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

transformers/tests/modeling_tf_xlnet_test.py

@@ -20,8 +20,6 @@ import os
 import unittest
 import json
 import random
-import shutil
-import pytest
 
 from transformers import XLNetConfig, is_tf_available
 
@@ -30,18 +28,21 @@ if is_tf_available():
 
     from transformers.modeling_tf_xlnet import (TFXLNetModel, TFXLNetLMHeadModel,
                                                         TFXLNetForSequenceClassification,
+                                                        TFXLNetForTokenClassification,
                                                         TFXLNetForQuestionAnsweringSimple,
                                                         TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
-else:
-    pytestmark = pytest.mark.skip("Require TensorFlow")
 
 from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import CACHE_DIR, require_tf, slow
 
+
+@require_tf
 class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
 
     all_model_classes=(TFXLNetModel, TFXLNetLMHeadModel,
                        TFXLNetForSequenceClassification,
+                       TFXLNetForTokenClassification,
                        TFXLNetForQuestionAnsweringSimple) if is_tf_available() else ()
     test_pruning = False
 
@@ -62,7 +63,6 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                      num_attention_heads=4,
                      d_inner=128,
                      num_hidden_layers=5,
-                     max_position_embeddings=10,
                      type_sequence_label_size=2,
                      untie_r=True,
                      bi_data=False,
@@ -86,7 +86,6 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
             self.num_attention_heads = num_attention_heads
             self.d_inner = d_inner
             self.num_hidden_layers = num_hidden_layers
-            self.max_position_embeddings = max_position_embeddings
             self.bi_data = bi_data
             self.untie_r = untie_r
             self.same_length = same_length
@@ -120,13 +119,12 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                 is_impossible_labels = ids_tensor([self.batch_size], 2, dtype=tf.float32)
 
             config = XLNetConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 d_model=self.hidden_size,
                 n_head=self.num_attention_heads,
                 d_inner=self.d_inner,
                 n_layer=self.num_hidden_layers,
                 untie_r=self.untie_r,
-                max_position_embeddings=self.max_position_embeddings,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 same_length=self.same_length,
@@ -258,6 +256,26 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                 list(list(mem.shape) for mem in result["mems_1"]),
                 [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
 
+        def create_and_check_xlnet_for_token_classification(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
+                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels):
+            config.num_labels = input_ids_1.shape[1]
+            model = TFXLNetForTokenClassification(config)
+            inputs = {'input_ids': input_ids_1,
+                      'attention_mask': input_mask,
+                      # 'token_type_ids': token_type_ids
+                      }
+            logits, mems_1 = model(inputs)
+            result = {
+                "mems_1": [mem.numpy() for mem in mems_1],
+                "logits": logits.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].shape),
+                [self.batch_size, self.seq_length, config.num_labels])
+            self.parent.assertListEqual(
+                list(list(mem.shape) for mem in result["mems_1"]),
+                [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
+
         def prepare_config_and_inputs_for_common(self):
             config_and_inputs = self.prepare_config_and_inputs()
             (config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
@@ -282,24 +300,26 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
     def test_xlnet_lm_head(self):
         self.model_tester.set_seed()
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
-        self.model_tester.create_and_check_xlnet_lm_head(*config_and_inputs) 
+        self.model_tester.create_and_check_xlnet_lm_head(*config_and_inputs)
 
     def test_xlnet_sequence_classif(self):
         self.model_tester.set_seed()
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlnet_sequence_classif(*config_and_inputs)
 
+    def test_xlnet_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xlnet_for_token_classification(*config_and_inputs)
+
     def test_xlnet_qa(self):
         self.model_tester.set_seed()
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlnet_qa(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TFXLNetModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TFXLNetModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_transfo_xl_test.py

@@ -18,8 +18,6 @@ from __future__ import print_function
 
 import unittest
 import random
-import shutil
-import pytest
 
 from transformers import is_torch_available
 
@@ -27,12 +25,13 @@ if is_torch_available():
     import torch
     from transformers import (TransfoXLConfig, TransfoXLModel, TransfoXLLMHeadModel)
     from transformers.modeling_transfo_xl import TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
+
+@require_torch
 class TransfoXLModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (TransfoXLModel, TransfoXLLMHeadModel) if is_torch_available() else ()
@@ -66,7 +65,7 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
             self.batch_size = batch_size
             self.seq_length = seq_length
             self.mem_len = mem_len
-            self.key_len = seq_length + mem_len
+            self.key_length = seq_length + mem_len
             self.clamp_len = clamp_len
             self.is_training = is_training
             self.use_labels = use_labels
@@ -91,7 +90,7 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
                 lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
             config = TransfoXLConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 cutoffs=self.cutoffs,
@@ -111,6 +110,7 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
 
         def create_transfo_xl_model(self, config, input_ids_1, input_ids_2, lm_labels):
             model = TransfoXLModel(config)
+            model.to(torch_device)
             model.eval()
 
             hidden_states_1, mems_1 = model(input_ids_1)
@@ -140,6 +140,7 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
 
         def create_transfo_xl_lm_head(self, config, input_ids_1, input_ids_2, lm_labels):
             model = TransfoXLLMHeadModel(config)
+            model.to(torch_device)
             model.eval()
 
             lm_logits_1, mems_1 = model(input_ids_1)
@@ -204,12 +205,10 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
         output_result = self.model_tester.create_transfo_xl_lm_head(*config_and_inputs)
         self.model_tester.check_transfo_xl_lm_head_output(output_result)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = TransfoXLModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = TransfoXLModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_xlm_test.py

@@ -17,8 +17,6 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
-import shutil
-import pytest
 
 from transformers import is_torch_available
 
@@ -26,13 +24,13 @@ if is_torch_available():
     from transformers import (XLMConfig, XLMModel, XLMWithLMHeadModel, XLMForQuestionAnswering,
                                       XLMForSequenceClassification, XLMForQuestionAnsweringSimple)
     from transformers.modeling_xlm import XLM_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
 
