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

transformers/modeling_bert.py

@@ -48,6 +48,12 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-pytorch_model.bin",
     'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-pytorch_model.bin",
     'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-pytorch_model.bin",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-pytorch_model.bin",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-pytorch_model.bin",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-pytorch_model.bin",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-pytorch_model.bin",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/pytorch_model.bin",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/pytorch_model.bin",
 }
 
 
@@ -138,7 +144,11 @@ def swish(x):
     return x * torch.sigmoid(x)
 
 
-ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish, "gelu_new": gelu_new}
+def mish(x):
+    return x * torch.tanh(nn.functional.softplus(x))
+
+
+ACT2FN = {"gelu": gelu, "relu": torch.nn.functional.relu, "swish": swish, "gelu_new": gelu_new, "mish": mish}
 
 
 BertLayerNorm = torch.nn.LayerNorm
@@ -278,7 +288,7 @@ class BertAttention(nn.Module):
         if len(heads) == 0:
             return
         mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
-        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        heads = set(heads) - self.pruned_heads  # Convert to set and remove already pruned heads
         for head in heads:
             # Compute how many pruned heads are before the head and move the index accordingly
             head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
@@ -597,7 +607,7 @@ class BertModel(BertPreTrainedModel):
 
         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
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -656,8 +666,6 @@ class BertModel(BertPreTrainedModel):
 
         if attention_mask is None:
             attention_mask = torch.ones(input_shape, device=device)
-        if encoder_attention_mask is None:
-            encoder_attention_mask = torch.ones(input_shape, device=device)
         if token_type_ids is None:
             token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
@@ -665,18 +673,20 @@ class BertModel(BertPreTrainedModel):
         # ourselves in which case we just need to make it broadcastable to all heads.
         if attention_mask.dim() == 3:
             extended_attention_mask = attention_mask[:, None, :, :]
-
-        # Provided a padding mask of dimensions [batch_size, seq_length]
-        # - if the model is a decoder, apply a causal mask in addition to the padding mask
-        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
-        if attention_mask.dim() == 2:
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
             if self.config.is_decoder:
                 batch_size, seq_length = input_shape
                 seq_ids = torch.arange(seq_length, device=device)
                 causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = causal_mask.to(torch.long)  # not converting to long will cause errors with pytorch version < 1.3
                 extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
             else:
                 extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError("Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(input_shape, attention_mask.shape))
 
         # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
         # masked positions, this operation will create a tensor which is 0.0 for
@@ -688,13 +698,24 @@ class BertModel(BertPreTrainedModel):
 
         # If a 2D ou 3D attention mask is provided for the cross-attention
         # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
-        if encoder_attention_mask.dim() == 3:
-            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
-        if encoder_attention_mask.dim() == 2:
-            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+        if self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+
+            if encoder_attention_mask.dim() == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            elif encoder_attention_mask.dim() == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+            else:
+                raise ValueError("Wrong shape for encoder_hidden_shape (shape {}) or encoder_attention_mask (shape {})".format(encoder_hidden_shape,
+                                                                                                                               encoder_attention_mask.shape))
 
-        encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
-        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
+            encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
+        else:
+            encoder_extended_attention_mask = None
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head
@@ -760,7 +781,7 @@ class BertForPreTraining(BertPreTrainedModel):
 
         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
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, seq_relationship_scores = outputs[:2]
 
@@ -836,7 +857,7 @@ class BertForMaskedLM(BertPreTrainedModel):
 
         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
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -919,7 +940,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
 
         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
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         seq_relationship_scores = outputs[0]
 
@@ -984,7 +1005,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
 
         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
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -1060,7 +1081,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
         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
+        input_ids = torch.tensor([tokenizer.encode(s, add_special_tokens=True) 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]
@@ -1134,7 +1155,7 @@ class BertForTokenClassification(BertPreTrainedModel):
 
