Merge branch 'master' into master

pytorch_transformers/modeling_bert.py

@@ -74,7 +74,7 @@ def load_tf_weights_in_bert(model, config, tf_checkpoint_path):
         import numpy as np
         import tensorflow as tf
     except ImportError:
-        logger.error("Loading a TensorFlow models in PyTorch, requires TensorFlow to be installed. Please see "
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
             "https://www.tensorflow.org/install/ for installation instructions.")
         raise
     tf_path = os.path.abspath(tf_checkpoint_path)
@@ -216,28 +216,15 @@ 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)")
 
 
 
 try:
     from apex.normalization.fused_layer_norm import FusedLayerNorm as BertLayerNorm
-except ImportError:
+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):
@@ -577,7 +564,9 @@ BERT_START_DOCSTRING = r"""    The BERT model was proposed in
         https://pytorch.org/docs/stable/nn.html#module
 
     Parameters:
-        config (:class:`~pytorch_transformers.BertConfig`): Model configuration class with all the parameters of the model.
+        config (:class:`~pytorch_transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
 BERT_INPUTS_DOCSTRING = r"""
@@ -597,23 +586,26 @@ BERT_INPUTS_DOCSTRING = r"""
                 ``tokens:         [CLS] the dog is hairy . [SEP]``
                 
                 ``token_type_ids:   0   0   0   0  0     0   0``
-    
+
+            Bert is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
             Indices can be obtained using :class:`pytorch_transformers.BertTokenizer`.
             See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
             :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
         **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of positions of each input sequence tokens in the position embeddings.
-            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
         **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
             (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **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.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -643,12 +635,11 @@ class BertModel(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> model = BertModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertModel.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
 
     """
     def __init__(self, config):
@@ -754,13 +745,11 @@ class BertForPreTraining(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForPreTraining(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids)
-        >>> prediction_scores, seq_relationship_scores = outputs[:2]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForPreTraining.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores, seq_relationship_scores = outputs[:2]
 
     """
     def __init__(self, config):
@@ -824,13 +813,11 @@ class BertForMaskedLM(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForMaskedLM(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, masked_lm_labels=input_ids)
-        >>> loss, prediction_scores = outputs[:2]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForMaskedLM.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
 
     """
     def __init__(self, config):
@@ -857,7 +844,7 @@ class BertForMaskedLM(BertPreTrainedModel):
         sequence_output = outputs[0]
         prediction_scores = self.cls(sequence_output)
 
-        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention is they are here
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
         if masked_lm_labels is not None:
             loss_fct = CrossEntropyLoss(ignore_index=-1)
             masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
@@ -891,13 +878,11 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForNextSentencePrediction(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids)
-        >>> seq_relationship_scores = outputs[0]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForNextSentencePrediction.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        seq_relationship_scores = outputs[0]
 
     """
     def __init__(self, config):
@@ -932,7 +917,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the sequence classification/regression loss.
-            Indices should be in ``[0, ..., config.num_labels]``.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
 
@@ -951,14 +936,12 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForSequenceClassification(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=labels)
-        >>> loss, logits = outputs[:2]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
 
     """
     def __init__(self, config):
@@ -1027,12 +1010,12 @@ class BertForMultipleChoice(BertPreTrainedModel):
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
             (see `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, num_choices, sequence_length)``:
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             The second dimension of the input (`num_choices`) indicates the number of choices to score.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -1057,15 +1040,13 @@ class BertForMultipleChoice(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForMultipleChoice(config)
-        >>> choices = ["Hello, my dog is cute", "Hello, my cat is amazing"]
-        >>> input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
-        >>> labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=labels)
-        >>> loss, classification_scores = outputs[:2]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForMultipleChoice.from_pretrained('bert-base-uncased')
+        choices = ["Hello, my dog is cute", "Hello, my cat is amazing"]
+        input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        labels = torch.tensor(1).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, classification_scores = outputs[:2]
 
     """
     def __init__(self, config):
@@ -1110,7 +1091,7 @@ class BertForTokenClassification(BertPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for computing the token classification loss.
-            Indices should be in ``[0, ..., config.num_labels]``.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
 
     Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
         **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
@@ -1127,14 +1108,12 @@ class BertForTokenClassification(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForTokenClassification(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=labels)
-        >>> loss, scores = outputs[:2]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForTokenClassification.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1] * input_ids.size(1)).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, scores = outputs[:2]
 
     """
     def __init__(self, config):
@@ -1203,15 +1182,13 @@ class BertForQuestionAnswering(BertPreTrainedModel):
 
     Examples::
 
-        >>> config = BertConfig.from_pretrained('bert-base-uncased')
-        >>> tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
-        >>> 
-        >>> model = BertForQuestionAnswering(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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)
-        >>> loss, start_scores, end_scores = outputs[:2]
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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)
+        loss, start_scores, end_scores = outputs[:2]
 
     """
     def __init__(self, config):