+@require_torch
 class XLMModelTest(CommonTestCases.CommonModelTester):
 
     all_model_classes = (XLMModel, XLMWithLMHeadModel, XLMForQuestionAnswering,
@@ -122,7 +120,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
                 is_impossible_labels = ids_tensor([self.batch_size], 2).float()
 
             config = XLMConfig(
-                 vocab_size_or_config_json_file=self.vocab_size,
+                 vocab_size=self.vocab_size,
                  n_special=self.n_special,
                  emb_dim=self.hidden_size,
                  n_layers=self.num_hidden_layers,
@@ -148,6 +146,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xlm_model(self, config, input_ids, token_type_ids, input_lengths, sequence_labels, token_labels, is_impossible_labels, input_mask):
             model = XLMModel(config=config)
+            model.to(torch_device)
             model.eval()
             outputs = model(input_ids, lengths=input_lengths, langs=token_type_ids)
             outputs = model(input_ids, langs=token_type_ids)
@@ -163,6 +162,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xlm_lm_head(self, config, input_ids, token_type_ids, input_lengths, sequence_labels, token_labels, is_impossible_labels, input_mask):
             model = XLMWithLMHeadModel(config)
+            model.to(torch_device)
             model.eval()
 
             loss, logits = model(input_ids, token_type_ids=token_type_ids, labels=token_labels)
@@ -182,6 +182,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xlm_simple_qa(self, config, input_ids, token_type_ids, input_lengths, sequence_labels, token_labels, is_impossible_labels, input_mask):
             model = XLMForQuestionAnsweringSimple(config)
+            model.to(torch_device)
             model.eval()
 
             outputs = model(input_ids)
@@ -206,6 +207,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xlm_qa(self, config, input_ids, token_type_ids, input_lengths, sequence_labels, token_labels, is_impossible_labels, input_mask):
             model = XLMForQuestionAnswering(config)
+            model.to(torch_device)
             model.eval()
 
             outputs = model(input_ids)
@@ -260,6 +262,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
 
         def create_and_check_xlm_sequence_classif(self, config, input_ids, token_type_ids, input_lengths, sequence_labels, token_labels, is_impossible_labels, input_mask):
             model = XLMForSequenceClassification(config)
+            model.to(torch_device)
             model.eval()
 
             (logits,) = model(input_ids)
@@ -312,12 +315,10 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlm_sequence_classif(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(XLM_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XLMModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = XLMModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/modeling_xlnet_test.py

@@ -20,25 +20,25 @@ import os
 import unittest
 import json
 import random
-import shutil
-import pytest
 
 from transformers import is_torch_available
 
 if is_torch_available():
     import torch
 
-    from transformers import (XLNetConfig, XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering)
+    from transformers import (XLNetConfig, XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification,
+                              XLNetForTokenClassification, XLNetForQuestionAnswering)
     from transformers.modeling_xlnet import XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
-else:
-    pytestmark = pytest.mark.skip("Require Torch")
 
 from .modeling_common_test import (CommonTestCases, ids_tensor)
 from .configuration_common_test import ConfigTester
+from .utils import CACHE_DIR, require_torch, slow, torch_device
 
+
+@require_torch
 class XLNetModelTest(CommonTestCases.CommonModelTester):
 
-    all_model_classes=(XLNetModel, XLNetLMHeadModel,
+    all_model_classes=(XLNetModel, XLNetLMHeadModel, XLNetForTokenClassification,
                     XLNetForSequenceClassification, XLNetForQuestionAnswering) if is_torch_available() else ()
     test_pruning = False
 
@@ -59,7 +59,6 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                      num_attention_heads=4,
                      d_inner=128,
                      num_hidden_layers=5,
-                     max_position_embeddings=10,
                      type_sequence_label_size=2,
                      untie_r=True,
                      bi_data=False,
@@ -83,7 +82,6 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
             self.num_attention_heads = num_attention_heads
             self.d_inner = d_inner
             self.num_hidden_layers = num_hidden_layers
-            self.max_position_embeddings = max_position_embeddings
             self.bi_data = bi_data
             self.untie_r = untie_r
             self.same_length = same_length
@@ -99,27 +97,28 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
             input_mask = ids_tensor([self.batch_size, self.seq_length], 2).float()
 
             input_ids_q = ids_tensor([self.batch_size, self.seq_length + 1], self.vocab_size)
-            perm_mask = torch.zeros(self.batch_size, self.seq_length + 1, self.seq_length + 1, dtype=torch.float)
+            perm_mask = torch.zeros(self.batch_size, self.seq_length + 1, self.seq_length + 1, dtype=torch.float, device=torch_device)
             perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
-            target_mapping = torch.zeros(self.batch_size, 1, self.seq_length + 1, dtype=torch.float)
+            target_mapping = torch.zeros(self.batch_size, 1, self.seq_length + 1, dtype=torch.float, device=torch_device)
             target_mapping[:, 0, -1] = 1.0  # predict last token
 
             sequence_labels = None
             lm_labels = None
             is_impossible_labels = None
+            token_labels = None
             if self.use_labels:
                 lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
                 sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
                 is_impossible_labels = ids_tensor([self.batch_size], 2).float()
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
 
             config = XLNetConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 d_model=self.hidden_size,
                 n_head=self.num_attention_heads,
                 d_inner=self.d_inner,
                 n_layer=self.num_hidden_layers,
                 untie_r=self.untie_r,
-                max_position_embeddings=self.max_position_embeddings,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 same_length=self.same_length,
@@ -129,15 +128,16 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 num_labels=self.type_sequence_label_size)
 
             return (config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
-                    target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels)
+                    target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels)
 
         def set_seed(self):
             random.seed(self.seed)
             torch.manual_seed(self.seed)
 
         def create_and_check_xlnet_base_model(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
-                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels):
+                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels):
             model = XLNetModel(config)
+            model.to(torch_device)
             model.eval()
 
             _, _ = model(input_ids_1, input_mask=input_mask)
@@ -152,6 +152,7 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
 
             config.mem_len = 0
             model = XLNetModel(config)
+            model.to(torch_device)
             model.eval()
             no_mems_outputs = model(input_ids_1)
             self.parent.assertEqual(len(no_mems_outputs), 1)
@@ -163,9 +164,23 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 list(list(mem.size()) for mem in result["mems_1"]),
                 [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
 
+        def create_and_check_xlnet_base_model_with_att_output(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
+                    target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels):
+            model = XLNetModel(config)
+            model.to(torch_device)
+            model.eval()
+
+            _, _, attentions = model(input_ids_1, target_mapping=target_mapping)
+
+            self.parent.assertEqual(len(attentions), config.n_layer)
+            self.parent.assertIsInstance(attentions[0], tuple)
+            self.parent.assertEqual(len(attentions[0]), 2)
+            self.parent.assertTrue(attentions[0][0].shape, attentions[0][0].shape)
+
         def create_and_check_xlnet_lm_head(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
-                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels):
+                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels):
             model = XLNetLMHeadModel(config)
+            model.to(torch_device)
             model.eval()
 
             loss_1, all_logits_1, mems_1 = model(input_ids_1, token_type_ids=segment_ids, labels=lm_labels)
@@ -204,8 +219,9 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 [[self.mem_len, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
 
         def create_and_check_xlnet_qa(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
-                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels):
+                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels):
             model = XLNetForQuestionAnswering(config)
+            model.to(torch_device)
             model.eval()
 
             outputs = model(input_ids_1)
@@ -261,9 +277,43 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 list(list(mem.size()) for mem in result["mems"]),
                 [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
 