         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
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).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]
@@ -1218,9 +1239,9 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
         input_text = "[CLS] " + question + " [SEP] " + text + " [SEP]"
         input_ids = tokenizer.encode(input_text)
-        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))] 
+        token_type_ids = [0 if i <= input_ids.index(102) else 1 for i in range(len(input_ids))]
         start_scores, end_scores = model(torch.tensor([input_ids]), token_type_ids=torch.tensor([token_type_ids]))
-        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)  
+        all_tokens = tokenizer.convert_ids_to_tokens(input_ids)
         print(' '.join(all_tokens[torch.argmax(start_scores) : torch.argmax(end_scores)+1]))
         # a nice puppet
 

transformers/modeling_camembert.py

+# coding=utf-8
+# Copyright 2019 Inria, Facebook AI Research 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 CamemBERT model. """
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import logging
+
+from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForMultipleChoice, RobertaForTokenClassification
+from .configuration_camembert import CamembertConfig
+from .file_utils import add_start_docstrings
+
+logger = logging.getLogger(__name__)
+
+CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'camembert-base': "https://s3.amazonaws.com/models.huggingface.co/bert/camembert-base-pytorch_model.bin",
+}
+
+
+CAMEMBERT_START_DOCSTRING = r"""    The CamemBERT model was proposed in
+    `CamemBERT: a Tasty French Language Model`_
+    by Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suárez, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah, and Benoît Sagot. It is based on Facebook's RoBERTa model released in 2019.
+    
+    It is a model trained on 138GB of French text.
+    
+    This implementation is the same as RoBERTa.
+
+    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.
+
+    .. _`CamemBERT: a Tasty French Language Model`:
+        https://arxiv.org/abs/1911.03894
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.CamembertConfig`): 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:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+CAMEMBERT_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, CamemBERT 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 CamembertTokenizer.encode function with 
+            the ``add_special_tokens`` parameter set to ``True``.
+
+            CamemBERT 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:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **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.
+        **token_type_ids**: (`optional` need to be trained) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Optional segment token indices to indicate first and second portions of the inputs.
+            This embedding matrice is not trained (not pretrained during CamemBERT pretraining), you will have to train it
+            during finetuning.
+            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).
+        **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[``.
+        **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**.
+        **inputs_embeds**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, embedding_dim)``:
+            Optionally, instead of passing ``input_ids`` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+"""
+
+@add_start_docstrings("The bare CamemBERT Model transformer outputting raw hidden-states without any specific head on top.",
+                      CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertModel(RobertaModel):
+    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)
+            eo match pre-training, CamemBERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] [SEP] no it is not . [SEP]``
+
+                ``token_type_ids:   0   0  0    0    0     0       0   0   0     1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertModel.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).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 = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a `language modeling` head on top. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForMaskedLM(RobertaForMaskedLM):
+    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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForMaskedLM.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, masked_lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model transformer with a sequence classification/regression head on top (a linear layer 
+    on top of the pooled output) e.g. for GLUE tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForSequenceClassification(RobertaForSequenceClassification):
+    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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForSequenceClassification.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !")).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 = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a multiple choice classification head on top (a linear layer on top of
+    the pooled output and a softmax) e.g. for RocStories/SWAG tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForMultipleChoice(RobertaForMultipleChoice):
+    r"""
+    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 loss.
+        **classification_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)`` where `num_choices` is the size of the second dimension
+            of the input tensors. (see `input_ids` above).
+            Classification 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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForMultipleChoice.from_pretrained('camembert-base')
+        choices = ["J'aime le camembert !", "Je deteste le camembert !"]
+        input_ids = torch.tensor([tokenizer.encode(s, add_special_tokens=True) 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]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@add_start_docstrings("""CamemBERT Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    CAMEMBERT_START_DOCSTRING, CAMEMBERT_INPUTS_DOCSTRING)
+class CamembertForTokenClassification(RobertaForTokenClassification):
+    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 - 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,)``:
+            Classification loss.
+        **scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification 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 = CamembertTokenizer.from_pretrained('camembert-base')
+        model = CamembertForTokenClassification.from_pretrained('camembert-base')
+        input_ids = torch.tensor(tokenizer.encode("J'aime le camembert !", add_special_tokens=True)).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]
+
+    """
+    config_class = CamembertConfig
+    pretrained_model_archive_map = CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP

transformers/modeling_ctrl.py

@@ -63,7 +63,8 @@ def scaled_dot_product_attention(q, k, v, mask, attention_mask=None, head_mask=N
     scaled_attention_logits = matmul_qk / np.sqrt(dk)
 
     if mask is not None:
-        scaled_attention_logits += (mask * -1e4)
+        nd, ns = scaled_attention_logits.size(-2), scaled_attention_logits.size(-1)
+        scaled_attention_logits += (mask[ns-nd:ns, :ns] * -1e4)
 
     if attention_mask is not None:
         # Apply the attention mask
@@ -251,7 +252,7 @@ class CTRLModel(CTRLPreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -267,7 +268,7 @@ class CTRLModel(CTRLPreTrainedModel):
 
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLModel.from_pretrained('ctrl')
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -373,7 +374,7 @@ class CTRLModel(CTRLPreTrainedModel):
             inputs_embeds = self.w(input_ids)
         # inputs_embeds = embedded.unsqueeze(0) if len(input_ids.shape)<2 else embedded
         seq_len = input_shape[-1]
-        mask = torch.triu(torch.ones(seq_len, seq_len), 1).to(inputs_embeds.device)
+        mask = torch.triu(torch.ones(seq_len + past_length, seq_len + past_length), 1).to(inputs_embeds.device)
 
         inputs_embeds *= np.sqrt(self.d_model_size)
 
@@ -437,7 +438,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         **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).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -457,7 +458,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = CTRLLMHeadModel.from_pretrained('ctrl')
 
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 

transformers/modeling_distilbert.py

@@ -42,7 +42,9 @@ 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-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-pytorch_model.bin",
+    'distilbert-base-german-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-german-cased-pytorch_model.bin",
+    'distilbert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-multilingual-cased-pytorch_model.bin",
 }
 
 
@@ -413,7 +415,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertModel.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -509,7 +511,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForMaskedLM.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         loss, prediction_scores = outputs[:2]
 
@@ -579,7 +581,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=labels)
         loss, logits = outputs[:2]
@@ -654,7 +656,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
 
         tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
         model = DistilBertForQuestionAnswering.from_pretrained('distilbert-base-uncased')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         start_positions = torch.tensor([1])
         end_positions = torch.tensor([3])
         outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)

transformers/modeling_encoder_decoder.py

@@ -59,12 +59,14 @@ class PreTrainedEncoderDecoder(nn.Module):
             encoder_pretrained_model_name_or_path: information necessary to initiate the encoder. Either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/encoder``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
             decoder_pretrained_model_name_or_path: information necessary to initiate the decoder. Either:
 
                 - a string with the `shortcut name` of a pre-trained model to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/decoder``.
                 - a path or url to a `tensorflow index checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In this case, ``from_tf`` should be set to True and a configuration object should be provided as ``config`` argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.
 
@@ -164,7 +166,39 @@ class PreTrainedEncoderDecoder(nn.Module):
 
         We save the encoder' and decoder's parameters in two separate directories.
         """
+
+        # If the root output directory does not exist, create it 
+        if not os.path.exists(save_directory):
+            os.mkdir(save_directory)
+
+        # Check whether the output directory is empty or not
+        sub_directories = [directory for directory in os.listdir(save_directory)
+            if os.path.isdir(os.path.join(save_directory, directory))]
+
+        if len(sub_directories) > 0:
+            if "encoder" in sub_directories and "decoder" in sub_directories:
+                print("WARNING: there is an older version of encoder-decoder saved in" +\
+                    " the output directory. The default behaviour is to overwrite them.")
+
+            # Empty the output directory
+            for directory_to_remove in sub_directories:
+                # Remove all files into the subdirectory
+                files_to_remove = os.listdir(os.path.join(save_directory, directory_to_remove))
+                for file_to_remove in files_to_remove:
+                    os.remove(os.path.join(save_directory, directory_to_remove, file_to_remove))
+                # Remove the subdirectory itself
+                os.rmdir(os.path.join(save_directory, directory_to_remove))
+
+            assert(len(os.listdir(save_directory)) == 0) # sanity check
+
+        # Create the "encoder" directory inside the output directory and save the encoder into it
+        if not os.path.exists(os.path.join(save_directory, "encoder")):
+            os.mkdir(os.path.join(save_directory, "encoder"))
         self.encoder.save_pretrained(os.path.join(save_directory, "encoder"))
+
+        # Create the "encoder" directory inside the output directory and save the decoder into it
+        if not os.path.exists(os.path.join(save_directory, "decoder")):
+            os.mkdir(os.path.join(save_directory, "decoder"))
         self.decoder.save_pretrained(os.path.join(save_directory, "decoder"))
 