pytorch_transformers/modeling_distilbert.py

+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team, The Google AI Language Team and Facebook, 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.
+""" PyTorch DistilBERT model
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+    and in part from HuggingFace PyTorch version of Google AI Bert model (https://github.com/google-research/bert)
+"""
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import copy
+import sys
+from io import open
+
+import itertools
+import numpy as np
+
+import torch
+import torch.nn as nn
+
+from pytorch_transformers.modeling_utils import PretrainedConfig, PreTrainedModel, add_start_docstrings, prune_linear_layer
+
+import logging
+logger = logging.getLogger(__name__)
+
+
+DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-pytorch_model.bin",
+    'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-pytorch_model.bin"
+}
+
+DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-config.json",
+    'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-config.json"
+}
+
+
+class DistilBertConfig(PretrainedConfig):
+    pretrained_config_archive_map = DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size_or_config_json_file=30522,
+                 max_position_embeddings=512,
+                 sinusoidal_pos_embds=True,
+                 n_layers=6,
+                 n_heads=12,
+                 dim=768,
+                 hidden_dim=4*768,
+                 dropout=0.1,
+                 attention_dropout=0.1,
+                 activation='gelu',
+                 initializer_range=0.02,
+                 tie_weights_=True,
+                 qa_dropout=0.1,
+                 seq_classif_dropout=0.2,
+                 **kwargs):
+        super(DistilBertConfig, self).__init__(**kwargs)
+
+        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
+                        and isinstance(vocab_size_or_config_json_file, unicode)):
+            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
+                json_config = json.loads(reader.read())
+            for key, value in json_config.items():
+                self.__dict__[key] = value
+        elif isinstance(vocab_size_or_config_json_file, int):
+            self.vocab_size = vocab_size_or_config_json_file
+            self.max_position_embeddings = max_position_embeddings
+            self.sinusoidal_pos_embds = sinusoidal_pos_embds
+            self.n_layers = n_layers
+            self.n_heads = n_heads
+            self.dim = dim
+            self.hidden_dim = hidden_dim
+            self.dropout = dropout
+            self.attention_dropout = attention_dropout
+            self.activation = activation
+            self.initializer_range = initializer_range
+            self.tie_weights_ = tie_weights_
+            self.qa_dropout = qa_dropout
+            self.seq_classif_dropout = seq_classif_dropout
+        else:
+            raise ValueError("First argument must be either a vocabulary size (int)"
+                             " or the path to a pretrained model config file (str)")
+    @property
+    def hidden_size(self):
+        return self.dim
+
+    @property
+    def num_attention_heads(self):
+        return self.n_heads
+
+    @property
+    def num_hidden_layers(self):
+        return self.n_layers
+
+
+### UTILS AND BUILDING BLOCKS OF THE ARCHITECTURE ###
+def gelu(x):
+    return 0.5 * x * (1.0 + torch.erf(x / math.sqrt(2.0)))
+
+def create_sinusoidal_embeddings(n_pos, dim, out):
+    position_enc = np.array([
+        [pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)]
+        for pos in range(n_pos)
+    ])
+    out[:, 0::2] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+    out[:, 1::2] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+    out.detach_()
+    out.requires_grad = False
+
+class Embeddings(nn.Module):
+    def __init__(self,
+                 config):
+        super(Embeddings, self).__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.dim, padding_idx=0)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.dim)
+        if config.sinusoidal_pos_embds:
+            create_sinusoidal_embeddings(n_pos=config.max_position_embeddings,
+                                         dim=config.dim,
+                                         out=self.position_embeddings.weight)
+
+        self.LayerNorm = nn.LayerNorm(config.dim, eps=1e-12)
+        self.dropout = nn.Dropout(config.dropout)
+
+    def forward(self, input_ids):
+        """
+        Parameters
+        ----------
+        input_ids: torch.tensor(bs, max_seq_length)
+            The token ids to embed.
+
+        Outputs
+        -------
+        embeddings: torch.tensor(bs, max_seq_length, dim)
+            The embedded tokens (plus position embeddings, no token_type embeddings)
+        """
+        seq_length = input_ids.size(1)
+        position_ids = torch.arange(seq_length, dtype=torch.long, device=input_ids.device) # (max_seq_length)
+        position_ids = position_ids.unsqueeze(0).expand_as(input_ids)                      # (bs, max_seq_length)
+
+        word_embeddings = self.word_embeddings(input_ids)                   # (bs, max_seq_length, dim)
+        position_embeddings = self.position_embeddings(position_ids)        # (bs, max_seq_length, dim)
+
+        embeddings = word_embeddings + position_embeddings  # (bs, max_seq_length, dim)
+        embeddings = self.LayerNorm(embeddings)             # (bs, max_seq_length, dim)
+        embeddings = self.dropout(embeddings)               # (bs, max_seq_length, dim)
+        return embeddings
+
+class MultiHeadSelfAttention(nn.Module):
+    def __init__(self, config):
+        super(MultiHeadSelfAttention, self).__init__()
+
+        self.n_heads = config.n_heads
+        self.dim = config.dim
+        self.dropout = nn.Dropout(p=config.attention_dropout)
+        self.output_attentions = config.output_attentions
+
+        assert self.dim % self.n_heads == 0
+
+        self.q_lin = nn.Linear(in_features=config.dim, out_features=config.dim)
+        self.k_lin = nn.Linear(in_features=config.dim, out_features=config.dim)
+        self.v_lin = nn.Linear(in_features=config.dim, out_features=config.dim)
+        self.out_lin = nn.Linear(in_features=config.dim, out_features=config.dim)
+
+    def prune_heads(self, heads):
+        attention_head_size = self.dim // self.n_heads
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.n_heads, attention_head_size)
+        for head in heads:
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+        # Prune linear layers
+        self.q_lin = prune_linear_layer(self.q_lin, index)
+        self.k_lin = prune_linear_layer(self.k_lin, index)
+        self.v_lin = prune_linear_layer(self.v_lin, index)
+        self.out_lin = prune_linear_layer(self.out_lin, index, dim=1)
+        # Update hyper params
+        self.n_heads = self.n_heads - len(heads)
+        self.dim = attention_head_size * self.n_heads
+
+    def forward(self, query, key, value, mask, head_mask = None):
+        """
+        Parameters
+        ----------
+        query: torch.tensor(bs, seq_length, dim)
+        key: torch.tensor(bs, seq_length, dim)
+        value: torch.tensor(bs, seq_length, dim)
+        mask: torch.tensor(bs, seq_length)
+
+        Outputs
+        -------
+        weights: torch.tensor(bs, n_heads, seq_length, seq_length)
+            Attention weights
+        context: torch.tensor(bs, seq_length, dim)
+            Contextualized layer. Optional: only if `output_attentions=True`
+        """
+        bs, q_length, dim = query.size()
+        k_length = key.size(1)
+        # assert dim == self.dim, 'Dimensions do not match: %s input vs %s configured' % (dim, self.dim)
+        # assert key.size() == value.size()
+
+        dim_per_head = self.dim // self.n_heads
+
+        assert 2 <= mask.dim() <= 3
+        causal = (mask.dim() == 3)
+        mask_reshp = (bs, 1, 1, k_length)
+
+        def shape(x):
+            """ separate heads """
+            return x.view(bs, -1, self.n_heads, dim_per_head).transpose(1, 2)
+
+        def unshape(x):
+            """ group heads """
+            return x.transpose(1, 2).contiguous().view(bs, -1, self.n_heads * dim_per_head)
+
+        q = shape(self.q_lin(query))           # (bs, n_heads, q_length, dim_per_head)
+        k = shape(self.k_lin(key))             # (bs, n_heads, k_length, dim_per_head)
+        v = shape(self.v_lin(value))           # (bs, n_heads, k_length, dim_per_head)
+
+        q = q / math.sqrt(dim_per_head)                     # (bs, n_heads, q_length, dim_per_head)
+        scores = torch.matmul(q, k.transpose(2,3))          # (bs, n_heads, q_length, k_length)
+        mask = (mask==0).view(mask_reshp).expand_as(scores) # (bs, n_heads, q_length, k_length)
+        scores.masked_fill_(mask, -float('inf'))            # (bs, n_heads, q_length, k_length)
+
+        weights = nn.Softmax(dim=-1)(scores)   # (bs, n_heads, q_length, k_length)
+        weights = self.dropout(weights)        # (bs, n_heads, q_length, k_length)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = torch.matmul(weights, v)     # (bs, n_heads, q_length, dim_per_head)
+        context = unshape(context)             # (bs, q_length, dim)
+        context = self.out_lin(context)        # (bs, q_length, dim)
+
+        if self.output_attentions:
+            return (context, weights)
+        else:
+            return (context,)
+
+class FFN(nn.Module):
+    def __init__(self, config):
+        super(FFN, self).__init__()
+        self.dropout = nn.Dropout(p=config.dropout)
+        self.lin1 = nn.Linear(in_features=config.dim, out_features=config.hidden_dim)
+        self.lin2 = nn.Linear(in_features=config.hidden_dim, out_features=config.dim)
+        assert config.activation in ['relu', 'gelu'], "activation ({}) must be in ['relu', 'gelu']".format(config.activation)
+        self.activation = gelu if config.activation == 'gelu' else nn.ReLU()
+
+    def forward(self, input):
+        x = self.lin1(input)
+        x = self.activation(x)
+        x = self.lin2(x)
+        x = self.dropout(x)
+        return x
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config):
+        super(TransformerBlock, self).__init__()
+
+        self.n_heads = config.n_heads
+        self.dim = config.dim
+        self.hidden_dim = config.hidden_dim
+        self.dropout = nn.Dropout(p=config.dropout)
+        self.activation = config.activation
+        self.output_attentions = config.output_attentions
+
+        assert config.dim % config.n_heads == 0
+
+        self.attention = MultiHeadSelfAttention(config)
+        self.sa_layer_norm = nn.LayerNorm(normalized_shape=config.dim, eps=1e-12)
+
+        self.ffn = FFN(config)
+        self.output_layer_norm = nn.LayerNorm(normalized_shape=config.dim, eps=1e-12)
+
+    def forward(self, x, attn_mask=None, head_mask=None):
+        """
+        Parameters
+        ----------
+        x: torch.tensor(bs, seq_length, dim)
+        attn_mask: torch.tensor(bs, seq_length)
+
+        Outputs
+        -------
+        sa_weights: torch.tensor(bs, n_heads, seq_length, seq_length)
+            The attention weights
+        ffn_output: torch.tensor(bs, seq_length, dim)
+            The output of the transformer block contextualization.
+        """
+        # Self-Attention
+        sa_output = self.attention(query=x, key=x, value=x, mask=attn_mask, head_mask=head_mask)
+        if self.output_attentions:
+            sa_output, sa_weights = sa_output                  # (bs, seq_length, dim), (bs, n_heads, seq_length, seq_length)
+        else: # To handle these `output_attention` or `output_hidden_states` cases returning tuples
+            assert type(sa_output) == tuple
+            sa_output = sa_output[0]
+        sa_output = self.sa_layer_norm(sa_output + x)          # (bs, seq_length, dim)
+
+        # Feed Forward Network
+        ffn_output = self.ffn(sa_output)                             # (bs, seq_length, dim)
+        ffn_output = self.output_layer_norm(ffn_output + sa_output)  # (bs, seq_length, dim)
+
+        output = (ffn_output,)
+        if self.output_attentions:
+            output = (sa_weights,) + output
+        return output
+
+
+class Transformer(nn.Module):
+    def __init__(self, config):
+        super(Transformer, self).__init__()
+        self.n_layers = config.n_layers
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+
+        layer = TransformerBlock(config)
+        self.layer = nn.ModuleList([copy.deepcopy(layer) for _ in range(config.n_layers)])
+
+    def forward(self, x, attn_mask=None, head_mask=None):
+        """
+        Parameters
+        ----------
+        x: torch.tensor(bs, seq_length, dim)
+            Input sequence embedded.
+        attn_mask: torch.tensor(bs, seq_length)
+            Attention mask on the sequence.
+
+        Outputs
+        -------
+        hidden_state: torch.tensor(bs, seq_length, dim)
+            Sequence of hiddens states in the last (top) layer
+        all_hidden_states: Tuple[torch.tensor(bs, seq_length, dim)]
+            Tuple of length n_layers with the hidden states from each layer.
+            Optional: only if output_hidden_states=True
+        all_attentions: Tuple[torch.tensor(bs, n_heads, seq_length, seq_length)]
+            Tuple of length n_layers with the attention weights from each layer
+            Optional: only if output_attentions=True
+        """
+        all_hidden_states = ()
+        all_attentions = ()
+
+        hidden_state = x
+        for i, layer_module in enumerate(self.layer):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_state,)
+
+            layer_outputs = layer_module(x=hidden_state,
+                                         attn_mask=attn_mask,
+                                         head_mask=head_mask[i])
+            hidden_state = layer_outputs[-1]
+
+            if self.output_attentions:
+                assert len(layer_outputs) == 2
+                attentions = layer_outputs[0]
+                all_attentions = all_attentions + (attentions,)
+            else:
+                assert len(layer_outputs) == 1
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_state,)
+
+        outputs = (hidden_state,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+### INTERFACE FOR ENCODER AND TASK SPECIFIC MODEL ###
+class DistilBertPreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for downloading and loading pretrained models.
+    """
+    config_class = DistilBertConfig
+    pretrained_model_archive_map = DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = None
+    base_model_prefix = "distilbert"
+
+    def __init__(self, *inputs, **kwargs):
+        super(DistilBertPreTrainedModel, self).__init__(*inputs, **kwargs)
+    
+    def init_weights(self, module):
+        """ Initialize the weights.
+        """
+        if isinstance(module, nn.Embedding):
+            if module.weight.requires_grad:
+                module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+DISTILBERT_START_DOCSTRING = r"""
+    DistilBERT is a small, fast, cheap and light Transformer model
+    trained by distilling Bert base. It has 40% less parameters than
+    `bert-base-uncased`, runs 60% faster while preserving over 95% of
+    Bert's performances as measured on the GLUE language understanding benchmark.
+
+    Here are the differences between the interface of Bert and DistilBert:
+
+    - DistilBert doesn't have `token_type_ids`, you don't need to indicate which token belongs to which segment. Just separate your segments with the separation token `tokenizer.sep_token` (or `[SEP]`)
+    - DistilBert doesn't have options to select the input positions (`position_ids` input). This could be added if necessary though, just let's us know if you need this option.
+
+    For more information on DistilBERT, please refer to our
+    `detailed blog post`_
+    
+    .. _`detailed blog post`:
+        https://medium.com/huggingface/distilbert-8cf3380435b5
+
+    Parameters:
+        config (:class:`~pytorch_transformers.DistilBertConfig`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+DISTILBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids** ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            The input sequences should start with `[CLS]` and end with `[SEP]` tokens.
+            
+            For now, ONLY BertTokenizer(`bert-base-uncased`) is supported and you should use this tokenizer when using DistilBERT.
+        **attention_mask**: (`optional`) ``torch.LongTensor`` 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]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare DistilBERT encoder/transformer outputing raw hidden-states without any specific head on top.",
+                      DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class DistilBertModel(DistilBertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        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
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(DistilBertModel, self).__init__(config)
+
+        self.embeddings = Embeddings(config)   # Embeddings
+        self.transformer = Transformer(config) # Encoder
+
+        self.apply(self.init_weights)
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        old_embeddings = self.embeddings.word_embeddings
+        new_embeddings = self._get_resized_embeddings(old_embeddings, new_num_tokens)
+        self.embeddings.word_embeddings = new_embeddings
+        return self.embeddings.word_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.transformer.layer[layer].attention.prune_heads(heads)
+
+    def forward(self,
+                input_ids, attention_mask=None, head_mask=None):
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids) # (bs, seq_length)
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_hidden_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_hidden_layers
+
+        embedding_output = self.embeddings(input_ids)   # (bs, seq_length, dim)
+        tfmr_output = self.transformer(x=embedding_output,
+                                       attn_mask=attention_mask,
+                                       head_mask=head_mask)
+        hidden_state = tfmr_output[0]
+        output = (hidden_state, ) + tfmr_output[1:]
+
+        return output # last-layer hidden-state, (all hidden_states), (all attentions)
+
+
+@add_start_docstrings("""DistilBert Model with a `masked language modeling` head on top. """,
+                      DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class DistilBertForMaskedLM(DistilBertPreTrainedModel):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        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
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(DistilBertForMaskedLM, self).__init__(config)
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+
+        self.distilbert = DistilBertModel(config)
+        self.vocab_transform = nn.Linear(config.dim, config.dim)
+        self.vocab_layer_norm = nn.LayerNorm(config.dim, eps=1e-12)
+        self.vocab_projector = nn.Linear(config.dim, config.vocab_size)
+
+        self.apply(self.init_weights)
+        self.tie_weights()
+
+        self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.vocab_projector,
+                                   self.distilbert.embeddings.word_embeddings)
+
+    def forward(self, input_ids, attention_mask=None, masked_lm_labels=None, head_mask=None):
+        dlbrt_output = self.distilbert(input_ids=input_ids,
+                                    attention_mask=attention_mask,
+                                    head_mask=head_mask)
+        hidden_states = dlbrt_output[0]                              # (bs, seq_length, dim)
+        prediction_logits = self.vocab_transform(hidden_states)      # (bs, seq_length, dim)
+        prediction_logits = gelu(prediction_logits)                  # (bs, seq_length, dim)
+        prediction_logits = self.vocab_layer_norm(prediction_logits) # (bs, seq_length, dim)
+        prediction_logits = self.vocab_projector(prediction_logits)  # (bs, seq_length, vocab_size)
+
+        outputs = (prediction_logits, ) + dlbrt_output[1:]
+        if masked_lm_labels is not None:
+            mlm_loss = self.mlm_loss_fct(prediction_logits.view(-1, prediction_logits.size(-1)),
+                                         masked_lm_labels.view(-1))
+            outputs = (mlm_loss,) + outputs     
+
+        return outputs # (mlm_loss), prediction_logits, (all hidden_states), (all attentions)
+
+
+@add_start_docstrings("""DistilBert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+                         the pooled output) e.g. for GLUE tasks. """,
+                      DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        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
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(DistilBertForSequenceClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.distilbert = DistilBertModel(config)
+        self.pre_classifier = nn.Linear(config.dim, config.dim)
+        self.classifier = nn.Linear(config.dim, config.num_labels)
+        self.dropout = nn.Dropout(config.seq_classif_dropout)
+
+        self.apply(self.init_weights)
+
+    def forward(self, input_ids,  attention_mask=None, labels=None, head_mask=None):
+        distilbert_output = self.distilbert(input_ids=input_ids,
+                                      attention_mask=attention_mask,
+                                      head_mask=head_mask)
+        hidden_state = distilbert_output[0]                    # (bs, seq_len, dim)
+        pooled_output = hidden_state[:, 0]                    # (bs, dim)
+        pooled_output = self.pre_classifier(pooled_output)   # (bs, dim)
+        pooled_output = nn.ReLU()(pooled_output)             # (bs, dim)
+        pooled_output = self.dropout(pooled_output)         # (bs, dim)
+        logits = self.classifier(pooled_output)              # (bs, dim)
+
+        outputs = (logits,) + distilbert_output[1:]
+        if labels is not None:
+            if self.num_labels == 1:
+                loss_fct = nn.MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""DistilBert Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear layers on top of
+                         the hidden-states output to compute `span start logits` and `span end logits`). """,
+                      DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
+    r"""
+        **start_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the start of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+        **end_positions**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for position (index) of the end of the labelled span for computing the token classification loss.
+            Positions are clamped to the length of the sequence (`sequence_length`).
+            Position outside of the sequence are not taken into account for computing the loss.
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
+        **start_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-start scores (before SoftMax).
+        **end_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length,)``
+            Span-end scores (before SoftMax).
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        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
+        start_positions = torch.tensor([1])
+        end_positions = torch.tensor([3])
+        outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
+        loss, start_scores, end_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(DistilBertForQuestionAnswering, self).__init__(config)
+
+        self.distilbert = DistilBertModel(config)
+        self.qa_outputs = nn.Linear(config.dim, config.num_labels)
+        assert config.num_labels == 2
+        self.dropout = nn.Dropout(config.qa_dropout)
+
+        self.apply(self.init_weights)
+        
+    def forward(self, input_ids, attention_mask=None, start_positions=None, end_positions=None, head_mask=None):
+        distilbert_output = self.distilbert(input_ids=input_ids,
+                                      attention_mask=attention_mask,
+                                      head_mask=head_mask)
+        hidden_states = distilbert_output[0]                                 # (bs, max_query_len, dim)
+
+        hidden_states = self.dropout(hidden_states)                       # (bs, max_query_len, dim)
+        logits = self.qa_outputs(hidden_states)                           # (bs, max_query_len, 2)
+        start_logits, end_logits = logits.split(1, dim=-1)
+        start_logits = start_logits.squeeze(-1)                           # (bs, max_query_len)
+        end_logits = end_logits.squeeze(-1)                               # (bs, max_query_len)
+
+        outputs = (start_logits, end_logits,) + distilbert_output[1:]
+        if start_positions is not None and end_positions is not None:
+            # If we are on multi-GPU, split add a dimension
+            if len(start_positions.size()) > 1:
+                start_positions = start_positions.squeeze(-1)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # sometimes the start/end positions are outside our model inputs, we ignore these terms
+            ignored_index = start_logits.size(1)
+            start_positions.clamp_(0, ignored_index)
+            end_positions.clamp_(0, ignored_index)
+
+            loss_fct = nn.CrossEntropyLoss(ignore_index=ignored_index)
+            start_loss = loss_fct(start_logits, start_positions)
+            end_loss = loss_fct(end_logits, end_positions)
+            total_loss = (start_loss + end_loss) / 2
+            outputs = (total_loss,) + outputs
+
+        return outputs  # (loss), start_logits, end_logits, (hidden_states), (attentions)