+        def create_and_check_xlnet_token_classif(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
+                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels):
+            model = XLNetForTokenClassification(config)
+            model.to(torch_device)
+            model.eval()
+
+            logits, mems_1 = model(input_ids_1)
+            loss, logits, mems_1 = model(input_ids_1, labels=token_labels)
+
+            result = {
+                "loss": loss,
+                "mems_1": mems_1,
+                "logits": logits,
+            }
+
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.seq_length, self.type_sequence_label_size])
+            self.parent.assertListEqual(
+                list(list(mem.size()) for mem in result["mems_1"]),
+                [[self.seq_length, self.batch_size, self.hidden_size]] * self.num_hidden_layers)
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
+                target_mapping, segment_ids, lm_labels,
+                sequence_labels, is_impossible_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids_1}
+            return config, inputs_dict
+
         def create_and_check_xlnet_sequence_classif(self, config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
-                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels):
+                target_mapping, segment_ids, lm_labels, sequence_labels, is_impossible_labels, token_labels):
             model = XLNetForSequenceClassification(config)
+            model.to(torch_device)
             model.eval()
 
             logits, mems_1 = model(input_ids_1)
@@ -289,7 +339,7 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
             config_and_inputs = self.prepare_config_and_inputs()
             (config, input_ids_1, input_ids_2, input_ids_q, perm_mask, input_mask,
                 target_mapping, segment_ids, lm_labels,
-                sequence_labels, is_impossible_labels) = config_and_inputs
+                sequence_labels, is_impossible_labels, token_labels) = config_and_inputs
             inputs_dict = {'input_ids': input_ids_1}
             return config, inputs_dict
 
@@ -306,27 +356,36 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlnet_base_model(*config_and_inputs)
 
+    def test_xlnet_base_model_with_att_output(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        config_and_inputs[0].output_attentions = True
+        self.model_tester.create_and_check_xlnet_base_model_with_att_output(*config_and_inputs)
+
     def test_xlnet_lm_head(self):
         self.model_tester.set_seed()
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
-        self.model_tester.create_and_check_xlnet_lm_head(*config_and_inputs) 
+        self.model_tester.create_and_check_xlnet_lm_head(*config_and_inputs)
 
     def test_xlnet_sequence_classif(self):
         self.model_tester.set_seed()
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlnet_sequence_classif(*config_and_inputs)
 
+    def test_xlnet_token_classif(self):
+        self.model_tester.set_seed()
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_xlnet_token_classif(*config_and_inputs)
+
     def test_xlnet_qa(self):
         self.model_tester.set_seed()
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_xlnet_qa(*config_and_inputs)
 
-    @pytest.mark.slow
+    @slow
     def test_model_from_pretrained(self):
-        cache_dir = "/tmp/transformers_test/"
         for model_name in list(XLNET_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
-            model = XLNetModel.from_pretrained(model_name, cache_dir=cache_dir)
-            shutil.rmtree(cache_dir)
+            model = XLNetModel.from_pretrained(model_name, cache_dir=CACHE_DIR)
             self.assertIsNotNone(model)
 
 

transformers/tests/optimization_test.py

@@ -18,7 +18,6 @@ from __future__ import print_function
 
 import unittest
 import os
-import pytest
 
 from transformers import is_torch_available
 
@@ -31,10 +30,9 @@ if is_torch_available():
                               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")
 
 from .tokenization_tests_commons import TemporaryDirectory
+from .utils import require_torch
 
 
 def unwrap_schedule(scheduler, num_steps=10):
@@ -58,6 +56,7 @@ def unwrap_and_save_reload_schedule(scheduler, num_steps=10):
                 scheduler.load_state_dict(state_dict)
     return lrs
 
+@require_torch
 class OptimizationTest(unittest.TestCase):
 
     def assertListAlmostEqual(self, list1, list2, tol):
@@ -80,6 +79,7 @@ class OptimizationTest(unittest.TestCase):
         self.assertListAlmostEqual(w.tolist(), [0.4, 0.2, -0.5], tol=1e-2)
 
 
+@require_torch
 class ScheduleInitTest(unittest.TestCase):
     m = torch.nn.Linear(50, 50) if is_torch_available() else None
     optimizer = AdamW(m.parameters(), lr=10.) if is_torch_available() else None

transformers/tests/optimization_tf_test.py

+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+
+from transformers import is_tf_available
+
+from .utils import require_tf
+
+if is_tf_available():
+    import tensorflow as tf
+    from tensorflow.python.eager import context
+    from tensorflow.python.framework import ops
+    from transformers import (create_optimizer, GradientAccumulator)
+
+
+@require_tf
+class OptimizationFTest(unittest.TestCase):
+    def assertListAlmostEqual(self, list1, list2, tol):
+        self.assertEqual(len(list1), len(list2))
+        for a, b in zip(list1, list2):
+            self.assertAlmostEqual(a, b, delta=tol)
+    
+    def testGradientAccumulator(self):
+        accumulator = GradientAccumulator()
+        accumulator([tf.constant([1.0, 2.0])])
+        accumulator([tf.constant([-2.0, 1.0])])
+        accumulator([tf.constant([-1.0, 2.0])])
+        with self.assertRaises(ValueError):
+            accumulator([tf.constant([1.0, 1.0]), tf.constant([2.0, 2.0])])
+        self.assertEqual(accumulator.step, 3)
+        self.assertEqual(len(accumulator.gradients), 1)
+        self.assertListAlmostEqual(accumulator.gradients[0].numpy().tolist(), [-2.0, 5.0], tol=1e-2)
+        accumulator.reset()
+        self.assertEqual(accumulator.step, 0)
+        self.assertListAlmostEqual(accumulator.gradients[0].numpy().tolist(), [0.0, 0.0], tol=1e-2)
+
+    def testGradientAccumulatorDistributionStrategy(self):
+        context._context = None
+        ops.enable_eager_execution_internal()
+        physical_devices = tf.config.experimental.list_physical_devices("CPU")
+        tf.config.experimental.set_virtual_device_configuration(
+            physical_devices[0],
+            [tf.config.experimental.VirtualDeviceConfiguration(),
+            tf.config.experimental.VirtualDeviceConfiguration()])
+
+        devices = tf.config.experimental.list_logical_devices(device_type="CPU")
+        strategy = tf.distribute.MirroredStrategy(devices=[device.name for device in devices])
+
+        with strategy.scope():
+            accumulator = GradientAccumulator()
+            variable = tf.Variable([4.0, 3.0])
+            optimizer = create_optimizer(5e-5, 10, 5)
+            gradient_placeholder = tf.Variable([0.0, 0.0], trainable=False)
+
+        def accumulate_on_replica(gradient):
+            accumulator([gradient])
+
+        def apply_on_replica():
+            optimizer.apply_gradients(list(zip(accumulator.gradients, [variable])), 1.0)
+
+        @tf.function
+        def accumulate(grad1, grad2):
+            with strategy.scope():
+                gradient_placeholder.values[0].assign(grad1)
+                gradient_placeholder.values[1].assign(grad2)
+                strategy.experimental_run_v2(accumulate_on_replica, args=(gradient_placeholder,))
+
+        @tf.function
+        def apply_grad():
+            with strategy.scope():
+                strategy.experimental_run_v2(apply_on_replica)
+
+        accumulate([1.0, 2.0], [-1.0, 1.0])
+        accumulate([3.0, -1.0], [-1.0, -1.0])
+        accumulate([-2.0, 2.0], [3.0, -2.0])
+        self.assertEqual(accumulator.step, 3)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[0].value().numpy().tolist(), [2.0, 3.0], tol=1e-2)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[1].value().numpy().tolist(), [1.0, -2.0], tol=1e-2)
+        apply_grad()
+        self.assertListAlmostEqual(variable.value().numpy().tolist(), [4.0, 3.0], tol=1e-2)
+        accumulator.reset()
+        self.assertEqual(accumulator.step, 0)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[0].value().numpy().tolist(), [0.0, 0.0], tol=1e-2)
+        self.assertListAlmostEqual(accumulator._gradients[0].values[1].value().numpy().tolist(), [0.0, 0.0], tol=1e-2)
+
+
+if __name__ == "__main__":
+    unittest.main()
\ No newline at end of file