     def forward(self, encoder_input_ids, decoder_input_ids, **kwargs):
@@ -217,9 +251,7 @@ class PreTrainedEncoderDecoder(nn.Module):
         encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
         if encoder_hidden_states is None:
             encoder_outputs = self.encoder(encoder_input_ids, **kwargs_encoder)
-            encoder_hidden_states = encoder_outputs[
-                0
-            ]  # output the last layer hidden state
+            encoder_hidden_states = encoder_outputs[0]
         else:
             encoder_outputs = ()
 

transformers/modeling_gpt2.py

@@ -329,7 +329,7 @@ class GPT2Model(GPT2PreTrainedModel):
         **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -345,7 +345,7 @@ class GPT2Model(GPT2PreTrainedModel):
 
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2Model.from_pretrained('gpt2')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -503,7 +503,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         **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).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -523,7 +523,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = GPT2LMHeadModel.from_pretrained('gpt2')
 
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -596,7 +596,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         **mc_prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, num_choices)``
             Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
         **past**:
-            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model 
             should not be passed as input ids as they have already been computed.
@@ -634,6 +634,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
     """
     def __init__(self, config):
         super(GPT2DoubleHeadsModel, self).__init__(config)
+        config.num_labels = 1
         self.transformer = GPT2Model(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)

transformers/modeling_openai.py

@@ -50,8 +50,10 @@ def load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path):
 
     logger.info("Loading weights from {}".format(openai_checkpoint_folder_path))
 
-    names = json.load(open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8'))
-    shapes = json.load(open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8'))
+    with open(openai_checkpoint_folder_path + '/parameters_names.json', "r", encoding='utf-8') as names_handle:
+        names = json.load(names_handle)
+    with open(openai_checkpoint_folder_path + '/params_shapes.json', "r", encoding='utf-8') as shapes_handle:
+        shapes = json.load(shapes_handle)
     offsets = np.cumsum([np.prod(shape) for shape in shapes])
     init_params = [np.load(openai_checkpoint_folder_path + '/params_{}.npy'.format(n)) for n in range(10)]
     init_params = np.split(np.concatenate(init_params, 0), offsets)[:-1]
@@ -347,7 +349,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -489,7 +491,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = OpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
-        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 
@@ -588,6 +590,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
     def __init__(self, config):
         super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
 
+        config.num_labels = 1
         self.transformer = OpenAIGPTModel(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)

