Merge branch 'master' into saving-and-resuming

transformers/modeling_ctrl.py

@@ -268,7 +268,7 @@ class CTRLModel(CTRLPreTrainedModel):
 
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLModel.from_pretrained('ctrl')
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -458,7 +458,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLLMHeadModel.from_pretrained('ctrl')
 
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 

transformers/modeling_distilbert.py

@@ -415,7 +415,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertModel.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -511,7 +511,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -581,7 +581,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -656,7 +656,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForQuestionAnswering.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/modeling_encoder_decoder.py

@@ -18,9 +18,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 import os
+import warnings
 
 import torch
 from torch import nn
+from tqdm import trange
 
 from .modeling_auto import AutoModel, AutoModelWithLMHead
 
@@ -59,12 +61,14 @@ class PreTrainedEncoderDecoder(nn.Module):
             encoder_pretrained_model_name_or_path: information necessary to initiate the encoder. Either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/encoder``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
             decoder_pretrained_model_name_or_path: information necessary to initiate the decoder. Either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/decoder``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -117,8 +121,7 @@ class PreTrainedEncoderDecoder(nn.Module):
         kwargs_common = {
             argument: value
             for argument, value in kwargs.items()
-            if not argument.startswith("encoder_")
-            and not argument.startswith("decoder_")
+            if not argument.startswith("encoder_") and not argument.startswith("decoder_")
         }
         kwargs_decoder = kwargs_common.copy()
         kwargs_encoder = kwargs_common.copy()
@@ -164,7 +167,39 @@ class PreTrainedEncoderDecoder(nn.Module):
 
         We save the encoder' and decoder's parameters in two separate directories.
         """
+
+        # If the root output directory does not exist, create it 
+        if not os.path.exists(save_directory):
+            os.mkdir(save_directory)
+
+        # Check whether the output directory is empty or not
+        sub_directories = [directory for directory in os.listdir(save_directory)
+            if os.path.isdir(os.path.join(save_directory, directory))]
+
+        if len(sub_directories) > 0:
+            if "encoder" in sub_directories and "decoder" in sub_directories:
+                print("WARNING: there is an older version of encoder-decoder saved in" +\
+                    " the output directory. The default behaviour is to overwrite them.")
+
+            # Empty the output directory
+            for directory_to_remove in sub_directories:
+                # Remove all files into the subdirectory
+                files_to_remove = os.listdir(os.path.join(save_directory, directory_to_remove))
+                for file_to_remove in files_to_remove:
+                    os.remove(os.path.join(save_directory, directory_to_remove, file_to_remove))
+                # Remove the subdirectory itself
+                os.rmdir(os.path.join(save_directory, directory_to_remove))
+
+            assert(len(os.listdir(save_directory)) == 0) # sanity check
+
+        # Create the "encoder" directory inside the output directory and save the encoder into it
+        if not os.path.exists(os.path.join(save_directory, "encoder")):
+            os.mkdir(os.path.join(save_directory, "encoder"))
         self.encoder.save_pretrained(os.path.join(save_directory, "encoder"))
+
+        # Create the "encoder" directory inside the output directory and save the decoder into it
+        if not os.path.exists(os.path.join(save_directory, "decoder")):
+            os.mkdir(os.path.join(save_directory, "decoder"))
         self.decoder.save_pretrained(os.path.join(save_directory, "decoder"))
 
     def forward(self, encoder_input_ids, decoder_input_ids, **kwargs):
@@ -186,51 +221,56 @@ class PreTrainedEncoderDecoder(nn.Module):
                 Indices of decoder input sequence tokens in the vocabulary.
             kwargs: (`optional`) Remaining dictionary of keyword arguments.
         """
-        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
-        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
-        # that apply to the model as whole.
-        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_encoder, kwargs_decoder = self.prepare_model_kwargs(**kwargs)
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            encoder_outputs = self.encoder(encoder_input_ids, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        decoder_outputs = self.decoder(decoder_input_ids, encoder_hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+    @staticmethod
+    def prepare_model_kwargs(**kwargs):
+        """ Prepare the encoder and decoder's keyword arguments.
+
+        Keyword arguments come in 3 flavors:
+        - encoder-specific (prefixed by `encoder_`)
+        - decoder-specific (prefixed by `decoder_`)
+        - those that apply to the model as whole.
+
+        We let the specific kwargs override the common ones in case of
+        conflict.
+        """
         kwargs_common = {
             argument: value
             for argument, value in kwargs.items()
-            if not argument.startswith("encoder_")
-            and not argument.startswith("decoder_")
+            if not argument.startswith("encoder_") and not argument.startswith("decoder_")
         }
-        kwargs_decoder = kwargs_common.copy()
-        kwargs_encoder = kwargs_common.copy()
-        kwargs_encoder.update(
+        decoder_kwargs = kwargs_common.copy()
+        encoder_kwargs = kwargs_common.copy()
+        encoder_kwargs.update(
             {
                 argument[len("encoder_") :]: value
                 for argument, value in kwargs.items()
                 if argument.startswith("encoder_")
             }
         )
-        kwargs_decoder.update(
+        decoder_kwargs.update(
             {
                 argument[len("decoder_") :]: value
                 for argument, value in kwargs.items()
                 if argument.startswith("decoder_")
             }
         )
-
-        # Encode if needed (training, first prediction pass)
-        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
-        if encoder_hidden_states is None:
-            encoder_outputs = self.encoder(encoder_input_ids, **kwargs_encoder)
-            encoder_hidden_states = encoder_outputs[
-                0
-            ]  # output the last layer hidden state
-        else:
-            encoder_outputs = ()
-
-        # Decode
-        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
-        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get(
-            "attention_mask", None
-        )
-        decoder_outputs = self.decoder(decoder_input_ids, **kwargs_decoder)
-
-        return decoder_outputs + encoder_outputs
+        decoder_kwargs["encoder_attention_mask"] = encoder_kwargs.get("attention_mask", None)
+        return encoder_kwargs, decoder_kwargs
 