transformers/tests/pipelines_test.py

+import unittest
+
+from typing import Iterable
+
+from transformers import pipeline
+from transformers.tests.utils import require_tf, require_torch
+
+QA_FINETUNED_MODELS = {
+    ('bert-base-uncased', 'bert-large-uncased-whole-word-masking-finetuned-squad', None),
+    ('bert-base-cased', 'bert-large-cased-whole-word-masking-finetuned-squad', None),
+    ('bert-base-uncased', 'distilbert-base-uncased-distilled-squad', None)
+}
+
+TF_QA_FINETUNED_MODELS = {
+    ('bert-base-uncased', 'bert-large-uncased-whole-word-masking-finetuned-squad', None),
+    ('bert-base-cased', 'bert-large-cased-whole-word-masking-finetuned-squad', None),
+    ('bert-base-uncased', 'distilbert-base-uncased-distilled-squad', None)
+}
+
+TF_NER_FINETUNED_MODELS = {
+    (
+        'bert-base-cased',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-tf_model.h5',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-config.json'
+    )
+}
+
+NER_FINETUNED_MODELS = {
+    (
+        'bert-base-cased',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-pytorch_model.bin',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-finetuned-conll03-english-config.json'
+    )
+}
+
+FEATURE_EXTRACT_FINETUNED_MODELS = {
+   ('bert-base-cased', 'bert-base-cased', None),
+   # ('xlnet-base-cased', 'xlnet-base-cased', None), # Disabled for now as it crash for TF2
+   ('distilbert-base-uncased', 'distilbert-base-uncased', None)
+}
+
+TF_FEATURE_EXTRACT_FINETUNED_MODELS = {
+   ('bert-base-cased', 'bert-base-cased', None),
+   # ('xlnet-base-cased', 'xlnet-base-cased', None), # Disabled for now as it crash for TF2
+   ('distilbert-base-uncased', 'distilbert-base-uncased', None)
+}
+
+TF_TEXT_CLASSIF_FINETUNED_MODELS = {
+    (
+        'bert-base-uncased',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-tf_model.h5',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-config.json'
+    )
+}
+
+TEXT_CLASSIF_FINETUNED_MODELS = {
+    (
+        'bert-base-uncased',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-pytorch_model.bin',
+        'https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-finetuned-sst-2-english-config.json'
+    )
+}
+
+
+class MonoColumnInputTestCase(unittest.TestCase):
+    def _test_mono_column_pipeline(self, nlp, valid_inputs: list, invalid_inputs: list, output_keys: Iterable[str]):
+        self.assertIsNotNone(nlp)
+
+        mono_result = nlp(valid_inputs[0])
+        self.assertIsInstance(mono_result, list)
+        self.assertIsInstance(mono_result[0], (dict, list))
+
+        if isinstance(mono_result[0], list):
+            mono_result = mono_result[0]
+
+        for key in output_keys:
+            self.assertIn(key, mono_result[0])
+
+        multi_result = nlp(valid_inputs)
+        self.assertIsInstance(multi_result, list)
+        self.assertIsInstance(multi_result[0], (dict, list))
+
+        if isinstance(multi_result[0], list):
+            multi_result = multi_result[0]
+
+        for result in multi_result:
+            for key in output_keys:
+                self.assertIn(key, result)
+
+        self.assertRaises(Exception, nlp, invalid_inputs)
+
+    @require_torch
+    def test_ner(self):
+        mandatory_keys = {'entity', 'word', 'score'}
+        valid_inputs = ['HuggingFace is solving NLP one commit at a time.', 'HuggingFace is based in New-York & Paris']
+        invalid_inputs = [None]
+        for tokenizer, model, config in NER_FINETUNED_MODELS:
+            nlp = pipeline(task='ner', model=model, config=config, tokenizer=tokenizer)
+            self._test_mono_column_pipeline(nlp, valid_inputs, invalid_inputs, mandatory_keys)
+
+    @require_tf
+    def test_tf_ner(self):
+        mandatory_keys = {'entity', 'word', 'score'}
+        valid_inputs = ['HuggingFace is solving NLP one commit at a time.', 'HuggingFace is based in New-York & Paris']
+        invalid_inputs = [None]
+        for tokenizer, model, config in TF_NER_FINETUNED_MODELS:
+            nlp = pipeline(task='ner', model=model, config=config, tokenizer=tokenizer)
+            self._test_mono_column_pipeline(nlp, valid_inputs, invalid_inputs, mandatory_keys)
+
+    @require_torch
+    def test_sentiment_analysis(self):
+        mandatory_keys = {'label'}
+        valid_inputs = ['HuggingFace is solving NLP one commit at a time.', 'HuggingFace is based in New-York & Paris']
+        invalid_inputs = [None]
+        for tokenizer, model, config in TEXT_CLASSIF_FINETUNED_MODELS:
+            nlp = pipeline(task='sentiment-analysis', model=model, config=config, tokenizer=tokenizer)
+            self._test_mono_column_pipeline(nlp, valid_inputs, invalid_inputs, mandatory_keys)
+
+    @require_tf
+    def test_tf_sentiment_analysis(self):
+        mandatory_keys = {'label'}
+        valid_inputs = ['HuggingFace is solving NLP one commit at a time.', 'HuggingFace is based in New-York & Paris']
+        invalid_inputs = [None]
+        for tokenizer, model, config in TF_TEXT_CLASSIF_FINETUNED_MODELS:
+            nlp = pipeline(task='sentiment-analysis', model=model, config=config, tokenizer=tokenizer)
+            self._test_mono_column_pipeline(nlp, valid_inputs, invalid_inputs, mandatory_keys)
+
+    @require_torch
+    def test_features_extraction(self):
+        valid_inputs = ['HuggingFace is solving NLP one commit at a time.', 'HuggingFace is based in New-York & Paris']
+        invalid_inputs = [None]
+        for tokenizer, model, config in FEATURE_EXTRACT_FINETUNED_MODELS:
+            nlp = pipeline(task='sentiment-analysis', model=model, config=config, tokenizer=tokenizer)
+            self._test_mono_column_pipeline(nlp, valid_inputs, invalid_inputs, {})
+
+    @require_tf
+    def test_tf_features_extraction(self):
+        valid_inputs = ['HuggingFace is solving NLP one commit at a time.', 'HuggingFace is based in New-York & Paris']
+        invalid_inputs = [None]
+        for tokenizer, model, config in TF_FEATURE_EXTRACT_FINETUNED_MODELS:
+            nlp = pipeline(task='sentiment-analysis', model=model, config=config, tokenizer=tokenizer)
+            self._test_mono_column_pipeline(nlp, valid_inputs, invalid_inputs, {})
+
+
+class MultiColumnInputTestCase(unittest.TestCase):
+    def _test_multicolumn_pipeline(self, nlp, valid_inputs: list, invalid_inputs: list, output_keys: Iterable[str]):
+        self.assertIsNotNone(nlp)
+
+        mono_result = nlp(valid_inputs[0])
+        self.assertIsInstance(mono_result, dict)
+
+        for key in output_keys:
+            self.assertIn(key, mono_result)
+
+        multi_result = nlp(valid_inputs)
+        self.assertIsInstance(multi_result, list)
+        self.assertIsInstance(multi_result[0], dict)
+
+        for result in multi_result:
+            for key in output_keys:
+                self.assertIn(key, result)
+
+        self.assertRaises(Exception, nlp, invalid_inputs[0])
+        self.assertRaises(Exception, nlp, invalid_inputs)
+
+    @require_torch
+    def test_question_answering(self):
+        mandatory_output_keys = {'score', 'answer', 'start', 'end'}
+        valid_samples = [
+            {'question': 'Where was HuggingFace founded ?', 'context': 'HuggingFace was founded in Paris.'},
+            {
+                'question': 'In what field is HuggingFace working ?',
+                'context': 'HuggingFace is a startup based in New-York founded in Paris which is trying to solve NLP.'
+            }
+        ]
+        invalid_samples = [
+            {'question': '', 'context': 'This is a test to try empty question edge case'},
+            {'question': None, 'context': 'This is a test to try empty question edge case'},
+            {'question': 'What is does with empty context ?', 'context': ''},
+            {'question': 'What is does with empty context ?', 'context': None},
+        ]
+
+        for tokenizer, model, config in QA_FINETUNED_MODELS:
+            nlp = pipeline(task='question-answering', model=model, config=config, tokenizer=tokenizer)
+            self._test_multicolumn_pipeline(nlp, valid_samples, invalid_samples, mandatory_output_keys)
+
+    @require_tf
+    def test_tf_question_answering(self):
+        mandatory_output_keys = {'score', 'answer', 'start', 'end'}
+        valid_samples = [
+            {'question': 'Where was HuggingFace founded ?', 'context': 'HuggingFace was founded in Paris.'},
+            {
+                'question': 'In what field is HuggingFace working ?',
+                'context': 'HuggingFace is a startup based in New-York founded in Paris which is trying to solve NLP.'
+            }
+        ]
+        invalid_samples = [
+            {'question': '', 'context': 'This is a test to try empty question edge case'},
+            {'question': None, 'context': 'This is a test to try empty question edge case'},
+            {'question': 'What is does with empty context ?', 'context': ''},
+            {'question': 'What is does with empty context ?', 'context': None},
+        ]
+
+        for tokenizer, model, config in TF_QA_FINETUNED_MODELS:
+            nlp = pipeline(task='question-answering', model=model, config=config, tokenizer=tokenizer)
+            self._test_multicolumn_pipeline(nlp, valid_samples, invalid_samples, mandatory_output_keys)
+
+
+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_auto_test.py