transformers/modeling_roberta.py

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

transformers/modeling_t5.py

+# coding=utf-8
+# Copyright 2018 Mesh TensorFlow authors, T5 Authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch T5 model. """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+import copy
+import itertools
+from io import open
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from .modeling_utils import PreTrainedModel, prune_linear_layer
+from .configuration_t5 import T5Config
+from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
+
+logger = logging.getLogger(__name__)
+
+####################################################
+# This dict contrains shortcut names and associated url
+# for the pretrained weights provided with the models
+####################################################
+T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-pytorch_model.bin",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-pytorch_model.bin",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-pytorch_model.bin",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-pytorch_model.bin",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-pytorch_model.bin",
+}
+
+####################################################
+# This is a conversion method from TF 1.0 to PyTorch
+# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
+####################################################
+def load_tf_weights_in_t5(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model.
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        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)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    tf_weights = {}
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        tf_weights[name] = array
+
+    for txt_name in names:
+        name = txt_name.split('/')
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
+            logger.info("Skipping {}".format("/".join(name)))
+            tf_weights.pop(txt_name, None)
+            continue
+        if '_slot_' in name[-1]:
+            logger.info("Skipping {}".format("/".join(name)))
+            tf_weights.pop(txt_name, None)
+            continue
+        pointer = model
+        array = tf_weights[txt_name]
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+            if l[0] in ['kernel', 'scale', 'embedding']:
+                pointer = getattr(pointer, 'weight')
+            # elif l[0] == 'scale':
+            #     pointer = getattr(pointer, 'weight')
+            # elif l[0] == 'output_bias' or l[0] == 'beta':
+            #     pointer = getattr(pointer, 'bias')
+            # elif l[0] == 'squad':
+            #     pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+        if l[0] not in ['kernel', 'scale', 'embedding']:
+            pointer = getattr(pointer, 'weight')
+        if l[0] != 'embedding':
+            logger.info("Transposing numpy weight of shape {} for {}".format(array.shape, name))
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array.astype(np.float32))
+        tf_weights.pop(txt_name, None)
+
+    logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
+    # logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
+    return model
+
+
+####################################################
+# PyTorch Models are constructed by sub-classing
+# - torch.nn.Module for the layers and
+# - PreTrainedModel for the models (it-self a sub-class of torch.nn.Module)
+####################################################
+
+class T5LayerNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """ Construct a layernorm module in the T5 style
+            No bias and no substraction of mean.
+        """
+        super(T5LayerNorm, self).__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, x):
+        variance = x.pow(2).mean(-1, keepdim=True)
+        x = x / torch.sqrt(variance + self.variance_epsilon)
+        return self.weight * x
+
+
+class T5DenseReluDense(nn.Module):
+    def __init__(self, config):
+        super(T5DenseReluDense, self).__init__()
+        self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states):
+        h = self.wi(hidden_states)
+        h = F.relu(h)
+        h = self.dropout(h)
+        h = self.wo(h)
+        return h
+
+
+class T5LayerFF(nn.Module):
+    def __init__(self, config):
+        super(T5LayerFF, self).__init__()
+        self.DenseReluDense = T5DenseReluDense(config)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states):
+        norm_x = self.layer_norm(hidden_states)
+        y = self.DenseReluDense(norm_x)
+        layer_output = hidden_states + self.dropout(y)
+        return layer_output
+
+
+class T5Attention(nn.Module):
+    NEW_ID = itertools.count()
+
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5Attention, self).__init__()
+        self.layer_id = next(T5Attention.NEW_ID)
+        self.is_decoder = config.is_decoder
+        self.has_relative_attention_bias = has_relative_attention_bias
+
+        self.output_attentions = config.output_attentions
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.d_model = config.d_model
+        self.d_kv = config.d_kv
+        self.n_heads = config.num_heads
+        self.dropout = config.dropout_rate
+        self.inner_dim = self.n_heads * self.d_kv
+
+        # Mesh TensorFlow initialization to avoid scaling before softmax
+        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
+
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.n_heads, self.d_kv)
+        heads = set(heads) - self.pruned_heads
+        for head in heads:
+            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+        # Prune linear layers
+        self.q = prune_linear_layer(self.q, index)
+        self.k = prune_linear_layer(self.k, index)
+        self.v = prune_linear_layer(self.v, index)
+        self.o = prune_linear_layer(self.o, index, dim=1)
+        # Update hyper params
