Merge branch 'automodels' of https://github.com/huggingface/pytorch-transformers into automodels

hubconf.py

 dependencies = ['torch', 'tqdm', 'boto3', 'requests', 'regex']
 
-from pytorch_transformers import (
-    AutoTokenizer, AutoModel, AutoModelWithLMHead, AutoModelForSequenceClassification, AutoModelForQuestionAnswering
+from hubconfs.automodels_hubconf import (
+    autoConfig,
+    autoModel,
+    autoModelForQuestionAnswering,
+    autoModelForSequenceClassification,
+    autoModelWithLMHead,
+    autoTokenizer,
 )

pytorch_transformers/__init__.py

@@ -7,6 +7,7 @@ from .tokenization_gpt2 import GPT2Tokenizer
 from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
 from .tokenization_xlm import XLMTokenizer
 from .tokenization_roberta import RobertaTokenizer
+from .tokenization_distilbert import DistilBertTokenizer
 
 from .tokenization_utils import (PreTrainedTokenizer)
 
@@ -41,6 +42,9 @@ from .modeling_xlm import (XLMConfig, XLMPreTrainedModel , XLMModel,
                            XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
 from .modeling_roberta import (RobertaConfig, RobertaForMaskedLM, RobertaModel, RobertaForSequenceClassification,
                                ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_distilbert import (DistilBertConfig, DistilBertForMaskedLM, DistilBertModel,
+                               DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
+                               DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
 from .modeling_utils import (WEIGHTS_NAME, CONFIG_NAME, TF_WEIGHTS_NAME,
                           PretrainedConfig, PreTrainedModel, prune_layer, Conv1D)
 

pytorch_transformers/modeling_auto.py

@@ -30,11 +30,13 @@ from .modeling_transfo_xl import TransfoXLConfig, TransfoXLModel, TransfoXLLMHea
 from .modeling_xlnet import XLNetConfig, XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering
 from .modeling_xlm import XLMConfig, XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering
 from .modeling_roberta import RobertaConfig, RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification
+from .modeling_distilbert import DistilBertConfig, DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, DistilBertForSequenceClassification
 
 from .modeling_utils import PreTrainedModel, SequenceSummary, add_start_docstrings
 
 logger = logging.getLogger(__name__)
 
+
 class AutoConfig(object):
     r""":class:`~pytorch_transformers.AutoConfig` is a generic configuration class
         that will be instantiated as one of the configuration classes of the library
@@ -46,6 +48,7 @@ class AutoConfig(object):
 
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertConfig (DistilBERT model)
             - contains `bert`: BertConfig (Bert model)
             - contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model)
             - contains `gpt2`: GPT2Config (OpenAI GPT-2 model)
@@ -67,6 +70,7 @@ class AutoConfig(object):
 
         The configuration class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertConfig (DistilBERT model)
             - contains `bert`: BertConfig (Bert model)
             - contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model)
             - contains `gpt2`: GPT2Config (OpenAI GPT-2 model)
@@ -116,7 +120,9 @@ class AutoConfig(object):
             assert unused_kwargs == {'foo': False}
 
         """
-        if 'roberta' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
             return RobertaConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
@@ -148,6 +154,7 @@ class AutoModel(object):
 
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertModel (DistilBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -169,6 +176,7 @@ class AutoModel(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertModel (DistilBERT model)
             - contains `roberta`: RobertaModel (RoBERTa model)
             - contains `bert`: BertModel (Bert model)
             - contains `openai-gpt`: OpenAIGPTModel (OpenAI GPT model)
@@ -233,7 +241,9 @@ class AutoModel(object):
             model = AutoModel.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'roberta' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
             return RobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
@@ -253,7 +263,6 @@ class AutoModel(object):
                          "'xlm', 'roberta'".format(pretrained_model_name_or_path))
 