@@ -18,15 +18,16 @@ from __future__ import print_function
 
 import unittest
 import shutil
-import pytest
 import logging
 
 from transformers import AutoTokenizer, BertTokenizer, AutoTokenizer, GPT2Tokenizer
 from transformers import BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
 
+from .utils import slow, SMALL_MODEL_IDENTIFIER
+
 
 class AutoTokenizerTest(unittest.TestCase):
-    @pytest.mark.slow
+    @slow
     def test_tokenizer_from_pretrained(self):
         logging.basicConfig(level=logging.INFO)
         for model_name in list(BERT_PRETRAINED_CONFIG_ARCHIVE_MAP.keys())[:1]:
@@ -41,6 +42,11 @@ class AutoTokenizerTest(unittest.TestCase):
             self.assertIsInstance(tokenizer, GPT2Tokenizer)
             self.assertGreater(len(tokenizer), 0)
 
+    def test_tokenizer_from_pretrained_identifier(self):
+        logging.basicConfig(level=logging.INFO)
+        tokenizer = AutoTokenizer.from_pretrained(SMALL_MODEL_IDENTIFIER)
+        self.assertIsInstance(tokenizer, BertTokenizer)
+        self.assertEqual(len(tokenizer), 12)
 
 if __name__ == "__main__":
     unittest.main()