+        self.n_heads = self.n_heads - len(heads)
+        self.inner_dim = self.d_kv * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    @staticmethod
+    def _relative_position_bucket(relative_position,
+                                  bidirectional=True,
+                                  num_buckets=32,
+                                  max_distance=128):
+        """
+        Adapted from Mesh Tensorflow:
+        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+        Translate relative position to a bucket number for relative attention.
+        The relative position is defined as memory_position - query_position, i.e.
+        the distance in tokens from the attending position to the attended-to
+        position.  If bidirectional=False, then positive relative positions are
+        invalid.
+        We use smaller buckets for small absolute relative_position and larger buckets
+        for larger absolute relative_positions.  All relative positions >=max_distance
+        map to the same bucket.  All relative positions <=-max_distance map to the
+        same bucket.  This should allow for more graceful generalization to longer
+        sequences than the model has been trained on.
+        Args:
+            relative_position: an int32 Tensor
+            bidirectional: a boolean - whether the attention is bidirectional
+            num_buckets: an integer
+            max_distance: an integer
+        Returns:
+            a Tensor with the same shape as relative_position, containing int32
+            values in the range [0, num_buckets)
+        """
+        ret = 0
+        n = -relative_position
+        if bidirectional:
+            num_buckets //= 2
+            ret += (n < 0).to(torch.long) * num_buckets  # mtf.to_int32(mtf.less(n, 0)) * num_buckets
+            n = torch.abs(n)
+        else:
+            n = torch.max(n, torch.zeros_like(n))
+        # now n is in the range [0, inf)
+
+        # half of the buckets are for exact increments in positions
+        max_exact = num_buckets // 2
+        is_small = (n < max_exact)
+
+        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
+        val_if_large = max_exact + (
+            torch.log(n.float() / max_exact)
+            / math.log(max_distance / max_exact) * (num_buckets - max_exact)).to(torch.long)
+        val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))
+
+        ret += torch.where(is_small, n, val_if_large)
+        return ret
+
+    def compute_bias(self, qlen, klen):
+        """ Compute binned relative position bias """
+        context_position = torch.arange(qlen, dtype=torch.long)[:, None]
+        memory_position = torch.arange(klen, dtype=torch.long)[None, :]
+        relative_position = memory_position - context_position  # shape (qlen, klen)
+        rp_bucket = self._relative_position_bucket(relative_position,  # shape (qlen, klen)
+                                                   bidirectional=not self.is_decoder,
+                                                   num_buckets=self.relative_attention_num_buckets)
+        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
+        values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, qlen, klen)
+        return values
+
+    def forward(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None):
+        """
+        Self-attention (if kv is None) or attention over source sentence (provided by kv).
+        """
+        # Input is (bs, qlen, dim)
+        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
+        bs, qlen, dim = input.size()
+        if kv is None:
+            klen = qlen if cache is None else cache['slen'] + qlen
+        else:
+            klen = kv.size(1)
+
+        def shape(x):
+            """  projection """
+            return x.view(bs, -1, self.n_heads, self.d_kv).transpose(1, 2)
+
+        def unshape(x):
+            """  compute context """
+            return x.transpose(1, 2).contiguous().view(bs, -1, self.inner_dim)
+
+        q = shape(self.q(input))                                          # (bs, n_heads, qlen, dim_per_head)
+        if kv is None:
+            k = shape(self.k(input))                                      # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(input))                                      # (bs, n_heads, qlen, dim_per_head)
+        elif cache is None or self.layer_id not in cache:
+            k = v = kv
+            k = shape(self.k(k))                                          # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(v))                                          # (bs, n_heads, qlen, dim_per_head)
+
+        if cache is not None:
+            if self.layer_id in cache:
+                if kv is None:
+                    k_, v_ = cache[self.layer_id]
+                    k = torch.cat([k_, k], dim=2)                             # (bs, n_heads, klen, dim_per_head)
+                    v = torch.cat([v_, v], dim=2)                             # (bs, n_heads, klen, dim_per_head)
+                else:
+                    k, v = cache[self.layer_id]
+            cache[self.layer_id] = (k, v)
+
+        # q = q / math.sqrt(dim_per_head)                                     # No scaling in T5
+        scores = torch.einsum('bnqd,bnkd->bnqk', q, k)                        # (bs, n_heads, qlen, klen)
+
+        if position_bias is None:
+            if not self.has_relative_attention_bias:
+                raise ValueError("No position_bias provided and no weights to compute position_bias")
+            position_bias = self.compute_bias(qlen, klen)
+            if mask is not None:
+                position_bias = position_bias + mask                          # (bs, n_heads, qlen, klen)
+
+        scores += position_bias
+        weights = F.softmax(scores.float(), dim=-1).type_as(scores)           # (bs, n_heads, qlen, klen)
