and piece[-1] == str(',') and piece[-2].isdigit():
                 cur_pieces = self.sp_model.EncodeAsPieces(
                     piece[:-1].replace(SPIECE_UNDERLINE, ''))
                 if piece[0] != SPIECE_UNDERLINE and cur_pieces[0][0] == SPIECE_UNDERLINE:
@@ -185,9 +187,9 @@ class XLNetTokenizer(PreTrainedTokenizer):
         """
         Build model inputs from a sequence or a pair of sequence for sequence classification tasks
         by concatenating and adding special tokens.
-        A RoBERTa sequence has the following format:
-            single sequence: <s> X </s>
-            pair of sequences: <s> A </s></s> B </s>
+        An XLNet sequence has the following format:
+            single sequence: X <sep> <cls>
+            pair of sequences: A <sep> B <sep> <cls>
         """
         sep = [self.sep_token_id]
         cls = [self.cls_token_id]
@@ -224,10 +226,10 @@ class XLNetTokenizer(PreTrainedTokenizer):
     def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
         """
         Creates a mask from the two sequences passed to be used in a sequence-pair classification task.
-        A BERT sequence pair mask has the following format:
+        An XLNet sequence pair mask has the following format:
         0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 2
         | first sequence    | second sequence     | CLS segment ID
-        
+
         if token_ids_1 is None, only returns the first portion of the mask (0's).
         """
         sep = [self.sep_token_id]

utils/download_glue_data.py

+''' Script for downloading all GLUE data.
+Original source: https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e
+
+Note: for legal reasons, we are unable to host MRPC.
+You can either use the version hosted by the SentEval team, which is already tokenized, 
+or you can download the original data from (https://download.microsoft.com/download/D/4/6/D46FF87A-F6B9-4252-AA8B-3604ED519838/MSRParaphraseCorpus.msi) and extract the data from it manually.
+For Windows users, you can run the .msi file. For Mac and Linux users, consider an external library such as 'cabextract' (see below for an example).
+You should then rename and place specific files in a folder (see below for an example).
+
+mkdir MRPC
+cabextract MSRParaphraseCorpus.msi -d MRPC
+cat MRPC/_2DEC3DBE877E4DB192D17C0256E90F1D | tr -d $'\r' > MRPC/msr_paraphrase_train.txt
+cat MRPC/_D7B391F9EAFF4B1B8BCE8F21B20B1B61 | tr -d $'\r' > MRPC/msr_paraphrase_test.txt
+rm MRPC/_*
+rm MSRParaphraseCorpus.msi
+
+1/30/19: It looks like SentEval is no longer hosting their extracted and tokenized MRPC data, so you'll need to download the data from the original source for now.
+2/11/19: It looks like SentEval actually *is* hosting the extracted data. Hooray!
+'''
+
+import os
+import sys
+import shutil
+import argparse
+import tempfile
+import urllib.request
+import zipfile
+
+TASKS = ["CoLA", "SST", "MRPC", "QQP", "STS", "MNLI", "SNLI", "QNLI", "RTE", "WNLI", "diagnostic"]
+TASK2PATH = {"CoLA":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FCoLA.zip?alt=media&token=46d5e637-3411-4188-bc44-5809b5bfb5f4',
+             "SST":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FSST-2.zip?alt=media&token=aabc5f6b-e466-44a2-b9b4-cf6337f84ac8',
+             "MRPC":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2Fmrpc_dev_ids.tsv?alt=media&token=ec5c0836-31d5-48f4-b431-7480817f1adc',
+             "QQP":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FQQP.zip?alt=media&token=700c6acf-160d-4d89-81d1-de4191d02cb5',
+             "STS":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FSTS-B.zip?alt=media&token=bddb94a7-8706-4e0d-a694-1109e12273b5',
+             "MNLI":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FMNLI.zip?alt=media&token=50329ea1-e339-40e2-809c-10c40afff3ce',
+             "SNLI":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FSNLI.zip?alt=media&token=4afcfbb2-ff0c-4b2d-a09a-dbf07926f4df',