transformers/tests/tokenization_bert_japanese_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, division, print_function, unicode_literals
+
+import os
+import unittest
+from io import open
+
+from transformers.tokenization_bert import WordpieceTokenizer
+from transformers.tokenization_bert_japanese import (BertJapaneseTokenizer,
+                                                     MecabTokenizer, CharacterTokenizer,
+                                                     VOCAB_FILES_NAMES)
+
+from .tokenization_tests_commons import CommonTestCases
+from .utils import slow, custom_tokenizers
+
+
+@custom_tokenizers
+class BertJapaneseTokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = BertJapaneseTokenizer
+
+    def setUp(self):
+        super(BertJapaneseTokenizationTest, self).setUp()
+
+        vocab_tokens = [u"[UNK]", u"[CLS]", u"[SEP]",
+            u"こんにちは", u"こん", u"にちは", u"ばんは", u"##こん", u"##にちは", u"##ばんは",
+            u"世界", u"##世界", u"、", u"##、", u"。", u"##。"]
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+    def get_tokenizer(self, **kwargs):
+        return BertJapaneseTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self):
+        input_text = u"こんにちは、世界。 \nこんばんは、世界。"
+        output_text = u"こんにちは 、 世界 。 こんばんは 、 世界 。"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file)
+
+        tokens = tokenizer.tokenize(u"こんにちは、世界。\nこんばんは、世界。")
+        self.assertListEqual(tokens,
+                             [u"こんにちは", u"、", u"世界", u"。",
+                              u"こん", u"##ばんは", u"、", u"世界", "。"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens),
+                             [3, 12, 10, 14, 4, 9, 12, 10, 14])
+
+    def test_mecab_tokenizer(self):
+        tokenizer = MecabTokenizer()
+
+        self.assertListEqual(
+            tokenizer.tokenize(u" \tｱｯﾌﾟﾙストアでiPhone８ が  \n 発売された　。  "),
+                               [u"アップルストア", u"で", u"iPhone", u"8", u"が",
+                                u"発売", u"さ", u"れ", u"た", u"。"])
+
+    def test_mecab_tokenizer_lower(self):
+        tokenizer = MecabTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(u" \tｱｯﾌﾟﾙストアでiPhone８ が  \n 発売された　。  "),
+                               [u"アップルストア", u"で", u"iphone", u"8", u"が",
+                                u"発売", u"さ", u"れ", u"た", u"。"])
+
+    def test_mecab_tokenizer_no_normalize(self):
+        tokenizer = MecabTokenizer(normalize_text=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(u" \tｱｯﾌﾟﾙストアでiPhone８ が  \n 発売された　。  "),
+                               [u"ｱｯﾌﾟﾙストア", u"で", u"iPhone", u"８", u"が",
+                                u"発売", u"さ", u"れ", u"た", u"　", u"。"])
+
+    def test_wordpiece_tokenizer(self):
+        vocab_tokens = [u"[UNK]", u"[CLS]", u"[SEP]",
+            u"こんにちは", u"こん", u"にちは" u"ばんは", u"##こん", u"##にちは", u"##ばんは"]
+
+        vocab = {}
+        for (i, token) in enumerate(vocab_tokens):
+            vocab[token] = i
+        tokenizer = WordpieceTokenizer(vocab=vocab, unk_token=u"[UNK]")
+
+        self.assertListEqual(tokenizer.tokenize(u""), [])
+
+        self.assertListEqual(tokenizer.tokenize(u"こんにちは"),
+                             [u"こんにちは"])
+
+        self.assertListEqual(tokenizer.tokenize(u"こんばんは"),
+                             [u"こん", u"##ばんは"])
+
+        self.assertListEqual(tokenizer.tokenize(u"こんばんは こんばんにちは こんにちは"),
+                             [u"こん", u"##ばんは", u"[UNK]", u"こんにちは"])
+
+    @slow
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class.from_pretrained("bert-base-japanese")
+
+        text = tokenizer.encode(u"ありがとう。", add_special_tokens=False)
+        text_2 = tokenizer.encode(u"どういたしまして。", add_special_tokens=False)
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        # 2 is for "[CLS]", 3 is for "[SEP]"
+        assert encoded_sentence == [2] + text + [3]
+        assert encoded_pair == [2] + text + [3] + text_2 + [3]
+
+
+class BertJapaneseCharacterTokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = BertJapaneseTokenizer
+
+    def setUp(self):
+        super(BertJapaneseCharacterTokenizationTest, self).setUp()
+
+        vocab_tokens = [u"[UNK]", u"[CLS]", u"[SEP]",
+            u"こ", u"ん", u"に", u"ち", u"は", u"ば", u"世", u"界", u"、", u"。"]
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+    def get_tokenizer(self, **kwargs):
+        return BertJapaneseTokenizer.from_pretrained(self.tmpdirname,
+                                                     subword_tokenizer_type="character",
+                                                     **kwargs)
+
+    def get_input_output_texts(self):
+        input_text = u"こんにちは、世界。 \nこんばんは、世界。"
+        output_text = u"こ ん に ち は 、 世 界 。 こ ん ば ん は 、 世 界 。"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file,
+                                         subword_tokenizer_type="character")
+
+        tokens = tokenizer.tokenize(u"こんにちは、世界。 \nこんばんは、世界。")
+        self.assertListEqual(tokens,
+            [u"こ", u"ん", u"に", u"ち", u"は", u"、", u"世", u"界", u"。",
+             u"こ", u"ん", u"ば", u"ん", u"は", u"、", u"世", u"界", u"。"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens),
+                             [3, 4, 5, 6, 7, 11, 9, 10, 12,
+                              3, 4, 8, 4, 7, 11, 9, 10, 12])
+
+    def test_character_tokenizer(self):
+        vocab_tokens = [u"[UNK]", u"[CLS]", u"[SEP]",
+            u"こ", u"ん", u"に", u"ち", u"は", u"ば", u"世", u"界"u"、", u"。"]
+
+        vocab = {}
+        for (i, token) in enumerate(vocab_tokens):
+            vocab[token] = i
+        tokenizer = CharacterTokenizer(vocab=vocab, unk_token=u"[UNK]")
+
+        self.assertListEqual(tokenizer.tokenize(u""), [])
+
+        self.assertListEqual(tokenizer.tokenize(u"こんにちは"),
+                             [u"こ", u"ん", u"に", u"ち", u"は"])
+
+        self.assertListEqual(tokenizer.tokenize(u"こんにちほ"),
+                             [u"こ", u"ん", u"に", u"ち", u"[UNK]"])
+
+    @slow
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class.from_pretrained("bert-base-japanese-char")
+
+        text = tokenizer.encode(u"ありがとう。", add_special_tokens=False)
+        text_2 = tokenizer.encode(u"どういたしまして。", add_special_tokens=False)
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        # 2 is for "[CLS]", 3 is for "[SEP]"
+        assert encoded_sentence == [2] + text + [3]
+        assert encoded_pair == [2] + text + [3] + text_2 + [3]
+
+
+

transformers/tests/tokenization_bert_test.py

@@ -16,7 +16,6 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import os
 import unittest
-import pytest
 from io import open
 
 from transformers.tokenization_bert import (BasicTokenizer,
@@ -26,6 +25,7 @@ from transformers.tokenization_bert import (BasicTokenizer,
                                                     _is_whitespace, VOCAB_FILES_NAMES)
 
 from .tokenization_tests_commons import CommonTestCases
+from .utils import slow
 
 class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
 
@@ -126,7 +126,7 @@ class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
         self.assertFalse(_is_punctuation(u"A"))
         self.assertFalse(_is_punctuation(u" "))
 
-    @pytest.mark.slow
+    @slow
     def test_sequence_builders(self):
         tokenizer = self.tokenizer_class.from_pretrained("bert-base-uncased")
 
@@ -139,5 +139,6 @@ class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
         assert encoded_sentence == [101] + text + [102]
         assert encoded_pair == [101] + text + [102] + text_2 + [102]
 
+
 if __name__ == '__main__':
     unittest.main()

transformers/tests/tokenization_distilbert_test.py

@@ -16,13 +16,13 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import os
 import unittest
-import pytest
 from io import open
 
 from transformers.tokenization_distilbert import (DistilBertTokenizer)
 
 from .tokenization_tests_commons import CommonTestCases
 from .tokenization_bert_test import BertTokenizationTest
+from .utils import slow
 
 class DistilBertTokenizationTest(BertTokenizationTest):
 
@@ -31,7 +31,7 @@ class DistilBertTokenizationTest(BertTokenizationTest):
     def get_tokenizer(self, **kwargs):
         return DistilBertTokenizer.from_pretrained(self.tmpdirname, **kwargs)
 
-    @pytest.mark.slow
+    @slow
     def test_sequence_builders(self):
         tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
 

transformers/tests/tokenization_gpt2_test.py

@@ -67,6 +67,5 @@ class GPT2TokenizationTest(CommonTestCases.CommonTokenizerTester):
         self.assertListEqual(
             tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
 