pytorch_transformers/modeling_gpt2.py

@@ -38,9 +38,11 @@ from .modeling_bert import BertLayerNorm as LayerNorm
 logger = logging.getLogger(__name__)
 
 GPT2_PRETRAINED_MODEL_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-pytorch_model.bin",
-                                     "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-pytorch_model.bin"}
+                                     "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-pytorch_model.bin",
+                                     "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-pytorch_model.bin"}
 GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP = {"gpt2": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-config.json",
-                                      "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-config.json"}
+                                      "gpt2-medium": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-medium-config.json",
+                                      "gpt2-large": "https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-large-config.json"}
 
 def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
     """ Load tf checkpoints in a pytorch model
@@ -50,7 +52,7 @@ def load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path):
         import numpy as np
         import tensorflow as tf
     except ImportError:
-        logger.error("Loading a TensorFlow models in PyTorch, requires TensorFlow to be installed. Please see "
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
             "https://www.tensorflow.org/install/ for installation instructions.")
         raise
     tf_path = os.path.abspath(gpt2_checkpoint_path)
@@ -137,7 +139,7 @@ class GPT2Config(PretrainedConfig):
         initializer_range=0.02,
 
         num_labels=1,
-        summary_type='token_ids',
+        summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
         summary_proj_to_labels=True,
@@ -383,17 +385,21 @@ GPT2_START_DOCSTRING = r"""    OpenAI GPT-2 model was proposed in
 
     Parameters:
         config (:class:`~pytorch_transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
 GPT2_INPUTS_DOCSTRING = r"""    Inputs:
         **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of input sequence tokens in the vocabulary.
+            GPT-2 is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
             Indices can be obtained using :class:`pytorch_transformers.BPT2Tokenizer`.
             See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
             :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
         **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of positions of each input sequence tokens in the position embeddings.
-            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
         **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
@@ -402,11 +408,7 @@ 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.Tensor`` 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.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -433,12 +435,11 @@ class GPT2Model(GPT2PreTrainedModel):
 
     Examples::
 
-        >>> config = GPT2Config.from_pretrained('gpt2')
-        >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
-        >>> model = GPT2Model(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
+        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
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
     """
     def __init__(self, config):
@@ -567,12 +568,11 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
 
     Examples::
 
-        >>> config = GPT2Config.from_pretrained('gpt2')
-        >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
-        >>> model = GPT2LMHeadModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=input_ids)
-        >>> loss, logits = outputs[:2]
+        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
+        outputs = model(input_ids, labels=input_ids)
+        loss, logits = outputs[:2]
 
     """
     def __init__(self, config):
@@ -614,7 +614,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
 @add_start_docstrings("""The GPT2 Model transformer with a language modeling and a multiple-choice classification
 head on top e.g. for RocStories/SWAG tasks. The two heads are two linear layers.
 The language modeling head has its weights tied to the input embeddings,
-the classification head takes as input the input of a specified classification token index in the intput sequence).
+the classification head takes as input the input of a specified classification token index in the input sequence).
 """, GPT2_START_DOCSTRING)
 class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
     r"""    Inputs:
@@ -629,7 +629,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             Selected in the range ``[0, input_ids.size(-1) - 1[``.
         **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             Indices of positions of each input sequence tokens in the position embeddings.
-            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
         **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
@@ -638,11 +638,7 @@ 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.Tensor`` 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.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -652,14 +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.
@@ -683,14 +676,14 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
 
     Examples::
 
-        >>> config = GPT2Config.from_pretrained('gpt2')
-        >>> tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
-        >>> model = GPT2DoubleHeadsModel(config)
-        >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]  # Assume you've added [CLS] to the vocabulary
-        >>> input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
-        >>> mc_token_ids = torch.tensor([-1, -1]).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, mc_token_ids)
-        >>> lm_prediction_scores, mc_prediction_scores = outputs[:2]
+        tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+        model = GPT2DoubleHeadsModel.from_pretrained('gpt2')
+        tokenizer.add_special_tokens({'cls_token': '[CLS]'})  # Add a [CLS] to the vocabulary (we should train it also!)
+        choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
+        input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        mc_token_ids = torch.tensor([input_ids.size(-1), input_ids.size(-1)]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, mc_token_ids)
+        lm_prediction_scores, mc_prediction_scores = outputs[:2]
 
     """
     def __init__(self, config):

pytorch_transformers/modeling_openai.py

@@ -171,7 +171,7 @@ class OpenAIGPTConfig(PretrainedConfig):
         predict_special_tokens=True,
 
         num_labels=1,
-        summary_type='token_ids',
+        summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
         summary_proj_to_labels=True,
@@ -397,26 +397,30 @@ OPENAI_GPT_START_DOCSTRING = r"""    OpenAI GPT model was proposed in
 
     Parameters:
         config (:class:`~pytorch_transformers.OpenAIGPTConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
 OPENAI_GPT_INPUTS_DOCSTRING = r"""    Inputs:
         **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of input sequence tokens in the vocabulary.
+            GPT is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
             Indices can be obtained using :class:`pytorch_transformers.BPT2Tokenizer`.
             See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
             :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
         **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of positions of each input sequence tokens in the position embeddings.
-            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
         **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **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.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -439,12 +443,11 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
     Examples::
 
-        >>> config = OpenAIGPTConfig.from_pretrained('openai-gpt')
-        >>> tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
-        >>> model = OpenAIGPTModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
+        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
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
     """
     def __init__(self, config):
@@ -538,7 +541,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Labels for language modeling.
-            Note that the labels **are shifted** inside the model, i.e. you can set ``lm_labels = input_ids``
+            Note that the labels **are shifted** inside the model, i.e. you can set ``labels = input_ids``
             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]``
@@ -558,12 +561,11 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
 
     Examples::
 
-        >>> config = OpenAIGPTConfig.from_pretrained('openai-gpt')
-        >>> tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
-        >>> model = OpenAIGPTLMHeadModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=input_ids)
-        >>> loss, logits = outputs[:2]
+        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
+        outputs = model(input_ids, labels=input_ids)
+        loss, logits = outputs[:2]
 
     """
     def __init__(self, config):
@@ -604,7 +606,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
 @add_start_docstrings("""OpenAI GPT Model transformer with a language modeling and a multiple-choice classification
 head on top e.g. for RocStories/SWAG tasks. The two heads are two linear layers.
 The language modeling head has its weights tied to the input embeddings,
-the classification head takes as input the input of a specified classification token index in the intput sequence).
+the classification head takes as input the input of a specified classification token index in the input sequence).
 """, OPENAI_GPT_START_DOCSTRING)
 class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
     r"""    Inputs:
@@ -619,16 +621,16 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
             Selected in the range ``[0, input_ids.size(-1) - 1[``.
         **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             Indices of positions of each input sequence tokens in the position embeddings.
-            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
         **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, num_choices, sequence_length)``:
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, num_choices, sequence_length)``:
+        **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.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -665,14 +667,14 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
 
     Examples::
 
-        >>> config = OpenAIGPTConfig.from_pretrained('openai-gpt')
-        >>> tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
-        >>> model = OpenAIGPTDoubleHeadsModel(config)
-        >>> choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]  # Assume you've added [CLS] to the vocabulary
-        >>> input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
-        >>> mc_token_ids = torch.tensor([-1, -1]).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, mc_token_ids)
-        >>> lm_prediction_scores, mc_prediction_scores = outputs[:2]
+        tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
+        model = OpenAIGPTDoubleHeadsModel.from_pretrained('openai-gpt')
+        tokenizer.add_special_tokens({'cls_token': '[CLS]'})  # Add a [CLS] to the vocabulary (we should train it also!)
+        choices = ["Hello, my dog is cute [CLS]", "Hello, my cat is cute [CLS]"]
+        input_ids = torch.tensor([tokenizer.encode(s) for s in choices]).unsqueeze(0)  # Batch size 1, 2 choices
+        mc_token_ids = torch.tensor([input_ids.size(-1), input_ids.size(-1)]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, mc_token_ids)
+        lm_prediction_scores, mc_prediction_scores = outputs[:2]
 
     """
     def __init__(self, config):