 
-
 class AutoModelWithLMHead(object):
     r"""
         :class:`~pytorch_transformers.AutoModelWithLMHead` is a generic model class
@@ -266,6 +275,7 @@ class AutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -290,6 +300,7 @@ class AutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
             - contains `roberta`: RobertaForMaskedLM (RoBERTa model)
             - contains `bert`: BertForMaskedLM (Bert model)
             - contains `openai-gpt`: OpenAIGPTLMHeadModel (OpenAI GPT model)
@@ -354,7 +365,9 @@ class AutoModelWithLMHead(object):
             model = AutoModelWithLMHead.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'roberta' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
@@ -386,6 +399,7 @@ class AutoModelForSequenceClassification(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -407,6 +421,7 @@ class AutoModelForSequenceClassification(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForSequenceClassification (DistilBERT model)
             - contains `roberta`: RobertaForSequenceClassification (RoBERTa model)
             - contains `bert`: BertForSequenceClassification (Bert model)
             - contains `xlnet`: XLNetForSequenceClassification (XLNet model)
@@ -468,7 +483,9 @@ class AutoModelForSequenceClassification(object):
             model = AutoModelForSequenceClassification.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'roberta' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
             return RobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
             return BertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
@@ -493,6 +510,7 @@ class AutoModelForQuestionAnswering(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForQuestionAnswering (DistilBERT model)
             - contains `bert`: BertForQuestionAnswering (Bert model)
             - contains `xlnet`: XLNetForQuestionAnswering (XLNet model)
             - contains `xlm`: XLMForQuestionAnswering (XLM model)
@@ -513,6 +531,7 @@ class AutoModelForQuestionAnswering(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `distilbert`: DistilBertForQuestionAnswering (DistilBERT model)
             - contains `bert`: BertForQuestionAnswering (Bert model)
             - contains `xlnet`: XLNetForQuestionAnswering (XLNet model)
             - contains `xlm`: XLMForQuestionAnswering (XLM model)
@@ -573,7 +592,9 @@ class AutoModelForQuestionAnswering(object):
             model = AutoModelForQuestionAnswering.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'bert' in pretrained_model_name_or_path:
+        if 'distilbert' in pretrained_model_name_or_path:
+            return DistilBertForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'bert' in pretrained_model_name_or_path:
             return BertForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'xlnet' in pretrained_model_name_or_path:
             return XLNetForQuestionAnswering.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

pytorch_transformers/modeling_bert.py

@@ -216,7 +216,7 @@ class BertConfig(PretrainedConfig):
             self.layer_norm_eps = layer_norm_eps
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
 
 
@@ -224,20 +224,7 @@ try:
     from apex.normalization.fused_layer_norm import FusedLayerNorm as BertLayerNorm
 except (ImportError, AttributeError) as e:
     logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
-    class BertLayerNorm(nn.Module):
-        def __init__(self, hidden_size, eps=1e-12):
-            """Construct a layernorm module in the TF style (epsilon inside the square root).
-            """
-            super(BertLayerNorm, self).__init__()
-            self.weight = nn.Parameter(torch.ones(hidden_size))
-            self.bias = nn.Parameter(torch.zeros(hidden_size))
-            self.variance_epsilon = eps
-
-        def forward(self, x):
-            u = x.mean(-1, keepdim=True)
-            s = (x - u).pow(2).mean(-1, keepdim=True)
-            x = (x - u) / torch.sqrt(s + self.variance_epsilon)
-            return self.weight * x + self.bias
+    BertLayerNorm = torch.nn.LayerNorm
 
 class BertEmbeddings(nn.Module):
     """Construct the embeddings from word, position and token_type embeddings.
@@ -449,7 +436,7 @@ class BertEncoder(nn.Module):
             outputs = outputs + (all_hidden_states,)
         if self.output_attentions:
             outputs = outputs + (all_attentions,)
-        return outputs  # outputs, (hidden states), (attentions)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
 
 
 class BertPooler(nn.Module):

pytorch_transformers/modeling_gpt2.py

@@ -408,10 +408,6 @@ GPT2_INPUTS_DOCSTRING = r"""    Inputs:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
-            Mask to avoid performing attention on padding token indices.
-            Mask values selected in ``[0, 1]``:
-            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
         **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
@@ -642,10 +638,6 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
-            Mask to avoid performing attention on padding token indices.
-            Mask values selected in ``[0, 1]``:
-            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
         **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
@@ -656,14 +648,11 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             Indices are selected in ``[-1, 0, ..., config.vocab_size]``
             All labels set to ``-1`` are ignored (masked), the loss is only
             computed for labels in ``[0, ..., config.vocab_size]``
-        **multiple_choice_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
+        **mc_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size)``:
             Labels for computing the multiple choice classification loss.
             Indices should be in ``[0, ..., num_choices]`` where `num_choices` is the size of the second dimension
             of the input tensors. (see `input_ids` above)
 