+             "QNLI": 'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FQNLIv2.zip?alt=media&token=6fdcf570-0fc5-4631-8456-9505272d1601',
+             "RTE":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FRTE.zip?alt=media&token=5efa7e85-a0bb-4f19-8ea2-9e1840f077fb',
+             "WNLI":'https://firebasestorage.googleapis.com/v0/b/mtl-sentence-representations.appspot.com/o/data%2FWNLI.zip?alt=media&token=068ad0a0-ded7-4bd7-99a5-5e00222e0faf',
+             "diagnostic":'https://storage.googleapis.com/mtl-sentence-representations.appspot.com/tsvsWithoutLabels%2FAX.tsv?GoogleAccessId=firebase-adminsdk-0khhl@mtl-sentence-representations.iam.gserviceaccount.com&Expires=2498860800&Signature=DuQ2CSPt2Yfre0C%2BiISrVYrIFaZH1Lc7hBVZDD4ZyR7fZYOMNOUGpi8QxBmTNOrNPjR3z1cggo7WXFfrgECP6FBJSsURv8Ybrue8Ypt%2FTPxbuJ0Xc2FhDi%2BarnecCBFO77RSbfuz%2Bs95hRrYhTnByqu3U%2FYZPaj3tZt5QdfpH2IUROY8LiBXoXS46LE%2FgOQc%2FKN%2BA9SoscRDYsnxHfG0IjXGwHN%2Bf88q6hOmAxeNPx6moDulUF6XMUAaXCSFU%2BnRO2RDL9CapWxj%2BDl7syNyHhB7987hZ80B%2FwFkQ3MEs8auvt5XW1%2Bd4aCU7ytgM69r8JDCwibfhZxpaa4gd50QXQ%3D%3D'}
+
+MRPC_TRAIN = 'https://dl.fbaipublicfiles.com/senteval/senteval_data/msr_paraphrase_train.txt'
+MRPC_TEST = 'https://dl.fbaipublicfiles.com/senteval/senteval_data/msr_paraphrase_test.txt'
+
+def download_and_extract(task, data_dir):
+    print("Downloading and extracting %s..." % task)
+    data_file = "%s.zip" % task
+    urllib.request.urlretrieve(TASK2PATH[task], data_file)
+    with zipfile.ZipFile(data_file) as zip_ref:
+        zip_ref.extractall(data_dir)
+    os.remove(data_file)
+    print("\tCompleted!")
+
+def format_mrpc(data_dir, path_to_data):
+    print("Processing MRPC...")
+    mrpc_dir = os.path.join(data_dir, "MRPC")
+    if not os.path.isdir(mrpc_dir):
+        os.mkdir(mrpc_dir)
+    if path_to_data:
+        mrpc_train_file = os.path.join(path_to_data, "msr_paraphrase_train.txt")
+        mrpc_test_file = os.path.join(path_to_data, "msr_paraphrase_test.txt")
+    else:
+        print("Local MRPC data not specified, downloading data from %s" % MRPC_TRAIN)
+        mrpc_train_file = os.path.join(mrpc_dir, "msr_paraphrase_train.txt")
+        mrpc_test_file = os.path.join(mrpc_dir, "msr_paraphrase_test.txt")
+        urllib.request.urlretrieve(MRPC_TRAIN, mrpc_train_file)
+        urllib.request.urlretrieve(MRPC_TEST, mrpc_test_file)
+    assert os.path.isfile(mrpc_train_file), "Train data not found at %s" % mrpc_train_file
+    assert os.path.isfile(mrpc_test_file), "Test data not found at %s" % mrpc_test_file
+    urllib.request.urlretrieve(TASK2PATH["MRPC"], os.path.join(mrpc_dir, "dev_ids.tsv"))
+
+    dev_ids = []
+    with open(os.path.join(mrpc_dir, "dev_ids.tsv"), encoding="utf8") as ids_fh:
+        for row in ids_fh:
+            dev_ids.append(row.strip().split('\t'))
+
+    with open(mrpc_train_file, encoding="utf8") as data_fh, \
+         open(os.path.join(mrpc_dir, "train.tsv"), 'w', encoding="utf8") as train_fh, \
+         open(os.path.join(mrpc_dir, "dev.tsv"), 'w', encoding="utf8") as dev_fh:
+        header = data_fh.readline()
+        train_fh.write(header)
+        dev_fh.write(header)
+        for row in data_fh:
+            label, id1, id2, s1, s2 = row.strip().split('\t')
+            if [id1, id2] in dev_ids:
+                dev_fh.write("%s\t%s\t%s\t%s\t%s\n" % (label, id1, id2, s1, s2))
+            else:
+                train_fh.write("%s\t%s\t%s\t%s\t%s\n" % (label, id1, id2, s1, s2))
+
+    with open(mrpc_test_file, encoding="utf8") as data_fh, \
+            open(os.path.join(mrpc_dir, "test.tsv"), 'w', encoding="utf8") as test_fh:
+        header = data_fh.readline()
+        test_fh.write("index\t#1 ID\t#2 ID\t#1 String\t#2 String\n")
+        for idx, row in enumerate(data_fh):
+            label, id1, id2, s1, s2 = row.strip().split('\t')
+            test_fh.write("%d\t%s\t%s\t%s\t%s\n" % (idx, id1, id2, s1, s2))
+    print("\tCompleted!")
