Merge branch 'master' into distilbert-german

transformers/tests/modeling_albert_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+
+from transformers import is_torch_available
+
+from .modeling_common_test import (CommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+
+if is_torch_available():
+    from transformers import (AlbertConfig, AlbertModel, AlbertForMaskedLM,
+                              AlbertForSequenceClassification, AlbertForQuestionAnswering,
+                              )
+    from transformers.modeling_albert import ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+else:
+    pytestmark = pytest.mark.skip("Require Torch")
+
+
+class AlbertModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (AlbertModel, AlbertForMaskedLM) if is_torch_available() else ()
+
+    class AlbertModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     embedding_size=16,
+                     hidden_size=36,
+                     num_hidden_layers=6,
+                     num_hidden_groups=6,
+                     num_attention_heads=6,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.embedding_size = embedding_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+            self.num_hidden_groups = num_hidden_groups
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = AlbertConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                hidden_size=self.hidden_size,
+                num_hidden_layers=self.num_hidden_layers,
+                num_attention_heads=self.num_attention_heads,
+                intermediate_size=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                hidden_dropout_prob=self.hidden_dropout_prob,
+                attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                type_vocab_size=self.type_vocab_size,
+                initializer_range=self.initializer_range,
+                num_hidden_groups=self.num_hidden_groups)
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_albert_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = AlbertModel(config=config)
+            model.eval()
+            sequence_output, pooled_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+            sequence_output, pooled_output = model(input_ids, token_type_ids=token_type_ids)
+            sequence_output, pooled_output = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output,
+                "pooled_output": pooled_output,
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].size()),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooled_output"].size()), [self.batch_size, self.hidden_size])
+
+
+        def create_and_check_albert_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = AlbertForMaskedLM(config=config)
+            model.eval()
+            loss, prediction_scores = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, masked_lm_labels=token_labels)
+            result = {
+                "loss": loss,
+                "prediction_scores": prediction_scores,
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].size()),
+                [self.batch_size, self.seq_length, self.vocab_size])
+            self.check_loss_output(result)
+
+        def create_and_check_albert_for_question_answering(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = AlbertForQuestionAnswering(config=config)
+            model.eval()
+            loss, start_logits, end_logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids,
+                                                   start_positions=sequence_labels, end_positions=sequence_labels)
+            result = {
+                "loss": loss,
+                "start_logits": start_logits,
+                "end_logits": end_logits,
+            }
+            self.parent.assertListEqual(
+                list(result["start_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.parent.assertListEqual(
+                list(result["end_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.check_loss_output(result)
+
+
+        def create_and_check_albert_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = AlbertForSequenceClassification(config)
+            model.eval()
+            loss, logits = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.num_labels])
+            self.check_loss_output(result)
+
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_mask,
+             sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = AlbertModelTest.AlbertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=AlbertConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_albert_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_sequence_classification(*config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in list(ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = AlbertModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()