pytorch_transformers/modeling_roberta.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+"""PyTorch RoBERTa model. """
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from pytorch_transformers.modeling_bert import (BertConfig, BertEmbeddings,
+                                                BertLayerNorm, BertModel,
+                                                BertPreTrainedModel, gelu)
+
+from pytorch_transformers.modeling_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-pytorch_model.bin",
+    'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-pytorch_model.bin",
+    'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-pytorch_model.bin",
+}
+
+ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    'roberta-base': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-base-config.json",
+    'roberta-large': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-config.json",
+    'roberta-large-mnli': "https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-mnli-config.json",
+}
+
+
+class RobertaEmbeddings(BertEmbeddings):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+    def __init__(self, config):
+        super(RobertaEmbeddings, self).__init__(config)
+        self.padding_idx = 1
+
+    def forward(self, input_ids, token_type_ids=None, position_ids=None):
+        seq_length = input_ids.size(1)
+        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=input_ids.device)
+            position_ids = position_ids.unsqueeze(0).expand_as(input_ids)
+        return super(RobertaEmbeddings, self).forward(input_ids, token_type_ids=token_type_ids, position_ids=position_ids)
+
+
+class RobertaConfig(BertConfig):
+    pretrained_config_archive_map = ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+
+ROBERTA_START_DOCSTRING = r"""    The RoBERTa model was proposed in
+    `RoBERTa: A Robustly Optimized BERT Pretraining Approach`_
+    by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer,
+    Veselin Stoyanov. It is based on Google's BERT model released in 2018.
+    
+    It builds on BERT and modifies key hyperparameters, removing the next-sentence pretraining
+    objective and training with much larger mini-batches and learning rates.
+    
+    This implementation is the same as BertModel with a tiny embeddings tweak as well as a setup for Roberta pretrained 
+    models.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`RoBERTa: A Robustly Optimized BERT Pretraining Approach`:
+        https://arxiv.org/abs/1907.11692
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~pytorch_transformers.RobertaConfig`): Model configuration class with all the parameters of the 
+            model. Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+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 <s> and </s> tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         <s> Is this Jacksonville ? </s> </s> No it is not . </s>``
+
+            (b) For single sequences:
+
+                ``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``.
+
+            RoBERTa is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
+            See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
+            :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+        **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]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare RoBERTa Model transformer outputing raw hidden-states without any specific head on top.",
+                      ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class RobertaModel(BertModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``torch.FloatTensor`` of shape ``(batch_size, hidden_size)``
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during Bert pretraining. This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        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
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super(RobertaModel, self).__init__(config)
+
+        self.embeddings = RobertaEmbeddings(config)
+        self.apply(self.init_weights)
+
+    def forward(self, input_ids, token_type_ids=None, attention_mask=None, position_ids=None, head_mask=None):
+        if input_ids[:, 0].sum().item() != 0:
+            logger.warning("A sequence with no special tokens has been passed to the RoBERTa model. "
+                           "This model requires special tokens in order to work. "
+                           "Please specify add_special_tokens=True in your encoding.")
+        return super(RobertaModel, self).forward(input_ids, token_type_ids, attention_mask, position_ids, head_mask)
+
+
+@add_start_docstrings("""RoBERTa Model with a `language modeling` head on top. """,
+    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class RobertaForMaskedLM(BertPreTrainedModel):
+    r"""
+        **masked_lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``masked_lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        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
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super(RobertaForMaskedLM, self).__init__(config)
+
+        self.roberta = RobertaModel(config)
+        self.lm_head = RobertaLMHead(config)
+
+        self.apply(self.init_weights)
+        self.tie_weights()
+
+    def tie_weights(self):
+        """ Make sure we are sharing the input and output embeddings.
+            Export to TorchScript can't handle parameter sharing so we are cloning them instead.
+        """
+        self._tie_or_clone_weights(self.lm_head.decoder, self.roberta.embeddings.word_embeddings)
+
+    def forward(self, input_ids, token_type_ids=None, attention_mask=None, masked_lm_labels=None, position_ids=None,
+                head_mask=None):
+        outputs = self.roberta(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
+                            attention_mask=attention_mask, head_mask=head_mask)
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        outputs = (prediction_scores,) + outputs[2:]  # Add hidden states and attention if they are here
+
+        if masked_lm_labels is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), masked_lm_labels.view(-1))
+            outputs = (masked_lm_loss,) + outputs
+
+        return outputs  # (masked_lm_loss), prediction_scores, (hidden_states), (attentions)
+
+
+class RobertaLMHead(nn.Module):
+    """Roberta Head for masked language modeling."""
+
+    def __init__(self, config):
+        super(RobertaLMHead, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.layer_norm = BertLayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+    def forward(self, features, **kwargs):
+        x = self.dense(features)
+        x = gelu(x)
+        x = self.layer_norm(x)
+
+        # project back to size of vocabulary with bias
+        x = self.decoder(x) + self.bias
+
+        return x
+
+
+@add_start_docstrings("""RoBERTa Model transformer with a sequence classification/regression head on top (a linear layer 
+    on top of the pooled output) e.g. for GLUE tasks. """,
+    ROBERTA_START_DOCSTRING, ROBERTA_INPUTS_DOCSTRING)
+class RobertaForSequenceClassification(BertPreTrainedModel):
+    r"""
+        **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in ``[0, ..., config.num_labels]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        **logits**: ``torch.FloatTensor`` of shape ``(batch_size, config.num_labels)``
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        **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)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = RoertaTokenizer.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
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
+
+    """
+    config_class = RobertaConfig
+    pretrained_model_archive_map = ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super(RobertaForSequenceClassification, self).__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = RobertaModel(config)
+        self.classifier = RobertaClassificationHead(config)
+    
+    def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None,
+                position_ids=None, head_mask=None):
+        outputs = self.roberta(input_ids, position_ids=position_ids, token_type_ids=token_type_ids,
+                            attention_mask=attention_mask, head_mask=head_mask)
+        sequence_output = outputs[0]
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            outputs = (loss,) + outputs
+
+        return outputs  # (loss), logits, (hidden_states), (attentions)
+
+
+
+class RobertaClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super(RobertaClassificationHead, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x

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):
@@ -394,8 +394,8 @@ class MultiHeadAttn(nn.Module):
         self.pre_lnorm = pre_lnorm
 
         if r_r_bias is None or r_w_bias is None: # Biases are not shared
-            self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-            self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
+            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
         else:
             self.r_r_bias = r_r_bias
             self.r_w_bias = r_w_bias
@@ -483,8 +483,8 @@ class RelMultiHeadAttn(nn.Module):
         self.pre_lnorm = pre_lnorm
 
         if r_r_bias is None or r_w_bias is None: # Biases are not shared
-            self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-            self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
+            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
         else:
             self.r_r_bias = r_r_bias
             self.r_w_bias = r_w_bias
@@ -803,13 +803,13 @@ class AdaptiveEmbedding(nn.Module):
                 nn.Embedding(n_token, d_embed, sparse=sample_softmax>0)
             )
             if d_proj != d_embed:
-                self.emb_projs.append(nn.Parameter(torch.Tensor(d_proj, d_embed)))
+                self.emb_projs.append(nn.Parameter(torch.FloatTensor(d_proj, d_embed)))
         else:
             for i in range(len(self.cutoffs)):
                 l_idx, r_idx = self.cutoff_ends[i], self.cutoff_ends[i+1]
                 d_emb_i = d_embed // (div_val ** i)
                 self.emb_layers.append(nn.Embedding(r_idx-l_idx, d_emb_i))
-                self.emb_projs.append(nn.Parameter(torch.Tensor(d_proj, d_emb_i)))
+                self.emb_projs.append(nn.Parameter(torch.FloatTensor(d_proj, d_emb_i)))
 
     def forward(self, inp):
         if self.div_val == 1:
@@ -928,12 +928,16 @@ TRANSFO_XL_START_DOCSTRING = r"""    The Transformer-XL model was proposed in
 
     Parameters:
         config (:class:`~pytorch_transformers.TransfoXLConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
 TRANSFO_XL_INPUTS_DOCSTRING = r"""
     Inputs:
         **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of input sequence tokens in the vocabulary.
+            Transformer-XL is a model with relative position embeddings so you can either pad the inputs on
+            the right or on the left.
             Indices can be obtained using :class:`pytorch_transformers.TransfoXLTokenizer`.
             See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
             :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
@@ -941,7 +945,7 @@ TRANSFO_XL_INPUTS_DOCSTRING = r"""
             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` output below). Can be used to speed up sequential decoding and attend to longer context.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -968,12 +972,11 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
 
     Examples::
 
-        >>> config = TransfoXLConfig.from_pretrained('transfo-xl-wt103')
-        >>> tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
-        >>> model = TransfoXLModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids)
-        >>> last_hidden_states, mems = outputs[:2]
+        tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
+        model = TransfoXLModel.from_pretrained('transfo-xl-wt103')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states, mems = outputs[:2]
 
     """
     def __init__(self, config):
@@ -1003,8 +1006,8 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
         self.attn_type = config.attn_type
 
         if not config.untie_r:
-            self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-            self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
+            self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+            self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
 
         self.layers = nn.ModuleList()
         if config.attn_type == 0: # the default attention
@@ -1046,14 +1049,14 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
         if self.attn_type == 0: # default attention
             self.pos_emb = PositionalEmbedding(self.d_model)
         elif self.attn_type == 1: # learnable
-            self.r_emb = nn.Parameter(torch.Tensor(
+            self.r_emb = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head, self.d_head))
-            self.r_bias = nn.Parameter(torch.Tensor(
+            self.r_bias = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head))
         elif self.attn_type == 2: # absolute standard
             self.pos_emb = PositionalEmbedding(self.d_model)
         elif self.attn_type == 3: # absolute deeper SA
-            self.r_emb = nn.Parameter(torch.Tensor(
+            self.r_emb = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head, self.d_head))
 
         self.apply(self.init_weights)
@@ -1139,10 +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 = []
@@ -1284,12 +1287,11 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
 
     Examples::
 
-        >>> config = TransfoXLConfig.from_pretrained('transfo-xl-wt103')
-        >>> tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
-        >>> model = TransfoXLLMHeadModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids)
-        >>> prediction_scores, mems = outputs[:2]
+        tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
+        model = TransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores, mems = outputs[:2]
 
     """
     def __init__(self, config):

pytorch_transformers/modeling_transfo_xl_utilities.py

@@ -56,7 +56,7 @@ class ProjectedAdaptiveLogSoftmax(nn.Module):
             for i in range(len(self.cutoffs)):
                 if d_proj != d_embed:
                     self.out_projs.append(
-                        nn.Parameter(torch.Tensor(d_proj, d_embed))
+                        nn.Parameter(torch.FloatTensor(d_proj, d_embed))
                     )
                 else:
                     self.out_projs.append(None)
@@ -68,7 +68,7 @@ class ProjectedAdaptiveLogSoftmax(nn.Module):
                 d_emb_i = d_embed // (div_val ** i)
 
                 self.out_projs.append(
-                    nn.Parameter(torch.Tensor(d_proj, d_emb_i))
+                    nn.Parameter(torch.FloatTensor(d_proj, d_emb_i))
                 )
 
                 self.out_layers.append(nn.Linear(d_emb_i, r_idx-l_idx))

pytorch_transformers/modeling_utils.py

@@ -39,6 +39,20 @@ WEIGHTS_NAME = "pytorch_model.bin"
 TF_WEIGHTS_NAME = 'model.ckpt'
 
 
+try:
+    from torch.nn import Identity
+except ImportError:
+    # Older PyTorch compatibility
+    class Identity(nn.Module):
+        r"""A placeholder identity operator that is argument-insensitive.
+        """
+        def __init__(self, *args, **kwargs):
+            super(Identity, self).__init__()
+
+        def forward(self, input):
+            return input
+
+
 if not six.PY2:
     def add_start_docstrings(*docstr):
         def docstring_decorator(fn):
@@ -54,8 +68,22 @@ else:
 
 
 class PretrainedConfig(object):
-    """ Base class for all configuration classes.
-        Handle a few common parameters and methods for loading/downloading/saving configurations.
+    r""" Base class for all configuration classes.
+        Handles a few parameters common to all models' configurations as well as methods for loading/downloading/saving configurations.
+
+        Note:
+            A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to initialize a model does **not** load the model weights.
+            It only affects the model's configuration.
+
+        Class attributes (overridden by derived classes):
+            - ``pretrained_config_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained model configurations as values.
+
+        Parameters:
+            ``finetuning_task``: string, default `None`. Name of the task used to fine-tune the model. This can be used when converting from an original (TensorFlow or PyTorch) checkpoint.
+            ``num_labels``: integer, default `2`. Number of classes to use when the model is a classification model (sequences/tokens)
+            ``output_attentions``: boolean, default `False`. Should the model returns attentions weights.
+            ``output_hidden_states``: string, default `False`. Should the model returns all hidden-states.
+            ``torchscript``: string, default `False`. Is the model used with Torchscript.
     """
     pretrained_config_archive_map = {}
 
@@ -67,8 +95,8 @@ class PretrainedConfig(object):
         self.torchscript = kwargs.pop('torchscript', False)
 
     def save_pretrained(self, save_directory):
-        """ Save a configuration object to a directory, so that it
-            can be re-loaded using the `from_pretrained(save_directory)` class method.
+        """ Save a configuration object to the directory `save_directory`, so that it
+            can be re-loaded using the :func:`~pytorch_transformers.PretrainedConfig.from_pretrained` class method.
         """
         assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"
 
@@ -78,33 +106,56 @@ class PretrainedConfig(object):
         self.to_json_file(output_config_file)
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path, *input, **kwargs):
-        r""" Instantiate a PretrainedConfig from a pre-trained model configuration.
-
-        Params:
-            **pretrained_model_name_or_path**: either:
-                - a string with the `shortcut name` of a pre-trained model configuration to load from cache
-                    or download and cache if not already stored in cache (e.g. 'bert-base-uncased').
-                - a path to a `directory` containing a configuration file saved
-                    using the `save_pretrained(save_directory)` method.
-                - a path or url to a saved configuration `file`.
-            **cache_dir**: (`optional`) string:
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r""" Instantiate a :class:`~pytorch_transformers.PretrainedConfig` (or a derived class) from a pre-trained model configuration.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model configuration to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing a configuration file saved using the :func:`~pytorch_transformers.PretrainedConfig.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                - a path or url to a saved configuration JSON `file`, e.g.: ``./my_model_directory/configuration.json``.
+
+            cache_dir: (`optional`) string:
                 Path to a directory in which a downloaded pre-trained model
                 configuration should be cached if the standard cache should not be used.
-            **kwargs**: (`optional`) dict:
-                Dictionnary of key, values to update the configuration object after loading.
-                Can be used to override selected configuration parameters.
+
+            kwargs: (`optional`) dict: key/value pairs with which to update the configuration object after loading.
+
+                - The values in kwargs of any keys which are configuration attributes will be used to override the loaded values.
+                - Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled by the `return_unused_kwargs` keyword parameter.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            return_unused_kwargs: (`optional`) bool:
+
+                - If False, then this function returns just the final configuration object.
+                - If True, then this functions returns a tuple `(config, unused_kwargs)` where `unused_kwargs` is a dictionary consisting of the key/value pairs whose keys are not configuration attributes: ie the part of kwargs which has not been used to update `config` and is otherwise ignored.
 