+        weights = F.dropout(weights, p=self.dropout, training=self.training)  # (bs, n_heads, qlen, klen)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = torch.matmul(weights, v)                                    # (bs, n_heads, qlen, dim_per_head)
+        context = unshape(context)                                            # (bs, qlen, dim)
+
+        context = self.o(context)
+
+        outputs = (context,)
+        if self.output_attentions:
+            outputs = outputs + (weights,)
+        if self.has_relative_attention_bias:
+            outputs = outputs + (position_bias,)
+        return outputs
+
+
+class T5LayerSelfAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5LayerSelfAttention, self).__init__()
+        self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states, attention_mask=None, position_bias=None, head_mask=None):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.SelfAttention(norm_x,
+                                              mask=attention_mask,
+                                              position_bias=position_bias,
+                                              head_mask=head_mask)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5LayerCrossAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5LayerCrossAttention, self).__init__()
+        self.EncDecAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states, kv, attention_mask=None, position_bias=None, head_mask=None):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.EncDecAttention(norm_x,
+                                                mask=attention_mask,
+                                                kv=kv,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5Block(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5Block, self).__init__()
+        self.is_decoder = config.is_decoder
+        self.layer = nn.ModuleList()
+        self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+        if self.is_decoder:
+            self.layer.append(T5LayerCrossAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+            self.layer.append(T5LayerFF(config))
+        else:
+            self.layer.append(T5LayerFF(config))
+
+    def forward(self, hidden_states, attention_mask=None, position_bias=None,
+                encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
+                head_mask=None):
+        self_attention_outputs = self.layer[0](hidden_states,
+                                                attention_mask=attention_mask,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask)
+        hidden_states = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # Keep self-attention outputs and relative position weights
+
+        if not self.is_decoder:
+            hidden_states = self.layer[1](hidden_states)
+        else:
+            cross_attention_outputs = self.layer[1](hidden_states,
+                                                    kv=encoder_hidden_states,
+                                                    attention_mask=encoder_attention_mask,
+                                                    position_bias=encoder_decoder_position_bias,
+                                                    head_mask=head_mask)
+            hidden_states = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]  # Keep cross-attention outputs and relative position weights
+            hidden_states = self.layer[2](hidden_states)
+
+        outputs = (hidden_states,) + outputs  # add attentions if we output them
+        return outputs  # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+
+
+class T5PreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = T5Config
+    pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_t5
+    base_model_prefix = "transformer"
+
+    @property
+    def dummy_inputs(self):
+        input_ids = torch.tensor(DUMMY_INPUTS)
+        input_mask = torch.tensor(DUMMY_MASK)
+        dummy_inputs = {'decoder_input_ids': input_ids,
+                        'encoder_input_ids': input_ids,
+                        'decoder_attention_mask': input_mask}
+        return dummy_inputs
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        factor = self.config.initializer_factor  # Used for testing weights initialization
+        if isinstance(module, T5LayerNorm):
+            module.weight.data.fill_(factor*1.0)
+        elif isinstance(module, (T5Model, T5WithLMHeadModel)):
+            # Mesh TensorFlow embeddings initialization
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
+            module.shared.weight.data.normal_(mean=0.0, std=factor*1.0)
+        elif isinstance(module, T5DenseReluDense):
+            # Mesh TensorFlow FF initialization
+            # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
+            # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
+            module.wi.weight.data.normal_(mean=0.0, std=factor*((self.config.d_model) ** -0.5))
+            if hasattr(module.wi, 'bias') and module.wi.bias is not None:
+                module.wi.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor*((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, 'bias') and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5Attention):
+            # Mesh TensorFlow attention initialization to avoid scaling before softmax
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
+            d_model = self.config.d_model
+            d_kv = self.config.d_kv
+            n_heads = self.config.num_heads
+            module.q.weight.data.normal_(mean=0.0, std=factor*((d_model * d_kv) ** -0.5))
+            module.k.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
+            module.v.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
+            module.o.weight.data.normal_(mean=0.0, std=factor*((n_heads * d_kv) ** -0.5))
+            if module.has_relative_attention_bias:
+                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor*((d_model) ** -0.5))