transformers/tests/modeling_tf_albert_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+import sys
+
+from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+
+from transformers import AlbertConfig, is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+    from transformers.modeling_tf_albert import (TFAlbertModel, TFAlbertForMaskedLM,
+                                                 TFAlbertForSequenceClassification,
+                                                 TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+else:
+    pytestmark = pytest.mark.skip("Require TensorFlow")
+
+
+class TFAlbertModelTest(TFCommonTestCases.TFCommonModelTester):
+
+    all_model_classes = (
+        TFAlbertModel,
+        TFAlbertForMaskedLM,
+        TFAlbertForSequenceClassification
+    ) if is_tf_available() else ()
+
+    class TFAlbertModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     embedding_size=16,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                     ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.embedding_size = embedding_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor(
+                [self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor(
+                    [self.batch_size, self.seq_length], vocab_size=2)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                token_type_ids = ids_tensor(
+                    [self.batch_size, self.seq_length], self.type_vocab_size)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor(
+                    [self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor(
+                    [self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = AlbertConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                hidden_size=self.hidden_size,
+                num_hidden_layers=self.num_hidden_layers,
+                num_attention_heads=self.num_attention_heads,
+                intermediate_size=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                hidden_dropout_prob=self.hidden_dropout_prob,
+                attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                type_vocab_size=self.type_vocab_size,
+                initializer_range=self.initializer_range)
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def create_and_check_albert_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = TFAlbertModel(config=config)
+            # inputs = {'input_ids': input_ids,
+            #           'attention_mask': input_mask,
+            #           'token_type_ids': token_type_ids}
+            # sequence_output, pooled_output = model(**inputs)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            sequence_output, pooled_output = model(inputs)
+
+            inputs = [input_ids, input_mask]
+            sequence_output, pooled_output = model(inputs)
+
+            sequence_output, pooled_output = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output.numpy(),
+                "pooled_output": pooled_output.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].shape),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooled_output"].shape), [
+                                        self.batch_size, self.hidden_size])
+
+        def create_and_check_albert_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = TFAlbertForMaskedLM(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            prediction_scores, = model(inputs)
+            result = {
+                "prediction_scores": prediction_scores.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].shape),
+                [self.batch_size, self.seq_length, self.vocab_size])
+
+        def create_and_check_albert_for_sequence_classification(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = TFAlbertForSequenceClassification(config=config)
+            inputs = {'input_ids': input_ids,
+                      'attention_mask': input_mask,
+                      'token_type_ids': token_type_ids}
+            logits, = model(inputs)
+            result = {
+                "logits": logits.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].shape),
+                [self.batch_size, self.num_labels])
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_mask,
+             sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids,
+                           'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = TFAlbertModelTest.TFAlbertModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=AlbertConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_albert_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_masked_lm(
+            *config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_albert_for_sequence_classification(
+            *config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        # for model_name in list(TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+        for model_name in ['albert-base-uncased']:
+            model = TFAlbertModel.from_pretrained(
+                model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+
+if __name__ == "__main__":
+    unittest.main()

transformers/tests/modeling_tf_common_test.py

@@ -426,9 +426,17 @@ class TFCommonTestCases:
                     try:
                         x = wte([input_ids], mode="embedding")
                     except:
-                        x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32)
+                        try:
+                            x = wte([input_ids, None, None, None], mode="embedding")
+                        except:
+                            if hasattr(self.model_tester, "embedding_size"):
+                                x = tf.ones(input_ids.shape + [self.model_tester.embedding_size], dtype=tf.dtypes.float32)
+                            else:
+                                x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32)
                 # ^^ In our TF models, the input_embeddings can take slightly different forms,
-                # so we try two of them and fall back to just synthetically creating a dummy tensor of ones.
+                # so we try a few of them.
+                # We used to fall back to just synthetically creating a dummy tensor of ones:
+                #
                 inputs_dict["inputs_embeds"] = x
                 outputs = model(inputs_dict)
 

transformers/tests/tokenization_albert_test.py

+# coding=utf-8
+# Copyright 2019 Hugging Face inc.
+#
+# 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.
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+import unittest
+
+from transformers.tokenization_albert import (AlbertTokenizer, SPIECE_UNDERLINE)
+
+from .tokenization_tests_commons import CommonTestCases
+
+SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)),
+                    'fixtures/spiece.model')
+
+class AlbertTokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = AlbertTokenizer
+
+    def setUp(self):
+        super(AlbertTokenizationTest, self).setUp()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = AlbertTokenizer(SAMPLE_VOCAB)
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return AlbertTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self):
+        input_text = u"this is a test"
+        output_text = u"this is a test"
+        return input_text, output_text
+
+
+    def test_full_tokenizer(self):
+        tokenizer = AlbertTokenizer(SAMPLE_VOCAB, keep_accents=True)
+
+        tokens = tokenizer.tokenize(u'This is a test')
+        self.assertListEqual(tokens, [u'▁this', u'▁is', u'▁a', u'▁test'])
+
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(tokens), [48, 25, 21, 1289])
+
+        tokens = tokenizer.tokenize(u"I was born in 92000, and this is falsé.")
+        self.assertListEqual(tokens, [u'▁i', u'▁was', u'▁born', u'▁in', u'▁9', u'2000', u',', u'▁and', u'▁this', u'▁is', u'▁fal', u's', u'é', u'.'])
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(ids, [31, 23, 386, 19, 561, 3050, 15, 17, 48, 25, 8256, 18, 1, 9])
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(back_tokens, ['▁i', '▁was', '▁born', '▁in', '▁9', '2000', ',', '▁and', '▁this', '▁is', '▁fal', 's', '<unk>', '.'])
+
+    def test_sequence_builders(self):
+        tokenizer = AlbertTokenizer(SAMPLE_VOCAB)
+
+        text = tokenizer.encode("sequence builders")
+        text_2 = tokenizer.encode("multi-sequence build")
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id]
+        assert encoded_pair == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id] + text_2 + [tokenizer.sep_token_id]
+
+
+if __name__ == '__main__':
+    unittest.main()

transformers/tests/tokenization_tests_commons.py

@@ -110,6 +110,36 @@ class CommonTestCases:
 
             self.assertListEqual(subwords, subwords_loaded)
 