         Examples::
 
-            >>> config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
-            >>> config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
-            >>> config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
-            >>> config = BertConfig.from_pretrained('bert-base-uncased', output_attention=True)
-            >>> assert config.output_attention == True
+            # We can't instantiate directly the base class `PretrainedConfig` so let's show the examples on a
+            # derived class: BertConfig
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            config = BertConfig.from_pretrained('./test/saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+            config = BertConfig.from_pretrained('./test/saved_model/my_configuration.json')
+            config = BertConfig.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
+            assert config.output_attention == True
+            config, unused_kwargs = BertConfig.from_pretrained('bert-base-uncased', output_attention=True,
+                                                               foo=False, return_unused_kwargs=True)
+            assert config.output_attention == True
+            assert unused_kwargs == {'foo': False}
 
         """
         cache_dir = kwargs.pop('cache_dir', None)
+        force_download = kwargs.pop('force_download', False)
+        proxies = kwargs.pop('proxies', None)
+        return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
 
         if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
             config_file = cls.pretrained_config_archive_map[pretrained_model_name_or_path]
@@ -114,8 +165,8 @@ class PretrainedConfig(object):
             config_file = pretrained_model_name_or_path
         # redirect to the cache, if necessary
         try:
-            resolved_config_file = cached_path(config_file, cache_dir=cache_dir)
-        except EnvironmentError:
+            resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+        except EnvironmentError as e:
             if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
                 logger.error(
                     "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
@@ -128,7 +179,7 @@ class PretrainedConfig(object):
                         pretrained_model_name_or_path,
                         ', '.join(cls.pretrained_config_archive_map.keys()),
                         config_file))
-            return None
+            raise e
         if resolved_config_file == config_file:
             logger.info("loading configuration file {}".format(config_file))
         else:
@@ -148,7 +199,10 @@ class PretrainedConfig(object):
             kwargs.pop(key, None)
 
         logger.info("Model config %s", config)
-        return config
+        if return_unused_kwargs:
+            return config, kwargs
+        else:
+            return config
 
     @classmethod
     def from_dict(cls, json_object):
@@ -187,14 +241,26 @@ class PretrainedConfig(object):
 
 
 class PreTrainedModel(nn.Module):
-    """ Base class for all models. Handle loading/storing model config and
-        a simple interface for dowloading and loading pretrained models.
+    r""" Base class for all models.
+
+        :class:`~pytorch_transformers.PreTrainedModel` takes care of storing the configuration of the models and handles methods for loading/downloading/saving models
+        as well as a few methods commons to all models to (i) resize the input embeddings and (ii) prune heads in the self-attention heads.
+
+        Class attributes (overridden by derived classes):
+            - ``config_class``: a class derived from :class:`~pytorch_transformers.PretrainedConfig` to use as configuration class for this model architecture.
+            - ``pretrained_model_archive_map``: a python ``dict`` of with `short-cut-names` (string) as keys and `url` (string) of associated pretrained weights as values.
+            - ``load_tf_weights``: a python ``method`` for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments:
+
+                - ``model``: an instance of the relevant subclass of :class:`~pytorch_transformers.PreTrainedModel`,
+                - ``config``: an instance of the relevant subclass of :class:`~pytorch_transformers.PretrainedConfig`,
+                - ``path``: a path (string) to the TensorFlow checkpoint.
+
+            - ``base_model_prefix``: a string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.
     """
-    config_class = PretrainedConfig
+    config_class = None
     pretrained_model_archive_map = {}
     load_tf_weights = lambda model, config, path: None
     base_model_prefix = ""
-    input_embeddings = None
 
     def __init__(self, config, *inputs, **kwargs):
         super(PreTrainedModel, self).__init__()
@@ -252,17 +318,16 @@ class PreTrainedModel(nn.Module):
 
     def resize_token_embeddings(self, new_num_tokens=None):
         """ Resize input token embeddings matrix of the model if new_num_tokens != config.vocab_size.
-            Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
+        Take care of tying weights embeddings afterwards if the model class has a `tie_weights()` method.
 
-        Args:
-            new_num_tokens: (`optional`) int
-                New number of tokens in the embedding matrix.
-                Increasing the size will add newly initialized vectors at the end
-                Reducing the size will remove vectors from the end
-                If not provided or None: does nothing and just returns a pointer to the input tokens Embedding Module of the model.
+        Arguments:
+
+            new_num_tokens: (`optional`) int:
+                New number of tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end. 
+                If not provided or None: does nothing and just returns a pointer to the input tokens ``torch.nn.Embeddings`` Module of the model.
 
         Return: ``torch.nn.Embeddings``
-            Pointer to the input tokens Embedding Module of the model
+            Pointer to the input tokens Embeddings Module of the model
         """
         base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
         model_embeds = base_model._resize_token_embeddings(new_num_tokens)
@@ -281,15 +346,17 @@ class PreTrainedModel(nn.Module):
 
     def prune_heads(self, heads_to_prune):
         """ Prunes heads of the base model.
-            Args:
-                heads_to_prune: dict of {layer_num (int): list of heads to prune in this layer (list of int)}
+
+            Arguments:
+
+                heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
         """
         base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
         base_model._prune_heads(heads_to_prune)
 
     def save_pretrained(self, save_directory):
-        """ Save a model with its configuration file to a directory, so that it
-            can be re-loaded using the `from_pretrained(save_directory)` class method.
+        """ Save a model and its configuration file to a directory, so that it
+            can be re-loaded using the `:func:`~pytorch_transformers.PreTrainedModel.from_pretrained`` class method.
         """
         assert os.path.isdir(save_directory), "Saving path should be a directory where the model and configuration can be saved"
 
@@ -305,61 +372,88 @@ class PreTrainedModel(nn.Module):
         torch.save(model_to_save.state_dict(), output_model_file)
 
     @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path, *inputs, **kwargs):
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         r"""Instantiate a pretrained pytorch model from a pre-trained model configuration.
 
-            The model is set in evaluation mode by default using `model.eval()` (Dropout modules are desactivated)
-            To train the model, you should first set it back in training mode with `model.train()`
-
-        Params:
-            **pretrained_model_name_or_path**: either:
-                - a string with the `shortcut name` of a pre-trained model to load from cache
-                    or download and cache if not already stored in cache (e.g. 'bert-base-uncased').
-                - a path to a `directory` containing a configuration file saved
-                    using the `save_pretrained(save_directory)` method.
-                - 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 option is slower than converting the TensorFlow
-                    checkpoint in a PyTorch model using the provided conversion scripts and loading
-                    the PyTorch model afterwards.
-            **config**: an optional configuration for the model to use instead of an automatically loaded configuation.
-                Configuration can be automatically loaded when:
-                - the model is a model provided by the library (loaded with a `shortcut name` of a pre-trained model), or
-                - the model was saved using the `save_pretrained(save_directory)` (loaded by suppling the save directory).
-            **state_dict**: an optional state dictionnary for the model to use instead of a state dictionary loaded
-                from saved weights file.
-                This option can be used if you want to create a model from a pretrained configuraton but load your own weights.
-                In this case though, you should check if using `save_pretrained(dir)` and `from_pretrained(save_directory)` is not
-                a simpler option.
-            **cache_dir**: (`optional`) string:
+        The model is set in evaluation mode by default using ``model.eval()`` (Dropout modules are deactivated)
+        To train the model, you should first set it back in training mode with ``model.train()``
+
+        The warning ``Weights from XXX not initialized from pretrained model`` means that the weights of XXX do not come pre-trained with the rest of the model.
+        It is up to you to train those weights with a downstream fine-tuning task.
+
+        The warning ``Weights from XXX not used in YYY`` means that the layer XXX is not used by YYY, therefore those weights are discarded.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a path to a `directory` containing model weights saved using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
+                - 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.
+
+            model_args: (`optional`) Sequence of positional arguments:
+                All remaning positional arguments will be passed to the underlying model's ``__init__`` method
+
+            config: (`optional`) instance of a class derived from :class:`~pytorch_transformers.PretrainedConfig`:
+                Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:
+
+                - the model is a model provided by the library (loaded with the ``shortcut-name`` string of a pretrained model), or
+                - the model was saved using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained` and is reloaded by suppling the save directory.
+                - the model is loaded by suppling a local directory as ``pretrained_model_name_or_path`` and a configuration JSON file named `config.json` is found in the directory.
+
+            state_dict: (`optional`) dict:
+                an optional state dictionnary for the model to use instead of a state dictionary loaded from saved weights file.
+                This option can be used if you want to create a model from a pretrained configuration but load your own weights.
+                In this case though, you should check if using :func:`~pytorch_transformers.PreTrainedModel.save_pretrained` and :func:`~pytorch_transformers.PreTrainedModel.from_pretrained` is not a simpler option.
+
+            cache_dir: (`optional`) string:
                 Path to a directory in which a downloaded pre-trained model
                 configuration should be cached if the standard cache should not be used.
-            **output_loading_info**: (`optional`) boolean:
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            output_loading_info: (`optional`) boolean:
                 Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages.
-            **kwargs**: (`optional`) dict:
-                Dictionnary of key, values to update the configuration object after loading.
-                Can be used to override selected configuration parameters. E.g. ``output_attention=True``
+
+            kwargs: (`optional`) Remaining dictionary of keyword arguments:
+                Can be used to update the configuration object (after it being loaded) and initiate the model. (e.g. ``output_attention=True``). Behave differently depending on whether a `config` is provided or automatically loaded:
+
+                - If a configuration is provided with ``config``, ``**kwargs`` will be directly passed to the underlying model's ``__init__`` method (we assume all relevant updates to the configuration have already been done)
+                - If a configuration is not provided, ``kwargs`` will be first passed to the configuration class initialization function (:func:`~pytorch_transformers.PretrainedConfig.from_pretrained`). Each key of ``kwargs`` that corresponds to a configuration attribute will be used to override said attribute with the supplied ``kwargs`` value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model's ``__init__`` function.
 