 
 class Model2Model(PreTrainedEncoderDecoder):

transformers/modeling_gpt2.py

@@ -345,7 +345,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2Model.from_pretrained('gpt2')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -523,7 +523,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2LMHeadModel.from_pretrained('gpt2')
 
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -634,6 +634,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
     """
     def __init__(self, config):
         super(GPT2DoubleHeadsModel, self).__init__(config)
+        config.num_labels = 1
         self.transformer = GPT2Model(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)

transformers/modeling_openai.py

@@ -349,7 +349,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -491,7 +491,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -590,6 +590,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
     def __init__(self, config):
         super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
 
+        config.num_labels = 1
         self.transformer = OpenAIGPTModel(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)

transformers/modeling_roberta.py

@@ -51,24 +51,44 @@ class RobertaEmbeddings(BertEmbeddings):
                                                 padding_idx=self.padding_idx)
 
     def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+        if position_ids is None:
             if input_ids is not None:
-            input_shape = input_ids.size()
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = self.create_position_ids_from_input_ids(input_ids).to(input_ids.device)
             else:
-            input_shape = inputs_embeds.size()[:-1]
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
 
-        seq_length = input_shape[1]
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
-
-        if position_ids is None:
-            # Position numbers begin at padding_idx+1. Padding symbols are ignored.
-            # cf. fairseq's `utils.make_positions`
-            position_ids = torch.arange(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=torch.long, device=device)
-            position_ids = position_ids.unsqueeze(0).expand(input_shape)
         return super(RobertaEmbeddings, self).forward(input_ids,
                                                       token_type_ids=token_type_ids,
                                                       position_ids=position_ids,
                                                       inputs_embeds=inputs_embeds)
 
+    def create_position_ids_from_input_ids(self, x):
+        """ Replace non-padding symbols with their position numbers. Position numbers begin at
+        padding_idx+1. Padding symbols are ignored. This is modified from fairseq's
+        `utils.make_positions`.
+
+        :param torch.Tensor x:
+        :return torch.Tensor:
+        """
+        mask = x.ne(self.padding_idx).long()
+        incremental_indicies = torch.cumsum(mask, dim=1) * mask
+        return incremental_indicies + self.padding_idx
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """ We are provided embeddings directly. We cannot infer which are padded so just generate
+        sequential position ids.
+
+        :param torch.Tensor inputs_embeds:
+        :return torch.Tensor:
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(self.padding_idx+1, sequence_length+self.padding_idx+1, dtype=torch.long,
+                                    device=inputs_embeds.device)
+        return position_ids.unsqueeze(0).expand(input_shape)
+
 
 ROBERTA_START_DOCSTRING = r"""    The RoBERTa model was proposed in
     `RoBERTa: A Robustly Optimized BERT Pretraining Approach`_
@@ -168,7 +188,7 @@ class RobertaModel(BertModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaModel.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -216,7 +236,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaForMaskedLM.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -307,7 +327,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = RobertaForSequenceClassification.from_pretrained('roberta-base')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]

transformers/modeling_tf_albert.py

@@ -587,8 +587,8 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
         import tensorflow as tf
         from transformers import AlbertTokenizer, TFAlbertModel
 
-        tokenizer = AlbertTokenizer.from_pretrained('bert-base-uncased')
-        model = TFAlbertModel.from_pretrained('bert-base-uncased')
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1')
+        model = TFAlbertModel.from_pretrained('albert-base-v1')
         input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple

transformers/modeling_tf_bert.py

@@ -48,6 +48,12 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-tf_model.h5",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-tf_model.h5",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-tf_model.h5",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-tf_model.h5",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-tf_model.h5",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/tf_model.h5",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/tf_model.h5",
 }
 
 
@@ -671,7 +677,7 @@ class TFBertModel(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertModel.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -710,7 +716,7 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForPreTraining.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, seq_relationship_scores = outputs[:2]
 
@@ -759,7 +765,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForMaskedLM.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores = outputs[0]
 
@@ -806,7 +812,7 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForNextSentencePrediction.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         seq_relationship_scores = outputs[0]
 
@@ -851,7 +857,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 
@@ -988,7 +994,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForTokenClassification.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         scores = outputs[0]
 
@@ -1041,7 +1047,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForQuestionAnswering.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         start_scores, end_scores = outputs[:2]