+
+def download_diagnostic(data_dir):
+    print("Downloading and extracting diagnostic...")
+    if not os.path.isdir(os.path.join(data_dir, "diagnostic")):
+        os.mkdir(os.path.join(data_dir, "diagnostic"))
+    data_file = os.path.join(data_dir, "diagnostic", "diagnostic.tsv")
+    urllib.request.urlretrieve(TASK2PATH["diagnostic"], data_file)
+    print("\tCompleted!")
+    return
+
+def get_tasks(task_names):
+    task_names = task_names.split(',')
+    if "all" in task_names:
+        tasks = TASKS
+    else:
+        tasks = []
+        for task_name in task_names:
+            assert task_name in TASKS, "Task %s not found!" % task_name
+            tasks.append(task_name)
+    return tasks
+
+def main(arguments):
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data_dir', help='directory to save data to', type=str, default='glue_data')
+    parser.add_argument('--tasks', help='tasks to download data for as a comma separated string',
+                        type=str, default='all')
+    parser.add_argument('--path_to_mrpc', help='path to directory containing extracted MRPC data, msr_paraphrase_train.txt and msr_paraphrase_text.txt',
+                        type=str, default='')
+    args = parser.parse_args(arguments)
+
+    if not os.path.isdir(args.data_dir):
+        os.mkdir(args.data_dir)
+    tasks = get_tasks(args.tasks)
+
+    for task in tasks:
+        if task == 'MRPC':
+            format_mrpc(args.data_dir, args.path_to_mrpc)
+        elif task == 'diagnostic':
+            download_diagnostic(args.data_dir)
+        else:
+            download_and_extract(task, args.data_dir)
+
+
+if __name__ == '__main__':
+    sys.exit(main(sys.argv[1:]))

utils/link_tester.py

+""" Link tester.
+
+This little utility reads all the python files in the repository,
+scans for links pointing to S3 and tests the links one by one. Raises an error
+at the end of the scan if at least one link was reported broken.
+"""
+import os
+import re
+import sys
+
+import requests
+
+
+REGEXP_FIND_S3_LINKS = r"""([\"'])(https:\/\/s3)(.*)?\1"""
+
+
+def list_python_files_in_repository():
+    """ List all python files in the repository.
+
+    This function assumes that the script is executed in the root folder.
+    """
+    source_code_files = []
+    for path, subdirs, files in os.walk("."):
+        if "templates" in path:
+            continue
+        for name in files:
+            if ".py" in name and ".pyc" not in name:
+                path_to_files = os.path.join(path, name)
+                source_code_files.append(path_to_files)
+
+    return source_code_files
+
+
+def find_all_links(file_paths):
+    links = []
+    for path in file_paths:
+        links += scan_code_for_links(path)
+
+    return links
+
+
+def scan_code_for_links(source):
+    """ Scans the file to find links using a regular expression.
+    Returns a list of links.