         Examples::
 
-            >>> model = BertModel.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
-            >>> model = BertModel.from_pretrained('./test/saved_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
-            >>> model = BertModel.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
-            >>> assert model.config.output_attention == True
-            >>> # Loading from a TF checkpoint file instead of a PyTorch model (slower)
-            >>> config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
-            >>> model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+            model = BertModel.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = BertModel.from_pretrained('./test/saved_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = BertModel.from_pretrained('bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
+            model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
         config = kwargs.pop('config', None)
         state_dict = kwargs.pop('state_dict', None)
         cache_dir = kwargs.pop('cache_dir', None)
         from_tf = kwargs.pop('from_tf', False)
+        force_download = kwargs.pop('force_download', False)
+        proxies = kwargs.pop('proxies', None)
         output_loading_info = kwargs.pop('output_loading_info', False)
 
         # Load config
         if config is None:
-            config = cls.config_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+            config, model_kwargs = cls.config_class.from_pretrained(
+                pretrained_model_name_or_path, *model_args,
+                cache_dir=cache_dir, return_unused_kwargs=True,
+                force_download=force_download,
+                **kwargs
+            )
+        else:
+            model_kwargs = kwargs
 
         # Load model
         if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
@@ -378,8 +472,8 @@ class PreTrainedModel(nn.Module):
                 archive_file = pretrained_model_name_or_path
         # redirect to the cache, if necessary
         try:
-            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir)
-        except EnvironmentError:
+            resolved_archive_file = cached_path(archive_file, cache_dir=cache_dir, force_download=force_download, proxies=proxies)
+        except EnvironmentError as e:
             if pretrained_model_name_or_path in cls.pretrained_model_archive_map:
                 logger.error(
                     "Couldn't reach server at '{}' to download pretrained weights.".format(
@@ -392,7 +486,7 @@ class PreTrainedModel(nn.Module):
                         pretrained_model_name_or_path,
                         ', '.join(cls.pretrained_model_archive_map.keys()),
                         archive_file))
-            return None
+            raise e
         if resolved_archive_file == archive_file:
             logger.info("loading weights file {}".format(archive_file))
         else:
@@ -400,7 +494,7 @@ class PreTrainedModel(nn.Module):
                 archive_file, resolved_archive_file))
 
         # Instantiate model.
-        model = cls(config)
+        model = cls(config, *model_args, **model_kwargs)
 
         if state_dict is None and not from_tf:
             state_dict = torch.load(resolved_archive_file, map_location='cpu')
@@ -530,7 +624,7 @@ class PoolerEndLogits(nn.Module):
             **start_states**: ``torch.LongTensor`` of shape identical to hidden_states
                 hidden states of the first tokens for the labeled span.
             **start_positions**: ``torch.LongTensor`` of shape ``(batch_size,)``
-                position of the first token for the labeled span: 
+                position of the first token for the labeled span:
             **p_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, seq_len)``
                 Mask of invalid position such as query and special symbols (PAD, SEP, CLS)
                 1.0 means token should be masked.
@@ -713,11 +807,11 @@ class SequenceSummary(nn.Module):
                 - 'last' => [default] take the last token hidden state (like XLNet)
                 - 'first' => take the first token hidden state (like Bert)
                 - 'mean' => take the mean of all tokens hidden states
-                - 'token_ids' => supply a Tensor of classification token indices (GPT/GPT-2)
+                - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2)
                 - 'attn' => Not implemented now, use multi-head attention
             summary_use_proj: Add a projection after the vector extraction
             summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False.
-            summary_activation: 'tanh' => add a tanh activation to the output, Other => no activation. Default 
+            summary_activation: 'tanh' => add a tanh activation to the output, Other => no activation. Default
             summary_first_dropout: Add a dropout before the projection and activation
             summary_last_dropout: Add a dropout after the projection and activation
     """
@@ -725,13 +819,13 @@ class SequenceSummary(nn.Module):
         super(SequenceSummary, self).__init__()
 
         self.summary_type = config.summary_type if hasattr(config, 'summary_use_proj') else 'last'
-        if config.summary_type == 'attn':
+        if self.summary_type == 'attn':
             # We should use a standard multi-head attention module with absolute positional embedding for that.
             # Cf. https://github.com/zihangdai/xlnet/blob/master/modeling.py#L253-L276
             # We can probably just use the multi-head attention module of PyTorch >=1.1.0
             raise NotImplementedError
 
-        self.summary = nn.Identity()
+        self.summary = Identity()
         if hasattr(config, 'summary_use_proj') and config.summary_use_proj:
             if hasattr(config, 'summary_proj_to_labels') and config.summary_proj_to_labels and config.num_labels > 0:
                 num_classes = config.num_labels
@@ -739,23 +833,23 @@ class SequenceSummary(nn.Module):
                 num_classes = config.hidden_size
             self.summary = nn.Linear(config.hidden_size, num_classes)
 
-        self.activation = nn.Identity()
+        self.activation = Identity()
         if hasattr(config, 'summary_activation') and config.summary_activation == 'tanh':
             self.activation = nn.Tanh()
 
-        self.first_dropout = nn.Identity()
+        self.first_dropout = Identity()
         if hasattr(config, 'summary_first_dropout') and config.summary_first_dropout > 0:
             self.first_dropout = nn.Dropout(config.summary_first_dropout)
 
-        self.last_dropout = nn.Identity()
+        self.last_dropout = Identity()
         if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0:
             self.last_dropout = nn.Dropout(config.summary_last_dropout)
 
-    def forward(self, hidden_states, token_ids=None):
+    def forward(self, hidden_states, cls_index=None):
         """ hidden_states: float Tensor in shape [bsz, seq_len, hidden_size], the hidden-states of the last layer.
-            token_ids: [optional] index of the classification token if summary_type == 'token_ids',
+            cls_index: [optional] position of the classification token if summary_type == 'cls_index',
                 shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states.
-                if summary_type == 'token_ids' and token_ids is None:
+                if summary_type == 'cls_index' and cls_index is None:
                     we take the last token of the sequence as classification token
         """
         if self.summary_type == 'last':
@@ -764,14 +858,14 @@ class SequenceSummary(nn.Module):
             output = hidden_states[:, 0]
         elif self.summary_type == 'mean':
             output = hidden_states.mean(dim=1)
-        elif self.summary_type == 'token_ids':
-            if token_ids is None:
-                token_ids = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2]-1, dtype=torch.long)
+        elif self.summary_type == 'cls_index':
+            if cls_index is None:
+                cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2]-1, dtype=torch.long)
             else:
-                token_ids = token_ids.unsqueeze(-1).unsqueeze(-1)
-                token_ids = token_ids.expand((-1,) * (token_ids.dim()-1) + (hidden_states.size(-1),))
-            # shape of token_ids: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
-            output = hidden_states.gather(-2, token_ids).squeeze(-2) # shape (bsz, XX, hidden_size)
+                cls_index = cls_index.unsqueeze(-1).unsqueeze(-1)
+                cls_index = cls_index.expand((-1,) * (cls_index.dim()-1) + (hidden_states.size(-1),))
+            # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states
+            output = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, XX, hidden_size)
         elif self.summary_type == 'attn':
             raise NotImplementedError
 

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):
@@ -416,27 +416,35 @@ XLM_START_DOCSTRING = r"""    The XLM model was proposed in
 
     Parameters:
         config (:class:`~pytorch_transformers.XLMConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
 XLM_INPUTS_DOCSTRING = r"""
     Inputs:
         **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of input sequence tokens in the vocabulary.
+
+            XLM is a model with absolute position embeddings so it's usually advised to pad the inputs on
+            the right rather than the left.
+
             Indices can be obtained using :class:`pytorch_transformers.XLMTokenizer`.
             See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
             :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
         **position_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of positions of each input sequence tokens in the position embeddings.
-            Selected in the range ``[0, config.max_position_embeddings - 1[``.
+            Selected in the range ``[0, config.max_position_embeddings - 1]``.
         **token_type_ids**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
         **langs**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             A parallel sequence of tokens to be used to indicate the language of each token in the input.
-            Indices are selected in the pre-trained language vocabulary,
-            i.e. in the range ``[0, config.n_langs - 1[``.
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices are languages ids which can be obtained from the language names by using two conversion mappings
+            provided in the configuration of the model (only provided for multilingual models).
+            More precisely, the `language name -> language id` mapping is in `model.config.lang2id` (dict str -> int) and
+            the `language id -> language name` mapping is `model.config.id2lang` (dict int -> str).
+        **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.
@@ -449,7 +457,7 @@ XLM_INPUTS_DOCSTRING = r"""
             hidden-states (key and values in the attention blocks) as computed by the model
             (see `cache` output below). Can be used to speed up sequential decoding.
             The dictionary object will be modified in-place during the forward pass to add newly computed hidden-states.
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -472,12 +480,11 @@ class XLMModel(XLMPreTrainedModel):
 
     Examples::
 
-        >>> config = XLMConfig.from_pretrained('xlm-mlm-en-2048')
-        >>> tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        >>> model = XLMModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMModel.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
 
     """
     ATTRIBUTES = ['encoder', 'eos_index', 'pad_index',  # 'with_output', 
@@ -745,12 +752,11 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
 
     Examples::
 
-        >>> config = XLMConfig.from_pretrained('xlm-mlm-en-2048')
-        >>> tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        >>> model = XLMWithLMHeadModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMWithLMHeadModel.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
 
     """
     def __init__(self, config):
@@ -786,7 +792,7 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the sequence classification/regression loss.
-            Indices should be in ``[0, ..., config.num_labels]``.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
 
@@ -805,14 +811,12 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
 
     Examples::
 
-        >>> config = XLMConfig.from_pretrained('xlm-mlm-en-2048')
-        >>> tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        >>> 
-        >>> model = XLMForSequenceClassification(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=labels)
-        >>> loss, logits = outputs[:2]
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMForSequenceClassification.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
 
     """
     def __init__(self, config):
@@ -885,15 +889,13 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
 
     Examples::
 
-        >>> config = XLMConfig.from_pretrained('xlm-mlm-en-2048')
-        >>> tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        >>> 
-        >>> model = XLMForQuestionAnswering(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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)
-        >>> loss, start_scores, end_scores = outputs[:2]
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMForQuestionAnswering.from_pretrained('xlm-mlm-en-2048')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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)
+        loss, start_scores, end_scores = outputs[:2]
 
     """
     def __init__(self, config):

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):
@@ -335,7 +335,7 @@ class XLNetConfig(PretrainedConfig):
 
 try:
     from apex.normalization.fused_layer_norm import FusedLayerNorm as XLNetLayerNorm
-except ImportError:
+except (ImportError, AttributeError) as e:
     logger.info("Better speed can be achieved with apex installed from https://www.github.com/nvidia/apex .")
     class XLNetLayerNorm(nn.Module):
         def __init__(self, d_model, eps=1e-12):
@@ -367,16 +367,16 @@ class XLNetRelativeAttention(nn.Module):
         self.d_model = config.d_model
         self.scale = 1 / (config.d_head ** 0.5)
 
-        self.q = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.k = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.v = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.o = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
-        self.r = nn.Parameter(torch.Tensor(config.d_model, self.n_head, self.d_head))
+        self.q = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.k = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.v = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.o = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
+        self.r = nn.Parameter(torch.FloatTensor(config.d_model, self.n_head, self.d_head))
 
-        self.r_r_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-        self.r_s_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-        self.r_w_bias = nn.Parameter(torch.Tensor(self.n_head, self.d_head))
-        self.seg_embed = nn.Parameter(torch.Tensor(2, self.n_head, self.d_head))
+        self.r_r_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+        self.r_s_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+        self.r_w_bias = nn.Parameter(torch.FloatTensor(self.n_head, self.d_head))
+        self.seg_embed = nn.Parameter(torch.FloatTensor(2, self.n_head, self.d_head))
 
         self.layer_norm = XLNetLayerNorm(config.d_model, eps=config.layer_norm_eps)
         self.dropout = nn.Dropout(config.dropout)
@@ -647,12 +647,16 @@ XLNET_START_DOCSTRING = r"""    The XLNet model was proposed in
 
     Parameters:
         config (:class:`~pytorch_transformers.XLNetConfig`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~pytorch_transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
 
 XLNET_INPUTS_DOCSTRING = r"""
     Inputs:
         **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
             Indices of input sequence tokens in the vocabulary.
+            XLNet is a model with relative position embeddings so you can either pad the inputs on
+            the right or on the left.
             Indices can be obtained using :class:`pytorch_transformers.XLNetTokenizer`.
             See :func:`pytorch_transformers.PreTrainedTokenizer.encode` and
             :func:`pytorch_transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
@@ -660,11 +664,11 @@ XLNET_INPUTS_DOCSTRING = r"""
             A parallel sequence of tokens (can be used to indicate various portions of the inputs).
             The embeddings from these tokens will be summed with the respective token embeddings.
             Indices are selected in the vocabulary (unlike BERT which has a specific vocabulary for segment indices).
-        **attention_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **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.
-        **input_mask**: (`optional`) ``torch.Tensor`` of shape ``(batch_size, sequence_length)``:
+        **input_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
             Mask to avoid performing attention on padding token indices.
             Negative of `attention_mask`, i.e. with 0 for real tokens and 1 for padding.
             Kept for compatibility with the original code base.
@@ -673,8 +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;
@@ -685,7 +692,7 @@ XLNET_INPUTS_DOCSTRING = r"""
             Mask to indicate the output tokens to use.
             If ``target_mapping[k, i, j] = 1``, the i-th predict in batch k is on the j-th token.
             Only used during pretraining for partial prediction or for sequential decoding (generation).
-        **head_mask**: (`optional`) ``torch.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+        **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]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
@@ -701,7 +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)``:
@@ -712,12 +720,11 @@ class XLNetModel(XLNetPreTrainedModel):
 