-            `multiple_choice_labels`: optional multiple choice labels: ``torch.LongTensor`` of shape [batch_size]
-                with indices selected in [0, ..., num_choices].
-
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **lm_loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
             Language modeling loss.

pytorch_transformers/modeling_roberta.py

@@ -98,15 +98,15 @@ ROBERTA_INPUTS_DOCSTRING = r"""
     Inputs:
         **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of input sequence tokens in the vocabulary.
-            To match pre-training, RoBERTa input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+            To match pre-training, RoBERTa input sequence should be formatted with <s> and </s> tokens as follows:
 
             (a) For sequence pairs:
 
-                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP][SEP] no it is not . [SEP]``
+                ``tokens:         <s> Is this Jacksonville ? </s> </s> No it is not . </s>``
 
             (b) For single sequences:
 
-                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                ``tokens:         <s> the dog is hairy . </s>``
 
             Fully encoded sequences or sequence pairs can be obtained using the RobertaTokenizer.encode function with 
             the ``add_special_tokens`` parameter set to ``True``.

pytorch_transformers/modeling_transfo_xl.py

@@ -285,7 +285,7 @@ class TransfoXLConfig(PretrainedConfig):
             self.init_std = init_std
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
     @property
     def max_position_embeddings(self):
@@ -1142,10 +1142,10 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
             else:
                 mask_shift_len = qlen
             dec_attn_mask = (torch.triu(all_ones, 1+mlen)
-                    + torch.tril(all_ones, -mask_shift_len)).byte()[:, :, None] # -1
+                    + torch.tril(all_ones, -mask_shift_len)).bool()[:, :, None] # -1
         else:
             dec_attn_mask = torch.triu(
-                word_emb.new_ones(qlen, klen), diagonal=1+mlen).byte()[:,:,None]
+                word_emb.new_ones(qlen, klen), diagonal=1+mlen).bool()[:,:,None]
 
         hids = []
         attentions = []

pytorch_transformers/modeling_xlm.py

@@ -178,7 +178,7 @@ class XLMConfig(PretrainedConfig):
             self.end_n_top = end_n_top
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
     @property
     def vocab_size(self):

pytorch_transformers/modeling_xlnet.py

@@ -306,7 +306,7 @@ class XLNetConfig(PretrainedConfig):
             self.end_n_top = end_n_top
         else:
             raise ValueError("First argument must be either a vocabulary size (int)"
-                             "or the path to a pretrained model config file (str)")
+                             " or the path to a pretrained model config file (str)")
 
     @property
     def max_position_embeddings(self):
@@ -677,8 +677,11 @@ XLNET_INPUTS_DOCSTRING = r"""
             ``1`` for tokens that are MASKED, ``0`` for tokens that are NOT MASKED.
         **mems**: (`optional`)
             list of ``torch.FloatTensor`` (one for each layer):
-            that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
+            that contains pre-computed hidden-states (key and values in the attention blocks) as output by the model
             (see `mems` output below). Can be used to speed up sequential decoding and attend to longer context.
+            To activate mems you need to set up config.mem_len to a positive value which will be the max number of tokens in
+            the memory output by the model. E.g. `model = XLNetModel.from_pretrained('xlnet-base-case, mem_len=1024)` will
+            instantiate a model which can use up to 1024 tokens of memory (in addition to the input it self).
         **perm_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, sequence_length)``:
             Mask to indicate the attention pattern for each input token with values selected in ``[0, 1]``:
             If ``perm_mask[k, i, j] = 0``, i attend to j in batch k;
@@ -705,7 +708,8 @@ class XLNetModel(XLNetPreTrainedModel):
         **mems**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `mems` input above). Can be used to speed up sequential decoding and attend to longer context.
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -859,7 +863,7 @@ class XLNetModel(XLNetPreTrainedModel):
         target_mapping = target_mapping.permute(1, 2, 0).contiguous() if target_mapping is not None else None
 
         qlen, bsz = input_ids.shape[0], input_ids.shape[1]
-        mlen = mems[0].shape[0] if mems is not None else 0
+        mlen = mems[0].shape[0] if mems is not None and mems[0] is not None else 0
         klen = mlen + qlen
 
