Merge branch 'master' into do_lower_case

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_bert_test.py

@@ -131,10 +131,6 @@ class TFBertModelTest(TFCommonTestCases.TFCommonModelTester):
 
         def create_and_check_bert_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels):
             model = TFBertModel(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}

transformers/tests/modeling_tf_common_test.py

@@ -360,6 +360,16 @@ class TFCommonTestCases:
             #     self.assertTrue(models_equal)
 
 
+        def test_model_common_attributes(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+                assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
+                x = model.get_output_embeddings()
+                assert x is None or isinstance(x, tf.keras.layers.Layer)
+
+
         def test_tie_model_weights(self):
             pass
             # config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -401,6 +411,35 @@ class TFCommonTestCases:
                 first, second = model(inputs_dict, training=False)[0], model(inputs_dict, training=False)[0]
                 self.assertTrue(tf.math.equal(first, second).numpy().all())
 
+        def test_inputs_embeds(self):
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            input_ids = inputs_dict["input_ids"]
+            del inputs_dict["input_ids"]
+
+            for model_class in self.all_model_classes:
+                model = model_class(config)
+
+                wte = model.get_input_embeddings()
+                try:
+                    x = wte(input_ids, mode="embedding")
+                except:
+                    try:
+                        x = wte([input_ids], mode="embedding")
+                    except:
+                        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 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)
+
 
 def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
     """Creates a random int32 tensor of the shape within the vocab size."""

transformers/tests/optimization_test.py

@@ -25,8 +25,12 @@ from transformers import is_torch_available
 if is_torch_available():
     import torch
 
-    from transformers import (AdamW, ConstantLRSchedule, WarmupConstantSchedule,
-                                    WarmupCosineSchedule, WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)
+    from transformers import (AdamW,
+                              get_constant_schedule,
+                              get_constant_schedule_with_warmup,
+                              get_cosine_schedule_with_warmup,
+                              get_cosine_with_hard_restarts_schedule_with_warmup,
+                              get_linear_schedule_with_warmup)
 else:
     pytestmark = pytest.mark.skip("Require Torch")
 
@@ -87,59 +91,60 @@ class ScheduleInitTest(unittest.TestCase):
             self.assertAlmostEqual(a, b, delta=tol)
 
     def test_constant_scheduler(self):
-        scheduler = ConstantLRSchedule(self.optimizer)
+        scheduler = get_constant_schedule(self.optimizer)
         lrs = unwrap_schedule(scheduler, self.num_steps)
         expected_learning_rates = [10.] * self.num_steps
         self.assertEqual(len(lrs[0]), 1)
         self.assertListEqual([l[0] for l in lrs], expected_learning_rates)
 
-        scheduler = ConstantLRSchedule(self.optimizer)
+        scheduler = get_constant_schedule(self.optimizer)
         lrs_2 = unwrap_and_save_reload_schedule(scheduler, self.num_steps)
         self.assertListEqual([l[0] for l in lrs], [l[0] for l in lrs_2])
 
     def test_warmup_constant_scheduler(self):
-        scheduler = WarmupConstantSchedule(self.optimizer, warmup_steps=4)
+        scheduler = get_constant_schedule_with_warmup(self.optimizer, num_warmup_steps=4)
         lrs = unwrap_schedule(scheduler, self.num_steps)
         expected_learning_rates = [2.5, 5.0, 7.5, 10.0, 10.0, 10.0, 10.0, 10.0, 10.0, 10.0]
         self.assertEqual(len(lrs[0]), 1)
         self.assertListEqual([l[0] for l in lrs], expected_learning_rates)
 
-        scheduler = WarmupConstantSchedule(self.optimizer, warmup_steps=4)
+        scheduler = get_constant_schedule_with_warmup(self.optimizer, num_warmup_steps=4)
         lrs_2 = unwrap_and_save_reload_schedule(scheduler, self.num_steps)
         self.assertListEqual([l[0] for l in lrs], [l[0] for l in lrs_2])
 
     def test_warmup_linear_scheduler(self):
-        scheduler = WarmupLinearSchedule(self.optimizer, warmup_steps=2, t_total=10)
+        scheduler = get_linear_schedule_with_warmup(self.optimizer, num_warmup_steps=2, num_training_steps=10)
         lrs = unwrap_schedule(scheduler, self.num_steps)
         expected_learning_rates = [5.0, 10.0, 8.75, 7.5, 6.25, 5.0, 3.75, 2.5, 1.25, 0.0]
         self.assertEqual(len(lrs[0]), 1)
         self.assertListEqual([l[0] for l in lrs], expected_learning_rates)
 
-        scheduler = WarmupLinearSchedule(self.optimizer, warmup_steps=2, t_total=10)
+        scheduler = get_linear_schedule_with_warmup(self.optimizer, num_warmup_steps=2, num_training_steps=10)
         lrs_2 = unwrap_and_save_reload_schedule(scheduler, self.num_steps)
         self.assertListEqual([l[0] for l in lrs], [l[0] for l in lrs_2])
 
     def test_warmup_cosine_scheduler(self):
-        scheduler = WarmupCosineSchedule(self.optimizer, warmup_steps=2, t_total=10)
+        scheduler = get_cosine_schedule_with_warmup(self.optimizer, num_warmup_steps=2, num_training_steps=10)
         lrs = unwrap_schedule(scheduler, self.num_steps)
         expected_learning_rates = [5.0, 10.0, 9.61, 8.53, 6.91, 5.0, 3.08, 1.46, 0.38, 0.0]
         self.assertEqual(len(lrs[0]), 1)
         self.assertListAlmostEqual([l[0] for l in lrs], expected_learning_rates, tol=1e-2)
 