     Examples::
 
-        >>> config = XLNetConfig.from_pretrained('xlnet-large-cased')
-        >>> tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
-        >>> model = XLNetModel(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
+        tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
+        model = XLNetModel.from_pretrained('xlnet-large-cased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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
 
     """
     def __init__(self, config):
@@ -735,7 +742,7 @@ class XLNetModel(XLNetPreTrainedModel):
         self.n_layer = config.n_layer
 
         self.word_embedding = nn.Embedding(config.n_token, config.d_model)
-        self.mask_emb = nn.Parameter(torch.Tensor(1, 1, config.d_model))
+        self.mask_emb = nn.Parameter(torch.FloatTensor(1, 1, config.d_model))
         self.layer = nn.ModuleList([XLNetLayer(config) for _ in range(config.n_layer)])
         self.dropout = nn.Dropout(config.dropout)
 
@@ -856,7 +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
@@ -1008,7 +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)``:
@@ -1019,17 +1027,16 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
 
     Examples::
 
-        >>> config = XLNetConfig.from_pretrained('xlnet-large-cased')
-        >>> tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
-        >>> model = XLNetLMHeadModel(config)
-        >>> # We show how to setup inputs to predict a next token using a bi-directional context.
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>")).unsqueeze(0)  # We will predict the masked token
-        >>> perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float)
-        >>> perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
-        >>> target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float)  # Shape [1, 1, seq_length] => let's predict one token
-        >>> target_mapping[0, 0, -1] = 1.0  # Our first (and only) prediction will be the last token of the sequence (the masked token)
-        >>> outputs = model(input_ids, perm_mask=perm_mask, target_mapping=target_mapping)
-        >>> next_token_logits = outputs[0]  # Output has shape [target_mapping.size(0), target_mapping.size(1), config.vocab_size]
+        tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
+        model = XLNetLMHeadModel.from_pretrained('xlnet-large-cased')
+        # We show how to setup inputs to predict a next token using a bi-directional context.
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is very <mask>")).unsqueeze(0)  # We will predict the masked token
+        perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float)
+        perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
+        target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float)  # Shape [1, 1, seq_length] => let's predict one token
+        target_mapping[0, 0, -1] = 1.0  # Our first (and only) prediction will be the last token of the sequence (the masked token)
+        outputs = model(input_ids, perm_mask=perm_mask, target_mapping=target_mapping)
+        next_token_logits = outputs[0]  # Output has shape [target_mapping.size(0), target_mapping.size(1), config.vocab_size]
 
     """
     def __init__(self, config):
@@ -1077,7 +1084,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
     r"""
         **labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size,)``:
             Labels for computing the sequence classification/regression loss.
-            Indices should be in ``[0, ..., config.num_labels]``.
+            Indices should be in ``[0, ..., config.num_labels - 1]``.
             If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
             If ``config.num_labels > 1`` a classification loss is computed (Cross-Entropy).
 
@@ -1089,7 +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)``:
@@ -1100,14 +1108,12 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
 
     Examples::
 
-        >>> config = XLNetConfig.from_pretrained('xlnet-large-cased')
-        >>> tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
-        >>> 
-        >>> model = XLNetForSequenceClassification(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
-        >>> labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
-        >>> outputs = model(input_ids, labels=labels)
-        >>> loss, logits = outputs[:2]
+        tokenizer = XLNetTokenizer.from_pretrained('xlnet-large-cased')
+        model = XLNetForSequenceClassification.from_pretrained('xlnet-large-cased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+        loss, logits = outputs[:2]
 
     """
     def __init__(self, config):
@@ -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)``:
@@ -1200,15 +1207,13 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
 
     Examples::
 
-        >>> config = XLMConfig.from_pretrained('xlm-mlm-en-2048')
-        >>> tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
-        >>> 
-        >>> model = XLMForQuestionAnswering(config)
-        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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)
-        >>> loss, start_scores, end_scores = outputs[:2]
+        tokenizer = XLMTokenizer.from_pretrained('xlm-mlm-en-2048')
+        model = XLMForQuestionAnswering.from_pretrained('xlnet-large-cased')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).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)
+        loss, start_scores, end_scores = outputs[:2]
 
     """
     def __init__(self, config):

pytorch_transformers/optimization.py

@@ -36,13 +36,13 @@ class WarmupConstantSchedule(LambdaLR):
         Keeps learning rate schedule equal to 1. after warmup_steps.
     """
     def __init__(self, optimizer, warmup_steps, last_epoch=-1):
+        self.warmup_steps = warmup_steps
+        super(WarmupConstantSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
 
-        def lr_lambda(step):
-            if step < warmup_steps:
-                return float(step) / float(max(1.0, warmup_steps))
-            return 1.
-
-        super(WarmupConstantSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)
+    def lr_lambda(self, step):
+        if step < self.warmup_steps:
+            return float(step) / float(max(1.0, self.warmup_steps))
+        return 1.
 
 
 class WarmupLinearSchedule(LambdaLR):
@@ -51,13 +51,14 @@ class WarmupLinearSchedule(LambdaLR):
         Linearly decreases learning rate from 1. to 0. over remaining `t_total - warmup_steps` steps.
     """
     def __init__(self, optimizer, warmup_steps, t_total, last_epoch=-1):
+        self.warmup_steps = warmup_steps
+        self.t_total = t_total
+        super(WarmupLinearSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
 
-        def lr_lambda(step):
-            if step < warmup_steps:
-                return float(step) / float(max(1, warmup_steps))
-            return max(0.0, float(t_total - step) / float(max(1.0, t_total - warmup_steps)))
-
-        super(WarmupLinearSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)
+    def lr_lambda(self, step):
+        if step < self.warmup_steps:
+            return float(step) / float(max(1, self.warmup_steps))
+        return max(0.0, float(self.t_total - step) / float(max(1.0, self.t_total - self.warmup_steps)))
 
 
 class WarmupCosineSchedule(LambdaLR):
@@ -66,17 +67,19 @@ class WarmupCosineSchedule(LambdaLR):
         Decreases learning rate from 1. to 0. over remaining `t_total - warmup_steps` steps following a cosine curve.
         If `cycles` (default=0.5) is different from default, learning rate follows cosine function after warmup.
     """
-    warn_t_total = True
     def __init__(self, optimizer, warmup_steps, t_total, cycles=.5, last_epoch=-1):
+        self.warmup_steps = warmup_steps
+        self.t_total = t_total
+        self.cycles = cycles
+        super(WarmupCosineSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
 
-        def lr_lambda(step):
-            if step < warmup_steps:
-                return float(step) / float(max(1.0, warmup_steps))
-            else:
-                progress = float(step - warmup_steps) / float(max(1, t_total - warmup_steps))   # progress after warmup
-                return max(0.0, 0.5 * (1. + math.cos(math.pi * float(cycles) * 2.0 * progress)))
+    def lr_lambda(self, step):
+        if step < self.warmup_steps:
+            return float(step) / float(max(1.0, self.warmup_steps))
+        # progress after warmup
+        progress = float(step - self.warmup_steps) / float(max(1, self.t_total - self.warmup_steps))
+        return max(0.0, 0.5 * (1. + math.cos(math.pi * float(self.cycles) * 2.0 * progress)))
 
-        super(WarmupCosineSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)
 
 class WarmupCosineWithHardRestartsSchedule(LambdaLR):
     """ Linear warmup and then cosine cycles with hard restarts.
@@ -85,17 +88,20 @@ class WarmupCosineWithHardRestartsSchedule(LambdaLR):
         learning rate (with hard restarts).
     """
     def __init__(self, optimizer, warmup_steps, t_total, cycles=1., last_epoch=-1):
+        self.warmup_steps = warmup_steps
+        self.t_total = t_total
+        self.cycles = cycles
+        super(WarmupCosineWithHardRestartsSchedule, self).__init__(optimizer, self.lr_lambda, last_epoch=last_epoch)
+
+    def lr_lambda(self, step):
+        if step < self.warmup_steps:
+            return float(step) / float(max(1, self.warmup_steps))
+        # progress after warmup
+        progress = float(step - self.warmup_steps) / float(max(1, self.t_total - self.warmup_steps))
+        if progress >= 1.0:
+            return 0.0
+        return max(0.0, 0.5 * (1. + math.cos(math.pi * ((float(self.cycles) * progress) % 1.0))))
 
-        def lr_lambda(step):
-            if step < warmup_steps:
-                return float(step) / float(max(1, warmup_steps))
-            else:
-                progress = float(step - warmup_steps) / float(max(1, t_total - warmup_steps))   # progress after warmup
-                if progress >= 1.0:
-                    return 0.0
-                return max(0.0, 0.5 * (1. + math.cos(math.pi * ((float(cycles) * progress) % 1.0))))
-
-        super(WarmupCosineWithHardRestartsSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)
 
 
 class AdamW(Optimizer):

pytorch_transformers/tests/modeling_auto_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+import logging
+
+from pytorch_transformers import AutoConfig, BertConfig, AutoModel, BertModel
+from pytorch_transformers.modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+from .modeling_common_test import (CommonTestCases, ConfigTester, ids_tensor)
+
+
+class AutoModelTest(unittest.TestCase):
+    def test_model_from_pretrained(self):
+        logging.basicConfig(level=logging.INFO)
+        for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            config = AutoConfig.from_pretrained(model_name)
+            self.assertIsNotNone(config)
+            self.assertIsInstance(config, BertConfig)
+
+            model = AutoModel.from_pretrained(model_name)
+            model, loading_info = AutoModel.from_pretrained(model_name, output_loading_info=True)
+            self.assertIsNotNone(model)
+            self.assertIsInstance(model, BertModel)
+            for value in loading_info.values():
+                self.assertEqual(len(value), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()

pytorch_transformers/tests/modeling_common_test.py

@@ -21,6 +21,7 @@ import os
 import shutil
 import json
 import random
+import uuid
 
 import unittest
 import logging
@@ -48,6 +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()
@@ -158,6 +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
@@ -281,6 +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):
@@ -527,7 +535,7 @@ class ConfigTester(object):
 
     def create_and_test_config_to_json_file(self):
         config_first = self.config_class(**self.inputs_dict)
-        json_file_path = "/tmp/config.json"
+        json_file_path = os.path.join(os.getcwd(), "config_" + str(uuid.uuid4()) + ".json")
         config_first.to_json_file(json_file_path)
         config_second = self.config_class.from_json_file(json_file_path)
         os.remove(json_file_path)