         dtype_float = next(self.parameters()).dtype
@@ -1011,7 +1015,8 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         **mems**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `mems` input above). Can be used to speed up sequential decoding and attend to longer context.
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -1091,7 +1096,8 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
         **mems**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `mems` input above). Can be used to speed up sequential decoding and attend to longer context.
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:
@@ -1189,7 +1195,8 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
         **mems**:
             list of ``torch.FloatTensor`` (one for each layer):
             that contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `mems` input above). Can be used to speed up sequential decoding and attend to longer context.
+            if config.mem_len > 0 else tuple of None. Can be used to speed up sequential decoding and attend to longer context.
+            See details in the docstring of the `mems` input above.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
             list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
             of shape ``(batch_size, sequence_length, hidden_size)``:

pytorch_transformers/tests/modeling_common_test.py

@@ -49,6 +49,7 @@ class CommonTestCases:
         test_torchscript = True
         test_pruning = True
         test_resize_embeddings = True
+        test_head_masking = True
 
         def test_initialization(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -159,6 +160,9 @@ class CommonTestCases:
 
 
         def test_headmasking(self):
+            if not self.test_head_masking:
+                return
+
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
             config.output_attentions = True
@@ -282,6 +286,9 @@ class CommonTestCases:
                 self.assertTrue(models_equal)
 
         def test_tie_model_weights(self):
+            if not self.test_torchscript:
+                return
+
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
             def check_same_values(layer_1, layer_2):

pytorch_transformers/tests/modeling_dilbert_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+
+from pytorch_transformers import (DistilBertConfig, DistilBertModel, DistilBertForMaskedLM,
+                                  DistilBertForQuestionAnswering, DistilBertForSequenceClassification)
+from pytorch_transformers.modeling_distilbert import DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+from .modeling_common_test import (CommonTestCases, ConfigTester, ids_tensor)
+
+
+class DistilBertModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (DistilBertModel, DistilBertForMaskedLM, DistilBertForQuestionAnswering,
+                         DistilBertForSequenceClassification)
+    test_pruning = True
+    test_torchscript = True
+    test_resize_embeddings = True
+    test_head_masking = True
+
+    class DistilBertModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=False,
+                     use_labels=True,
+                     vocab_size=99,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     intermediate_size=37,
+                     hidden_act="gelu",
+                     hidden_dropout_prob=0.1,
+                     attention_probs_dropout_prob=0.1,
+                     max_position_embeddings=512,
+                     type_vocab_size=16,
+                     type_sequence_label_size=2,
+                     initializer_range=0.02,
+                     num_labels=3,
+                     num_choices=4,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_token_type_ids = use_token_type_ids
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.intermediate_size = intermediate_size
+            self.hidden_act = hidden_act
+            self.hidden_dropout_prob = hidden_dropout_prob
+            self.attention_probs_dropout_prob = attention_probs_dropout_prob
+            self.max_position_embeddings = max_position_embeddings
+            self.type_vocab_size = type_vocab_size
+            self.type_sequence_label_size = type_sequence_label_size
+            self.initializer_range = initializer_range
+            self.num_labels = num_labels
+            self.num_choices = num_choices
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = DistilBertConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                dim=self.hidden_size,
+                n_layers=self.num_hidden_layers,
+                n_heads=self.num_attention_heads,
+                hidden_dim=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                dropout=self.hidden_dropout_prob,
+                attention_dropout=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                initializer_range=self.initializer_range)
+
+            return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_distilbert_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = DistilBertModel(config=config)
+            model.eval()
+            (sequence_output,) = model(input_ids, input_mask)
+            (sequence_output,) = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output,
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].size()),
+                [self.batch_size, self.seq_length, self.hidden_size])
+
+        def create_and_check_distilbert_for_masked_lm(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = DistilBertForMaskedLM(config=config)
+            model.eval()
+            loss, prediction_scores = model(input_ids, attention_mask=input_mask, masked_lm_labels=token_labels)
+            result = {
+                "loss": loss,
+                "prediction_scores": prediction_scores,
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].size()),
+                [self.batch_size, self.seq_length, self.vocab_size])
+            self.check_loss_output(result)
+
+        def create_and_check_distilbert_for_question_answering(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            model = DistilBertForQuestionAnswering(config=config)
+            model.eval()
+            loss, start_logits, end_logits = model(input_ids, input_mask, sequence_labels, sequence_labels)
+            result = {
+                "loss": loss,
+                "start_logits": start_logits,
+                "end_logits": end_logits,
+            }
+            self.parent.assertListEqual(
+                list(result["start_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.parent.assertListEqual(
+                list(result["end_logits"].size()),
+                [self.batch_size, self.seq_length])
+            self.check_loss_output(result)
+
+        def create_and_check_distilbert_for_sequence_classification(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
+            config.num_labels = self.num_labels
+            model = DistilBertForSequenceClassification(config)
+            model.eval()
+            loss, logits = model(input_ids, input_mask, sequence_labels)
+            result = {
+                "loss": loss,
+                "logits": logits,
+            }
+            self.parent.assertListEqual(
+                list(result["logits"].size()),
+                [self.batch_size, self.num_labels])
+            self.check_loss_output(result)
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, input_mask, sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = DistilBertModelTest.DistilBertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=DistilBertConfig, dim=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_distilbert_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_distilbert_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_distilbert_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_distilbert_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_distilbert_for_sequence_classification(*config_and_inputs)
+
+    # @pytest.mark.slow
+    # def test_model_from_pretrained(self):
+    #     cache_dir = "/tmp/pytorch_transformers_test/"
+    #     for model_name in list(DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+    #         model = DistilBertModel.from_pretrained(model_name, cache_dir=cache_dir)
+    #         shutil.rmtree(cache_dir)
+    #         self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()