+    """
+    with open(source, 'r') as content:
+        content = content.read()
+        raw_links = re.findall(REGEXP_FIND_S3_LINKS, content)
+        links = [prefix + suffix for _, prefix, suffix in raw_links]
+
+    return links
+
+
+def check_all_links(links):
+    """ Check that the provided links are valid.
+
+    Links are considered valid if a HEAD request to the server
+    returns a 200 status code.
+    """
+    broken_links = []
+    for link in links:
+        head = requests.head(link)
+        if head.status_code != 200:
+            broken_links.append(link)
+
+    return broken_links
+
+
+if __name__ == "__main__":
+    file_paths = list_python_files_in_repository()
+    links = find_all_links(file_paths)
+    broken_links = check_all_links(links)
+    print("Looking for broken links to pre-trained models/configs/tokenizers...")
+    if broken_links:
+        print("The following links did not respond:")
+        for link in broken_links:
+            print("- {}".format(link))
+        sys.exit(1)
+    print("All links are ok.")

valohai.yaml

+---
+
+- step:
+    name: Execute python examples/run_glue.py
+    image: pytorch/pytorch:nightly-devel-cuda10.0-cudnn7
+    command:
+      - python /valohai/repository/utils/download_glue_data.py --data_dir=/glue_data
+      - pip install -e .
+      - pip install -r examples/requirements.txt
+      - python examples/run_glue.py --do_train --data_dir=/glue_data/{parameter-value:task_name} {parameters}
+    parameters:
+      - name: model_type
+        pass-as: --model_type={v}
+        type: string
+        default: bert
+      - name: model_name_or_path
+        pass-as: --model_name_or_path={v}
+        type: string
+        default: bert-base-uncased
+      - name: task_name
+        pass-as: --task_name={v}
+        type: string
+        default: MRPC
+      - name: max_seq_length
+        pass-as: --max_seq_length={v}
+        description: The maximum total input sequence length after tokenization. Sequences longer than this will be truncated, sequences shorter will be padded.
+        type: integer
+        default: 128
+      - name: per_gpu_train_batch_size
+        pass-as: --per_gpu_train_batch_size={v}
+        description: Batch size per GPU/CPU for training.
+        type: integer
+        default: 8
+      - name: per_gpu_eval_batch_size
+        pass-as: --per_gpu_eval_batch_size={v}
+        description: Batch size per GPU/CPU for evaluation.
+        type: integer
+        default: 8
+      - name: gradient_accumulation_steps
+        pass-as: --gradient_accumulation_steps={v}
+        description: Number of updates steps to accumulate before performing a backward/update pass.
+        type: integer
+        default: 1
+      - name: learning_rate
+        pass-as: --learning_rate={v}
+        description: The initial learning rate for Adam.
+        type: float
+        default: 0.00005
+      - name: adam_epsilon
+        pass-as: --adam_epsilon={v}
+        description: Epsilon for Adam optimizer.
+        type: float
+        default: 0.00000001
+      - name: max_grad_norm
+        pass-as: --max_grad_norm={v}
+        description: Max gradient norm.
+        type: float
+        default: 1.0
+      - name: num_train_epochs
+        pass-as: --num_train_epochs={v}
+        description: Total number of training epochs to perform.
+        type: integer
+        default: 3
+      - name: max_steps
+        pass-as: --max_steps={v}
+        description: If > 0, set total number of training steps to perform. Override num_train_epochs.
+        type: integer
+        default: -1
+      - name: warmup_steps
+        pass-as: --warmup_steps={v}
+        description: Linear warmup over warmup_steps.
+        type: integer
+        default: -1
+      - name: logging_steps
+        pass-as: --logging_steps={v}
+        description: Log every X updates steps.
+        type: integer
+        default: 25
+      - name: save_steps
+        pass-as: --save_steps={v}
+        description: Save checkpoint every X updates steps.
+        type: integer
+        default: -1
+      - name: output_dir
+        pass-as: --output_dir={v}
+        type: string
+        default: /valohai/outputs
+      - name: evaluate_during_training
+        description: Run evaluation during training at each logging step.
+        type: flag
+        default: true
+      - name: do_lower_case
+        description: Set this flag if you are using an uncased model.
+        type: flag