+        def test_added_tokens_do_lower_case(self):
+            tokenizer = self.get_tokenizer(do_lower_case=True)
+
+            text = "aaaaa bbbbbb low cccccccccdddddddd l"
+            text2 = "AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l"
+
+            toks0 = tokenizer.tokenize(text)  # toks before adding new_toks
+
+            new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd", 'AAAAA BBBBBB', 'CCCCCCCCCDDDDDDDD']
+            added = tokenizer.add_tokens(new_toks)
+            self.assertEqual(added, 2)
+
+            toks = tokenizer.tokenize(text)
+            toks2 = tokenizer.tokenize(text2)
+
+            self.assertEqual(len(toks), len(toks2))
+            self.assertNotEqual(len(toks), len(toks0))  # toks0 should be longer
+            self.assertListEqual(toks, toks2)
+
+            tokenizer = self.get_tokenizer(do_lower_case=False)
+
+            added = tokenizer.add_tokens(new_toks)
+            self.assertEqual(added, 4)
+
+            toks = tokenizer.tokenize(text)
+            toks2 = tokenizer.tokenize(text2)
+
+            self.assertEqual(len(toks), len(toks2))  # Length should still be the same
+            self.assertNotEqual(len(toks), len(toks0))
+            self.assertNotEqual(toks[0], toks2[0])  # But at least the first tokens should differ
 
         def test_add_tokens_tokenizer(self):
             tokenizer = self.get_tokenizer()
@@ -243,7 +273,11 @@ class CommonTestCases:
             sequence = tokenizer.encode(seq_0, add_special_tokens=False)
             num_added_tokens = tokenizer.num_added_tokens()
             total_length = len(sequence) + num_added_tokens
-            information = tokenizer.encode_plus(seq_0, max_length=total_length - 2, add_special_tokens=True, stride=stride)
+            information = tokenizer.encode_plus(seq_0,
+                                                max_length=total_length - 2,
+                                                add_special_tokens=True,
+                                                stride=stride,
+                                                return_overflowing_tokens=True)
 
             truncated_sequence = information["input_ids"]
             overflowing_tokens = information["overflowing_tokens"]
@@ -270,10 +304,12 @@ class CommonTestCases:
             )
 
             information = tokenizer.encode_plus(seq_0, seq_1, max_length=len(sequence) - 2, add_special_tokens=True,
-                                                stride=stride, truncation_strategy='only_second')
+                                                stride=stride, truncation_strategy='only_second',
+                                                return_overflowing_tokens=True)
             information_first_truncated = tokenizer.encode_plus(seq_0, seq_1, max_length=len(sequence) - 2,
                                                                 add_special_tokens=True, stride=stride,
-                                                                truncation_strategy='only_first')
+                                                                truncation_strategy='only_first',
+                                                                return_overflowing_tokens=True)
 
             truncated_sequence = information["input_ids"]
             overflowing_tokens = information["overflowing_tokens"]
@@ -305,7 +341,7 @@ class CommonTestCases:
 
             # Testing single inputs
             encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
-            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, add_special_tokens=True)
+            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, add_special_tokens=True, return_special_tokens_mask=True)
             encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
             special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
             self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
@@ -317,7 +353,8 @@ class CommonTestCases:
             # Testing inputs pairs
             encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False) + tokenizer.encode(sequence_1,
                                                                                                          add_special_tokens=False)
-            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, sequence_1, add_special_tokens=True)
+            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, sequence_1, add_special_tokens=True,
+                                                          return_special_tokens_mask=True)
             encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
             special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
             self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
@@ -329,7 +366,9 @@ class CommonTestCases:
             # Testing with already existing special tokens
             if tokenizer.cls_token_id == tokenizer.unk_token_id and tokenizer.cls_token_id == tokenizer.unk_token_id:
                 tokenizer.add_special_tokens({'cls_token': '</s>', 'sep_token': '<s>'})
-            encoded_sequence_dict = tokenizer.encode_plus(sequence_0, add_special_tokens=True)
+            encoded_sequence_dict = tokenizer.encode_plus(sequence_0,
+                                                          add_special_tokens=True,
+                                                          return_special_tokens_mask=True)
             encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
             special_tokens_mask_orig = encoded_sequence_dict["special_tokens_mask"]
             special_tokens_mask = tokenizer.get_special_tokens_mask(encoded_sequence_w_special, already_has_special_tokens=True)