-
 if __name__ == '__main__':
     unittest.main()

transformers/tests/tokenization_roberta_test.py

@@ -17,11 +17,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 import os
 import json
 import unittest
-import pytest
 from io import open
 
 from transformers.tokenization_roberta import RobertaTokenizer, VOCAB_FILES_NAMES
 from .tokenization_tests_commons import CommonTestCases
+from .utils import slow
 
 
 class RobertaTokenizationTest(CommonTestCases.CommonTokenizerTester):
@@ -79,7 +79,7 @@ class RobertaTokenizationTest(CommonTestCases.CommonTokenizerTester):
             [0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]
         )
 
-    @pytest.mark.slow
+    @slow
     def test_sequence_builders(self):
         tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
 

transformers/tests/tokenization_t5_test.py

+# coding=utf-8
+# Copyright 2018 Google T5 Authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+import unittest
+
+from transformers.tokenization_t5 import (T5Tokenizer)
+from transformers.tokenization_xlnet import SPIECE_UNDERLINE
+
+from .tokenization_tests_commons import CommonTestCases
+
+SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)),
+                    'fixtures/test_sentencepiece.model')
+
+class T5TokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = T5Tokenizer
+
+    def setUp(self):
+        super(T5TokenizationTest, self).setUp()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = T5Tokenizer(SAMPLE_VOCAB)
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return T5Tokenizer.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 = T5Tokenizer(SAMPLE_VOCAB)
+
+        tokens = tokenizer.tokenize(u'This is a test')
+        self.assertListEqual(tokens, [u'▁This', u'▁is', u'▁a', u'▁t', u'est'])
+
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(tokens), [285, 46, 10, 170, 382])
+
+        tokens = tokenizer.tokenize(u"I was born in 92000, and this is falsé.")
+        self.assertListEqual(tokens, [SPIECE_UNDERLINE + u'I', SPIECE_UNDERLINE + u'was', SPIECE_UNDERLINE + u'b',
+                                    u'or', u'n', SPIECE_UNDERLINE + u'in', SPIECE_UNDERLINE + u'',
+                                    u'9', u'2', u'0', u'0', u'0', u',', SPIECE_UNDERLINE + u'and', SPIECE_UNDERLINE + u'this',
+                                    SPIECE_UNDERLINE + u'is', SPIECE_UNDERLINE + u'f', u'al', u's', u'é', u'.'])
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(
+            ids, [8, 21, 84, 55, 24, 19, 7, 0,
+                602, 347, 347, 347, 3, 12, 66,
+                46, 72, 80, 6, 0, 4])
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(back_tokens, [SPIECE_UNDERLINE + u'I', SPIECE_UNDERLINE + u'was', SPIECE_UNDERLINE + u'b',
+                                        u'or', u'n', SPIECE_UNDERLINE + u'in',
+                                        SPIECE_UNDERLINE + u'', u'<unk>', u'2', u'0', u'0', u'0', u',',
+                                        SPIECE_UNDERLINE + u'and', SPIECE_UNDERLINE + u'this',
+                                        SPIECE_UNDERLINE + u'is', SPIECE_UNDERLINE + u'f', u'al', u's',
+                                        u'<unk>', u'.'])
+
+
+if __name__ == '__main__':
+    unittest.main()

transformers/tests/tokenization_tests_commons.py

@@ -102,14 +102,55 @@ class CommonTestCases:
             with TemporaryDirectory() as tmpdirname:
 
                 filename = os.path.join(tmpdirname, u"tokenizer.bin")
-                pickle.dump(tokenizer, open(filename, "wb"))
+                with open(filename, "wb") as handle:
+                    pickle.dump(tokenizer, handle)
 
-                tokenizer_new = pickle.load(open(filename, "rb"))
+                with open(filename, "rb") as handle:
+                    tokenizer_new = pickle.load(handle)
 
             subwords_loaded = tokenizer_new.tokenize(text)
 
             self.assertListEqual(subwords, subwords_loaded)
 
+        def test_added_tokens_do_lower_case(self):
+            tokenizer = self.get_tokenizer(do_lower_case=True)
+
+            special_token = tokenizer.all_special_tokens[0]
+
+            text = special_token + " aaaaa bbbbbb low cccccccccdddddddd l " + special_token
+            text2 = special_token + " AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l " + special_token
+
+            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)
+
+            # Check that none of the special tokens are lowercased
+            sequence_with_special_tokens = "A " + " yEs ".join(tokenizer.all_special_tokens) + " B"
+            tokenized_sequence = tokenizer.tokenize(sequence_with_special_tokens)
+
+            for special_token in tokenizer.all_special_tokens:
+                self.assertTrue(special_token in tokenized_sequence)
+
+            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[1], toks2[1])  # But at least the first non-special tokens should differ
 
         def test_add_tokens_tokenizer(self):
             tokenizer = self.get_tokenizer()
@@ -198,6 +239,15 @@ class CommonTestCases:
             self.assertNotEqual(len(tokens_2), 0)
             self.assertIsInstance(text_2, (str, unicode))
 
+        def test_encode_decode_with_spaces(self):
+            tokenizer = self.get_tokenizer()
+
+            new_toks = ['[ABC]', '[DEF]', 'GHI IHG']
+            tokenizer.add_tokens(new_toks)
+            input = "[ABC] [DEF] [ABC] GHI IHG [DEF]"
+            encoded = tokenizer.encode(input, add_special_tokens=False)
+            decoded = tokenizer.decode(encoded)
+            self.assertEqual(decoded, input)
 