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/modeling_roberta_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+import torch
+
+from pytorch_transformers import (RobertaConfig, RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification)
+from pytorch_transformers.modeling_roberta import ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+
+from .modeling_common_test import (CommonTestCases, ConfigTester, ids_tensor)
+
+
+class RobertaModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (RobertaForMaskedLM, RobertaModel)
+
+    class RobertaModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_token_type_ids=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     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)
+
+            token_type_ids = None
+            if self.use_token_type_ids:
+                token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+            sequence_labels = None
+            token_labels = None
+            choice_labels = None
+            if self.use_labels:
+                sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+                choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+            config = RobertaConfig(
+                vocab_size_or_config_json_file=self.vocab_size,
+                hidden_size=self.hidden_size,
+                num_hidden_layers=self.num_hidden_layers,
+                num_attention_heads=self.num_attention_heads,
+                intermediate_size=self.intermediate_size,
+                hidden_act=self.hidden_act,
+                hidden_dropout_prob=self.hidden_dropout_prob,
+                attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+                max_position_embeddings=self.max_position_embeddings,
+                type_vocab_size=self.type_vocab_size,
+                initializer_range=self.initializer_range)
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_roberta_model(self, config, input_ids, token_type_ids, input_mask, sequence_labels,
+                                           token_labels, choice_labels):
+            model = RobertaModel(config=config)
+            model.eval()
+            sequence_output, pooled_output = model(input_ids, token_type_ids, input_mask)
+            sequence_output, pooled_output = model(input_ids, token_type_ids)
+            sequence_output, pooled_output = model(input_ids)
+
+            result = {
+                "sequence_output": sequence_output,
+                "pooled_output": pooled_output,
+            }
+            self.parent.assertListEqual(
+                list(result["sequence_output"].size()),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(list(result["pooled_output"].size()), [self.batch_size, self.hidden_size])
+
+        def create_and_check_roberta_for_masked_lm(self, config, input_ids, token_type_ids, input_mask, sequence_labels,
+                                                   token_labels, choice_labels):
+            model = RobertaForMaskedLM(config=config)
+            model.eval()
+            loss, prediction_scores = model(input_ids, token_type_ids, input_mask, 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 prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, token_type_ids, input_mask,
+             sequence_labels, token_labels, choice_labels) = config_and_inputs
+            inputs_dict = {'input_ids': input_ids, 'token_type_ids': token_type_ids, 'attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = RobertaModelTest.RobertaModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=RobertaConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_roberta_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_roberta_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_roberta_for_masked_lm(*config_and_inputs)
+
+    @pytest.mark.slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/pytorch_transformers_test/"
+        for model_name in list(ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = RobertaModel.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+
+
+class RobertaModelIntegrationTest(unittest.TestCase):
+
+    @pytest.mark.slow
+    def test_inference_masked_lm(self):
+        model = RobertaForMaskedLM.from_pretrained('roberta-base')
+        
+        input_ids = torch.tensor([[    0, 31414,   232,   328,   740,  1140, 12695,    69, 46078,  1588,   2]])
+        output = model(input_ids)[0]
+        expected_shape = torch.Size((1, 11, 50265))
+        self.assertEqual(
+            output.shape,
+            expected_shape
+        )
+        # compare the actual values for a slice.
+        expected_slice = torch.Tensor(
+            [[[33.8843, -4.3107, 22.7779],
+              [ 4.6533, -2.8099, 13.6252],
+              [ 1.8222, -3.6898,  8.8600]]]
+        )
+        self.assertTrue(
+            torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3)
+        )
+
+    @pytest.mark.slow
+    def test_inference_no_head(self):
+        model = RobertaModel.from_pretrained('roberta-base')
+        
+        input_ids = torch.tensor([[    0, 31414,   232,   328,   740,  1140, 12695,    69, 46078,  1588,   2]])
+        output = model(input_ids)[0]
+        # compare the actual values for a slice.
+        expected_slice = torch.Tensor(
+            [[[-0.0231,  0.0782,  0.0074],
+              [-0.1854,  0.0539, -0.0174],
+              [ 0.0548,  0.0799,  0.1687]]]
+        )
+        self.assertTrue(
+            torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3)
+        )
+
+    @pytest.mark.slow
+    def test_inference_classification_head(self):
+        model = RobertaForSequenceClassification.from_pretrained('roberta-large-mnli')
+        
+        input_ids = torch.tensor([[    0, 31414,   232,   328,   740,  1140, 12695,    69, 46078,  1588,   2]])
+        output = model(input_ids)[0]
+        expected_shape = torch.Size((1, 3))
+        self.assertEqual(
+            output.shape,
+            expected_shape
+        )
+        expected_tensor = torch.Tensor([[-0.9469,  0.3913,  0.5118]])
+        self.assertTrue(
+            torch.allclose(output, expected_tensor, atol=1e-3)
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

pytorch_transformers/tests/optimization_test.py

@@ -17,13 +17,14 @@ from __future__ import division
 from __future__ import print_function
 
 import unittest
+import os
 
 import torch
 
 from pytorch_transformers import (AdamW, ConstantLRSchedule, WarmupConstantSchedule,
                                   WarmupCosineSchedule, WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)
 
-import numpy as np
+from .tokenization_tests_commons import TemporaryDirectory
 
 
 def unwrap_schedule(scheduler, num_steps=10):
@@ -33,6 +34,20 @@ def unwrap_schedule(scheduler, num_steps=10):
         lrs.append(scheduler.get_lr())
     return lrs
 
+def unwrap_and_save_reload_schedule(scheduler, num_steps=10):
+    lrs = []
+    for step in range(num_steps):
+        scheduler.step()
+        lrs.append(scheduler.get_lr())
+        if step == num_steps // 2:
+            with TemporaryDirectory() as tmpdirname:
+                file_name = os.path.join(tmpdirname, 'schedule.bin')
+                torch.save(scheduler.state_dict(), file_name)
+
+                state_dict = torch.load(file_name)
+                scheduler.load_state_dict(state_dict)
+    return lrs
+
 class OptimizationTest(unittest.TestCase):
 
     def assertListAlmostEqual(self, list1, list2, tol):
@@ -72,6 +87,10 @@ class ScheduleInitTest(unittest.TestCase):
         self.assertEqual(len(lrs[0]), 1)
         self.assertListEqual([l[0] for l in lrs], expected_learning_rates)
 
+        scheduler = ConstantLRSchedule(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)
         lrs = unwrap_schedule(scheduler, self.num_steps)
@@ -79,6 +98,10 @@ class ScheduleInitTest(unittest.TestCase):
         self.assertEqual(len(lrs[0]), 1)
         self.assertListEqual([l[0] for l in lrs], expected_learning_rates)
 
+        scheduler = WarmupConstantSchedule(self.optimizer, 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)
         lrs = unwrap_schedule(scheduler, self.num_steps)
@@ -86,6 +109,10 @@ class ScheduleInitTest(unittest.TestCase):
         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)
+        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)
         lrs = unwrap_schedule(scheduler, self.num_steps)
@@ -93,6 +120,10 @@ class ScheduleInitTest(unittest.TestCase):
         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)
+        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)
         lrs = unwrap_schedule(scheduler, self.num_steps)
@@ -100,6 +131,9 @@ class ScheduleInitTest(unittest.TestCase):
         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)
+        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()

pytorch_transformers/tests/tokenization_auto_test.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import pytest
+import logging
+
+from pytorch_transformers import AutoTokenizer, BertTokenizer, AutoTokenizer, GPT2Tokenizer
+from pytorch_transformers.modeling_bert import BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from pytorch_transformers.modeling_gpt2 import GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+class AutoTokenizerTest(unittest.TestCase):
+    def test_tokenizer_from_pretrained(self):
+        logging.basicConfig(level=logging.INFO)
+        for model_name in list(BERT_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            tokenizer = AutoTokenizer.from_pretrained(model_name)
+            self.assertIsNotNone(tokenizer)
+            self.assertIsInstance(tokenizer, BertTokenizer)
+            self.assertGreater(len(tokenizer), 0)
+
+        for model_name in list(GPT2_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            tokenizer = AutoTokenizer.from_pretrained(model_name)
+            self.assertIsNotNone(tokenizer)
+            self.assertIsInstance(tokenizer, GPT2Tokenizer)
+            self.assertGreater(len(tokenizer), 0)
+
+
+if __name__ == "__main__":
+    unittest.main()

pytorch_transformers/tests/tokenization_bert_test.py

@@ -24,30 +24,37 @@ from pytorch_transformers.tokenization_bert import (BasicTokenizer,
                                                     _is_control, _is_punctuation,
                                                     _is_whitespace, VOCAB_FILES_NAMES)
 
-from .tokenization_tests_commons import create_and_check_tokenizer_commons, TemporaryDirectory
+from .tokenization_tests_commons import CommonTestCases
 
-class TokenizationTest(unittest.TestCase):
+class BertTokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = BertTokenizer
+
+    def setUp(self):
+        super(BertTokenizationTest, self).setUp()
 
-    def test_full_tokenizer(self):
         vocab_tokens = [
             "[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn",
             "##ing", ",", "low", "lowest",
         ]
-        with TemporaryDirectory() as tmpdirname:
-            vocab_file = os.path.join(tmpdirname, VOCAB_FILES_NAMES['vocab_file'])
-            with open(vocab_file, "w", encoding='utf-8') as vocab_writer:
-                vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+        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]))
 
-            input_text = u"UNwant\u00E9d,running"
-            output_text = u"unwanted, running"
+    def get_tokenizer(self):
+        return self.tokenizer_class.from_pretrained(self.tmpdirname)
 
-            create_and_check_tokenizer_commons(self, input_text, output_text, BertTokenizer, tmpdirname)
+    def get_input_output_texts(self):
+        input_text = u"UNwant\u00E9d,running"
+        output_text = u"unwanted, running"
+        return input_text, output_text
 
-            tokenizer = BertTokenizer(vocab_file)
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file)
 
-            tokens = tokenizer.tokenize(u"UNwant\u00E9d,running")
-            self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
-            self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])
+        tokens = tokenizer.tokenize(u"UNwant\u00E9d,running")
+        self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])
 
     def test_chinese(self):
         tokenizer = BasicTokenizer()
@@ -118,6 +125,17 @@ class TokenizationTest(unittest.TestCase):
         self.assertFalse(_is_punctuation(u"A"))
         self.assertFalse(_is_punctuation(u" "))
 
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class.from_pretrained("bert-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/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/tests/tokenization_gpt2_test.py

@@ -20,42 +20,49 @@ import json
 
 from pytorch_transformers.tokenization_gpt2 import GPT2Tokenizer, VOCAB_FILES_NAMES
 
-from .tokenization_tests_commons import create_and_check_tokenizer_commons, TemporaryDirectory
+from .tokenization_tests_commons import CommonTestCases
 
-class GPT2TokenizationTest(unittest.TestCase):
+class GPT2TokenizationTest(CommonTestCases.CommonTokenizerTester):
 
-    def test_full_tokenizer(self):
-        """ Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt """
+    tokenizer_class = GPT2Tokenizer
+
+    def setUp(self):
+        super(GPT2TokenizationTest, self).setUp()
+
+        # Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt
         vocab = ["l", "o", "w", "e", "r", "s", "t", "i", "d", "n",
                  "lo", "low", "er",
                  "low", "lowest", "newer", "wider", "<unk>"]
         vocab_tokens = dict(zip(vocab, range(len(vocab))))
         merges = ["#version: 0.2", "l o", "lo w", "e r", ""]
-        special_tokens_map = {"unk_token": "<unk>"}
-
-        with TemporaryDirectory() as tmpdirname:
-            vocab_file = os.path.join(tmpdirname, VOCAB_FILES_NAMES['vocab_file'])
-            merges_file = os.path.join(tmpdirname, VOCAB_FILES_NAMES['merges_file'])
-            with open(vocab_file, "w") as fp:
-                fp.write(json.dumps(vocab_tokens))
-            with open(merges_file, "w") as fp:
-                fp.write("\n".join(merges))
-
-            input_text = u"lower newer"
-            output_text = u"lower<unk>newer"
-
-            create_and_check_tokenizer_commons(self, input_text, output_text, GPT2Tokenizer, tmpdirname, **special_tokens_map)
-
-            tokenizer = GPT2Tokenizer(vocab_file, merges_file, **special_tokens_map)
-            text = "lower"
-            bpe_tokens = ["low", "er"]
-            tokens = tokenizer.tokenize(text)
-            self.assertListEqual(tokens, bpe_tokens)
-
-            input_tokens = tokens + [tokenizer.unk_token]
-            input_bpe_tokens = [13, 12, 17]
-            self.assertListEqual(
-                tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+        self.special_tokens_map = {"unk_token": "<unk>"}
+
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES['vocab_file'])
+        self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES['merges_file'])
+        with open(self.vocab_file, "w") as fp:
+            fp.write(json.dumps(vocab_tokens))
+        with open(self.merges_file, "w") as fp:
+            fp.write("\n".join(merges))
+
+    def get_tokenizer(self):
+        return GPT2Tokenizer.from_pretrained(self.tmpdirname, **self.special_tokens_map)
+
+    def get_input_output_texts(self):
+        input_text = u"lower newer"
+        output_text = u"lower<unk>newer"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = GPT2Tokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
+        text = "lower"
+        bpe_tokens = ["low", "er"]
+        tokens = tokenizer.tokenize(text)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens + [tokenizer.unk_token]
+        input_bpe_tokens = [13, 12, 17]
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
 
 
 if __name__ == '__main__':