-        scheduler = WarmupCosineSchedule(self.optimizer, warmup_steps=2, t_total=10)
+        scheduler = get_cosine_schedule_with_warmup(self.optimizer, num_warmup_steps=2, num_training_steps=10)
         lrs_2 = unwrap_and_save_reload_schedule(scheduler, self.num_steps)
         self.assertListEqual([l[0] for l in lrs], [l[0] for l in lrs_2])
 
     def test_warmup_cosine_hard_restart_scheduler(self):
-        scheduler = WarmupCosineWithHardRestartsSchedule(self.optimizer, warmup_steps=2, cycles=2, t_total=10)
+        scheduler = get_cosine_with_hard_restarts_schedule_with_warmup(self.optimizer, num_warmup_steps=2, num_cycles=2, num_training_steps=10)
         lrs = unwrap_schedule(scheduler, self.num_steps)
         expected_learning_rates = [5.0, 10.0, 8.53, 5.0, 1.46, 10.0, 8.53, 5.0, 1.46, 0.0]
         self.assertEqual(len(lrs[0]), 1)
         self.assertListAlmostEqual([l[0] for l in lrs], expected_learning_rates, tol=1e-2)
 
-        scheduler = WarmupCosineWithHardRestartsSchedule(self.optimizer, warmup_steps=2, cycles=2, t_total=10)
+        scheduler = get_cosine_with_hard_restarts_schedule_with_warmup(self.optimizer, num_warmup_steps=2, num_cycles=2, num_training_steps=10)
         lrs_2 = unwrap_and_save_reload_schedule(scheduler, self.num_steps)
         self.assertListEqual([l[0] for l in lrs], [l[0] for l in lrs_2])
 
+
 if __name__ == "__main__":
     unittest.main()

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

@@ -190,6 +190,27 @@ class CommonTestCases:
             self.assertEqual(tokens[0], tokenizer.eos_token_id)
             self.assertEqual(tokens[-2], tokenizer.pad_token_id)
 
+        def test_add_special_tokens(self):
+            tokenizer = self.get_tokenizer()
+            input_text, output_text = self.get_input_output_texts()
+
+            special_token = "[SPECIAL TOKEN]"
+
+            tokenizer.add_special_tokens({"cls_token": special_token})
+            encoded_special_token = tokenizer.encode(special_token, add_special_tokens=False)
+            assert len(encoded_special_token) == 1
+
+            text = " ".join([input_text, special_token, output_text])
+            encoded = tokenizer.encode(text, add_special_tokens=False)
+
+            input_encoded = tokenizer.encode(input_text, add_special_tokens=False)
+            output_encoded = tokenizer.encode(output_text, add_special_tokens=False)
+            special_token_id = tokenizer.encode(special_token, add_special_tokens=False)
+            assert encoded == input_encoded + special_token_id + output_encoded
+
+            decoded = tokenizer.decode(encoded, skip_special_tokens=True)
+            assert special_token not in decoded
+
         def test_required_methods_tokenizer(self):
             tokenizer = self.get_tokenizer()
             input_text, output_text = self.get_input_output_texts()

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

@@ -27,6 +27,7 @@ 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
 
 logger = logging.getLogger(__name__)
 
@@ -41,6 +42,7 @@ 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 `camembert`: CamembertTokenizer (CamemBERT model)
             - contains `distilbert`: DistilBertTokenizer (DistilBert model)
             - contains `roberta`: RobertaTokenizer (RoBERTa model)
             - contains `bert`: BertTokenizer (Bert model)
@@ -64,8 +66,9 @@ 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 `camembert`: CamembertTokenizer (CamemBERT model)
             - contains `distilbert`: DistilBertTokenizer (DistilBert model)
-            - contains `roberta`: RobertaTokenizer (XLM 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)
@@ -103,6 +106,8 @@ class AutoTokenizer(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return DistilBertTokenizer.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 'roberta' in pretrained_model_name_or_path:
             return RobertaTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -121,4 +126,4 @@ 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', 'camembert', 'ctrl'".format(pretrained_model_name_or_path))

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
+
+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 self.fairseq_offset + 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]
+        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 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_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

@@ -21,6 +21,7 @@ import os
 import json
 import six
 import copy
+import itertools
 from io import open
 
 from .file_utils import cached_path, is_tf_available, is_torch_available
@@ -646,9 +647,9 @@ class PreTrainedTokenizer(object):
                         tokenized_text += [sub_text]
                 text_list = tokenized_text
 
-            return sum((self._tokenize(token, **kwargs) if token not \
+            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 \
-                    else [token] for token in tokenized_text), [])
+                    else [token] for token in tokenized_text)))
 
         added_tokens = list(self.added_tokens_encoder.keys()) + self.all_special_tokens
         tokenized_text = split_on_tokens(added_tokens, text)
@@ -676,10 +677,6 @@ class PreTrainedTokenizer(object):
         ids = []
         for token in tokens:
             ids.append(self._convert_token_to_id_with_added_voc(token))
-        if len(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))
         return ids
 
     def _convert_token_to_id_with_added_voc(self, token):
@@ -882,6 +879,11 @@ class PreTrainedTokenizer(object):
             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 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))
+                           
         return encoded_inputs
 
     def truncate_sequences(self, ids, pair_ids=None, num_tokens_to_remove=0, truncation_strategy='longest_first', stride=0):
@@ -1060,7 +1062,7 @@ class PreTrainedTokenizer(object):
             class attributes (cls_token, unk_token...).
         """
         all_toks = self.all_special_tokens
-        all_ids = list(self._convert_token_to_id(t) for t in all_toks)
+        all_ids = self.convert_tokens_to_ids(all_toks)
         return all_ids
 
     @staticmethod

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