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] == ',' 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

@@ -90,6 +90,9 @@ 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.

transformers/tokenization_camembert.py

@@ -16,9 +16,14 @@
 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
 
+logger = logging.getLogger(__name__)
 
 VOCAB_FILES_NAMES = {'vocab_file': 'sentencepiece.bpe.model'}
 
@@ -55,6 +60,7 @@ class CamembertTokenizer(PreTrainedTokenizer):
         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}
@@ -135,3 +141,17 @@ class CamembertTokenizer(PreTrainedTokenizer):
         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 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_distilbert.py

@@ -34,6 +34,7 @@ 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",
     }
 }
 
@@ -41,6 +42,7 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     'distilbert-base-uncased': 512,
     'distilbert-base-uncased-distilled-squad': 512,
     'distilbert-base-german-cased': 512,
+    'distilbert-base-multilingual-cased': 512,
 }
 
 

transformers/tokenization_gpt2.py

@@ -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
@@ -184,7 +184,7 @@ 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

transformers/tokenization_utils.py

@@ -252,6 +252,9 @@ 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.
@@ -287,6 +290,7 @@ 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())
@@ -353,7 +357,7 @@ 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."
@@ -513,6 +517,8 @@ 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):
+                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:
@@ -606,6 +612,9 @@ class PreTrainedTokenizer(object):
 
             Take care of added tokens.
         """
+        if self.init_kwargs.get('do_lower_case', False):
+            text = text.lower()
+
         def split_on_token(tok, text):
             result = []
             split_text = text.split(tok)
@@ -741,6 +750,9 @@ class PreTrainedTokenizer(object):
                     stride=0,
                     truncation_strategy='longest_first',
                     return_tensors=None,
+                    return_token_type_ids=True,
+                    return_overflowing_tokens=False,
+                    return_special_tokens_mask=False,
                     **kwargs):
         """
         Returns a dictionary containing the encoded sequence or sequence pair and additional informations:
@@ -767,7 +779,30 @@ class PreTrainedTokenizer(object):
                 - '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.
+            return_token_type_ids: (optional) Set to False to avoid returning token_type_ids (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)
+                    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
+
+                ``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):
@@ -789,10 +824,17 @@ class PreTrainedTokenizer(object):
                                       add_special_tokens=add_special_tokens,
                                       stride=stride,
                                       truncation_strategy=truncation_strategy,
-                                      return_tensors=return_tensors)
+                                      return_tensors=return_tensors,
+                                      return_token_type_ids=return_token_type_ids,
+                                      return_overflowing_tokens=return_overflowing_tokens,
+                                      return_special_tokens_mask=return_special_tokens_mask)
 
     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',
+                          return_tensors=None,
+                          return_token_type_ids=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
@@ -817,21 +859,27 @@ class PreTrainedTokenizer(object):
                 - '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.
+            return_token_type_ids: (optional) Set to False to avoid returning token_type_ids (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.
         """
@@ -840,23 +888,31 @@ 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)
+            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 [])
+            special_tokens_mask = [0] * (len(ids) + (len(pair_ids) if pair else 0))
+        if return_special_tokens_mask:
+            encoded_inputs["special_tokens_mask"] = self.get_special_tokens_mask(ids, pair_ids)
 
+        # Prepare inputs as tensors if asked
         if return_tensors == 'tf' and is_tf_available():
             sequence = tf.constant([sequence])
             token_type_ids = tf.constant([token_type_ids])
@@ -867,12 +923,15 @@ class PreTrainedTokenizer(object):
             logger.warning("Unable to convert output to tensors format {}, PyTorch or TensorFlow is not available.".format(return_tensors))
 
         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 "

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)
 
@@ -758,9 +758,9 @@ 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_xlnet.py

@@ -185,9 +185,9 @@ 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]
@@ -224,7 +224,7 @@ 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