Merge branch 'master' into do_lower_case

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_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_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