         def test_pretrained_model_lists(self):
             weights_list = list(self.tokenizer_class.max_model_input_sizes.keys())
@@ -243,7 +293,11 @@ 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 +324,12 @@ 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 +361,7 @@ 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 +373,8 @@ 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))
@@ -329,9 +386,98 @@ 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)
             self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
             self.assertEqual(special_tokens_mask_orig, special_tokens_mask)
+
+        def test_padding_to_max_length(self):
+            tokenizer = self.get_tokenizer()
+
+            sequence = "Sequence"
+            padding_size = 10
+            padding_idx = tokenizer.pad_token_id
+
+            # RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
+            tokenizer.padding_side = "right"
+            encoded_sequence = tokenizer.encode(sequence)
+            sequence_length = len(encoded_sequence)
+            padded_sequence = tokenizer.encode(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True)
+            padded_sequence_length = len(padded_sequence)
+            assert sequence_length + padding_size == padded_sequence_length
+            assert encoded_sequence + [padding_idx] * padding_size == padded_sequence
+
+            # LEFT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
+            tokenizer.padding_side = "left"
+            encoded_sequence = tokenizer.encode(sequence)
+            sequence_length = len(encoded_sequence)
+            padded_sequence = tokenizer.encode(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True)
+            padded_sequence_length = len(padded_sequence)
+            assert sequence_length + padding_size == padded_sequence_length
+            assert [padding_idx] * padding_size + encoded_sequence == padded_sequence
+
+            # RIGHT & LEFT PADDING - Check that nothing is done when a maximum length is not specified
+            encoded_sequence = tokenizer.encode(sequence)
+            sequence_length = len(encoded_sequence)
+
+            tokenizer.padding_side = "right"
+            padded_sequence_right = tokenizer.encode(sequence, pad_to_max_length=True)
+            padded_sequence_right_length = len(padded_sequence_right)
+
+            tokenizer.padding_side = "left"
+            padded_sequence_left = tokenizer.encode(sequence, pad_to_max_length=True)
+            padded_sequence_left_length = len(padded_sequence_left)
+
+            assert sequence_length == padded_sequence_right_length
+            assert encoded_sequence == padded_sequence_right
+            assert sequence_length == padded_sequence_left_length
+            assert encoded_sequence == padded_sequence_left
+
+        def test_encode_plus_with_padding(self):
+            tokenizer = self.get_tokenizer()
+
+            sequence = "Sequence"
+            padding_size = 10
+            padding_idx = tokenizer.pad_token_id
+            token_type_padding_idx = tokenizer.pad_token_type_id
+
+            encoded_sequence = tokenizer.encode_plus(sequence, return_special_tokens_mask=True)
+            input_ids = encoded_sequence['input_ids']
+            token_type_ids = encoded_sequence['token_type_ids']
+            attention_mask = encoded_sequence['attention_mask']
+            special_tokens_mask = encoded_sequence['special_tokens_mask']
+            sequence_length = len(input_ids)
+
+            # Test right padding
+            tokenizer.padding_side = "right"
+            padded_sequence = tokenizer.encode_plus(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True, return_special_tokens_mask=True)
+            padded_input_ids = padded_sequence['input_ids']
+            padded_token_type_ids = padded_sequence['token_type_ids']
+            padded_attention_mask = padded_sequence['attention_mask']
+            padded_special_tokens_mask = padded_sequence['special_tokens_mask']
+            padded_sequence_length = len(padded_input_ids)
+
+            assert sequence_length + padding_size == padded_sequence_length
+            assert input_ids + [padding_idx] * padding_size == padded_input_ids
+            assert token_type_ids + [token_type_padding_idx] * padding_size == padded_token_type_ids
+            assert attention_mask + [0] * padding_size == padded_attention_mask 
+            assert special_tokens_mask + [1] * padding_size == padded_special_tokens_mask 
+
+            # Test left padding
+            tokenizer.padding_side = "left"
+            padded_sequence = tokenizer.encode_plus(sequence, max_length=sequence_length + padding_size, pad_to_max_length=True, return_special_tokens_mask=True)
+            padded_input_ids = padded_sequence['input_ids']
+            padded_token_type_ids = padded_sequence['token_type_ids']
+            padded_attention_mask = padded_sequence['attention_mask']
+            padded_special_tokens_mask = padded_sequence['special_tokens_mask']
+            padded_sequence_length = len(padded_input_ids)
+
+            assert sequence_length + padding_size == padded_sequence_length
+            assert [padding_idx] * padding_size + input_ids == padded_input_ids
+            assert [token_type_padding_idx] * padding_size + token_type_ids == padded_token_type_ids
+            assert [0] * padding_size + attention_mask == padded_attention_mask 
+            assert [1] * padding_size + special_tokens_mask == padded_special_tokens_mask 
\ No newline at end of file

transformers/tests/tokenization_transfo_xl_test.py

@@ -16,7 +16,6 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import os
 import unittest
-import pytest
 from io import open
 
 from transformers import is_torch_available
@@ -24,11 +23,12 @@ from transformers import is_torch_available
 if is_torch_available():
     import torch
     from transformers.tokenization_transfo_xl import TransfoXLTokenizer, VOCAB_FILES_NAMES
-else:
-    pytestmark = pytest.mark.skip("Require Torch")  # TODO: untangle Transfo-XL tokenizer from torch.load and torch.save
 
 from .tokenization_tests_commons import CommonTestCases
+from .utils import require_torch
 
+
+@require_torch
 class TransfoXLTokenizationTest(CommonTestCases.CommonTokenizerTester):
 
     tokenizer_class = TransfoXLTokenizer if is_torch_available() else None

transformers/tests/tokenization_utils_test.py

@@ -18,13 +18,14 @@ from __future__ import print_function
 
 import unittest
 import six
-import pytest
 
 from transformers import PreTrainedTokenizer
 from transformers.tokenization_gpt2 import GPT2Tokenizer
 
+from .utils import slow
+
 class TokenizerUtilsTest(unittest.TestCase):
-    @pytest.mark.slow
+
     def check_tokenizer_from_pretrained(self, tokenizer_class):
         s3_models = list(tokenizer_class.max_model_input_sizes.keys())
         for model_name in s3_models[:1]:
@@ -41,6 +42,7 @@ class TokenizerUtilsTest(unittest.TestCase):
                 special_tok_id = tokenizer.convert_tokens_to_ids(special_tok)
                 self.assertIsInstance(special_tok_id, int)
 
+    @slow
     def test_pretrained_tokenizers(self):
         self.check_tokenizer_from_pretrained(GPT2Tokenizer)
 

transformers/tests/tokenization_xlm_test.py

@@ -17,11 +17,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 import os
 import unittest
 import json
-import pytest
 
 from transformers.tokenization_xlm import XLMTokenizer, VOCAB_FILES_NAMES
 
 from .tokenization_tests_commons import CommonTestCases
+from .utils import slow
 
 class XLMTokenizationTest(CommonTestCases.CommonTokenizerTester):
 
@@ -67,7 +67,7 @@ class XLMTokenizationTest(CommonTestCases.CommonTokenizerTester):
         self.assertListEqual(
             tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
 
-    @pytest.mark.slow
+    @slow
     def test_sequence_builders(self):
         tokenizer = XLMTokenizer.from_pretrained("xlm-mlm-en-2048")
 

transformers/tests/tokenization_xlnet_test.py

@@ -16,11 +16,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import os
 import unittest
-import pytest
 
 from transformers.tokenization_xlnet import (XLNetTokenizer, SPIECE_UNDERLINE)
 
 from .tokenization_tests_commons import CommonTestCases
+from .utils import slow
 
 SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)),
                     'fixtures/test_sentencepiece.model')
@@ -90,7 +90,7 @@ class XLNetTokenizationTest(CommonTestCases.CommonTokenizerTester):
                                       u'9', u'2', u'0', u'0', u'0', u',', SPIECE_UNDERLINE + u'and', SPIECE_UNDERLINE + u'this',
                                       SPIECE_UNDERLINE + u'is', SPIECE_UNDERLINE + u'f', u'al', u'se', u'.'])
 
-    @pytest.mark.slow
+    @slow
     def test_sequence_builders(self):
         tokenizer = XLNetTokenizer.from_pretrained("xlnet-base-cased")