pytorch_transformers/tests/tokenization_bert_test.py

@@ -42,7 +42,7 @@ class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
             vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
 
     def get_tokenizer(self):
-        return BertTokenizer.from_pretrained(self.tmpdirname)
+        return self.tokenizer_class.from_pretrained(self.tmpdirname)
 
     def get_input_output_texts(self):
         input_text = u"UNwant\u00E9d,running"
@@ -50,7 +50,7 @@ class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
         return input_text, output_text
 
     def test_full_tokenizer(self):
-        tokenizer = BertTokenizer(self.vocab_file)
+        tokenizer = self.tokenizer_class(self.vocab_file)
 
         tokens = tokenizer.tokenize(u"UNwant\u00E9d,running")
         self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
@@ -126,7 +126,7 @@ class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
         self.assertFalse(_is_punctuation(u" "))
 
     def test_sequence_builders(self):
-        tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+        tokenizer = self.tokenizer_class.from_pretrained("bert-base-uncased")
 
         text = tokenizer.encode("sequence builders")
         text_2 = tokenizer.encode("multi-sequence build")

pytorch_transformers/tests/tokenization_dilbert_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 pytorch_transformers.tokenization_distilbert import (DistilBertTokenizer)
+
+from .tokenization_tests_commons import CommonTestCases
+from .tokenization_bert_test import BertTokenizationTest
+
+class DistilBertTokenizationTest(BertTokenizationTest):
+
+    tokenizer_class = DistilBertTokenizer
+
+    def get_tokenizer(self):
+        return DistilBertTokenizer.from_pretrained(self.tmpdirname)
+
+    def test_sequence_builders(self):
+        tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
+
+        text = tokenizer.encode("sequence builders")
+        text_2 = tokenizer.encode("multi-sequence build")
+
+        encoded_sentence = tokenizer.add_special_tokens_single_sentence(text)
+        encoded_pair = tokenizer.add_special_tokens_sentences_pair(text, text_2)
+
+        assert encoded_sentence == [101] + text + [102]
+        assert encoded_pair == [101] + text + [102] + text_2 + [102]
+
+if __name__ == '__main__':
+    unittest.main()

pytorch_transformers/tokenization_bert.py

@@ -125,6 +125,9 @@ class BertTokenizer(PreTrainedTokenizer):
         super(BertTokenizer, self).__init__(unk_token=unk_token, sep_token=sep_token,
                                             pad_token=pad_token, cls_token=cls_token,
                                             mask_token=mask_token, **kwargs)
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 3  # take into account special tokens
+
         if not os.path.isfile(vocab_file):
             raise ValueError(
                 "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "

pytorch_transformers/tokenization_distilbert.py

+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tokenization classes for DistilBERT."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import collections
+import logging
+import os
+import unicodedata
+from io import open
+
+from .tokenization_bert import BertTokenizer
+
+logger = logging.getLogger(__name__)
+
+VOCAB_FILES_NAMES = {'vocab_file': 'vocab.txt'}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
+        'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    'distilbert-base-uncased': 512,
+    'distilbert-base-uncased-distilled-squad': 512,
+}
+
+
+class DistilBertTokenizer(BertTokenizer):
+    r"""
+    Constructs a DistilBertTokenizer.
+    :class:`~pytorch_transformers.DistilBertTokenizer` is identical to BertTokenizer and runs end-to-end tokenization: punctuation splitting + wordpiece
+
+    Args:
+        vocab_file: Path to a one-wordpiece-per-line vocabulary file
+        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_basic_tokenize: Whether to do basic tokenization before wordpiece.
+        max_len: An artificial maximum length to truncate tokenized sequences to; Effective maximum length is always the
+            minimum of this value (if specified) and the underlying BERT model's sequence length.
+        never_split: List of tokens which will never be split during tokenization. Only has an effect when
+            do_wordpiece_only=False
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES

pytorch_transformers/tokenization_gpt2.py

@@ -108,6 +108,8 @@ class GPT2Tokenizer(PreTrainedTokenizer):
     def __init__(self, vocab_file, merges_file, errors='replace', unk_token="<|endoftext|>",
                  bos_token="<|endoftext|>", eos_token="<|endoftext|>", **kwargs):
         super(GPT2Tokenizer, self).__init__(bos_token=bos_token, eos_token=eos_token, unk_token=unk_token, **kwargs)
+        self.max_len_single_sentence = self.max_len # no default special tokens - you can update this value if you add special tokens
+        self.max_len_sentences_pair = self.max_len # no default special tokens - you can update this value if you add special tokens
 
         self.encoder = json.load(open(vocab_file))
         self.decoder = {v:k for k,v in self.encoder.items()}

pytorch_transformers/tokenization_openai.py

@@ -87,6 +87,9 @@ class OpenAIGPTTokenizer(PreTrainedTokenizer):
     def __init__(self, vocab_file, merges_file, unk_token="<unk>", **kwargs):
         super(OpenAIGPTTokenizer, self).__init__(unk_token=unk_token, **kwargs)
 
+        self.max_len_single_sentence = self.max_len # no default special tokens - you can update this value if you add special tokens
+        self.max_len_sentences_pair = self.max_len # no default special tokens - you can update this value if you add special tokens
+
         try:
             import ftfy
             from spacy.lang.en import English

pytorch_transformers/tokenization_roberta.py

@@ -77,6 +77,9 @@ class RobertaTokenizer(PreTrainedTokenizer):
                                                sep_token=sep_token, cls_token=cls_token, pad_token=pad_token,
                                                mask_token=mask_token, **kwargs)
 
+        self.max_len_single_sentence = self.max_len - 2  # take into account special tokens
+        self.max_len_sentences_pair = self.max_len - 4  # take into account special tokens
+
         self.encoder = json.load(open(vocab_file, encoding="utf-8"))
         self.decoder = {v: k for k, v in self.encoder.items()}
         self.errors = errors  # how to handle errors in decoding
@@ -163,14 +166,14 @@ class RobertaTokenizer(PreTrainedTokenizer):
     def add_special_tokens_single_sentence(self, token_ids):
         """
         Adds special tokens to a sequence for sequence classification tasks.
-        A RoBERTa sequence has the following format: [CLS] X [SEP]
+        A RoBERTa sequence has the following format: <s> X </s>
         """
         return [self._convert_token_to_id(self.cls_token)] + token_ids + [self._convert_token_to_id(self.sep_token)]
 
     def add_special_tokens_sentences_pair(self, token_ids_0, token_ids_1):
         """
         Adds special tokens to a sequence pair for sequence classification tasks.
-        A RoBERTa sequence pair has the following format: [CLS] A [SEP][SEP] B [SEP]
+        A RoBERTa sequence pair has the following format: <s> A </s></s> B </s>
         """
         sep = [self._convert_token_to_id(self.sep_token)]
         cls = [self._convert_token_to_id(self.cls_token)]