+
+
+class T5Stack(T5PreTrainedModel):
+    def __init__(self, config):
+        super(T5Stack, self).__init__(config)
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.is_decoder = config.is_decoder
+
+        self.block = nn.ModuleList([T5Block(config, has_relative_attention_bias=bool(i == 0))
+                                    for i in range(config.num_layers)])
+        self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        self.init_weights()
+
+    def forward(self,
+                hidden_states,
+                attention_mask=None,
+                encoder_hidden_states=None,
+                encoder_attention_mask=None,
+                head_mask=None):
+
+        batch_size, seq_length = hidden_states.shape[0], hidden_states.shape[1]
+        if attention_mask is None:
+            attention_mask = torch.ones(batch_size, seq_length).to(hidden_states.device)
+        if self.is_decoder and encoder_attention_mask is None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = torch.ones(batch_size, encoder_seq_length).to(hidden_states.device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+        # Provided a padding mask of dimensions [batch_size, seq_length]
+        # - if the model is a decoder, apply a causal mask in addition to the padding mask
+        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                seq_ids = torch.arange(seq_length, device=hidden_states.device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = causal_mask.to(attention_mask)
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -1e9 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+        # extended_attention_mask = (extended_attention_mask == extended_attention_mask.transpose(-1, -2))
+
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
+
+        if self.is_decoder:
+            # If a 2D ou 3D attention mask is provided for the cross-attention
+            # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+            if encoder_attention_mask.dim() == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            if encoder_attention_mask.dim() == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+            # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+            # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+            # encoder_extended_attention_mask = (encoder_extended_attention_mask == encoder_extended_attention_mask.transpose(-1, -2))
+
+            encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            encoder_extended_attention_mask = None
+
+        # 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_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_layers
+
+        all_hidden_states = ()
+        all_attentions = ()
+        position_bias = None
+        encoder_decoder_position_bias = None
+
+        hidden_states = self.dropout(hidden_states)
+        for i, layer_module in enumerate(self.block):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states,
+                                         attention_mask=extended_attention_mask,
+                                         position_bias=position_bias,
+                                         encoder_hidden_states=encoder_hidden_states,
+                                         encoder_attention_mask=encoder_extended_attention_mask,
+                                         encoder_decoder_position_bias=encoder_decoder_position_bias,
+                                         head_mask=head_mask[i])
+            # layer_outputs is a tuple with:
+            # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+            hidden_states = layer_outputs[0]
+            if i == 0:
+                # We share the position biases between the layers - the first layer store them
+                # layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+                position_bias = layer_outputs[2 if self.output_attentions else 1]
+                if self.is_decoder:
+                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)  # We keep only self-attention weights for now
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        layer_output = self.dropout(hidden_states)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        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)
+
+
+T5_START_DOCSTRING = r"""    The T5 model was proposed in
+    `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
+    by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
+    It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
+
+    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.
+
+    .. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
+        https://arxiv.org/abs/1910.10683
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.T5Config`): 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:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+T5_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, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+            T5 is a model with relative position embeddings so you should be able to pad the inputs on
+            the right or the left.
+
+            Indices can be obtained using :class:`transformers.T5Tokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **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 T5 Model transformer outputting raw hidden-states"
+                      "without any specific head on top.",
+                      T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class T5Model(T5PreTrainedModel):
+    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 = T5Tokenizer.from_pretrained('t5-small')
+        model = T5Model.from_pretrained('t5-small')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(T5Model, self).__init__(config)
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = T5Stack(encoder_config)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = T5Stack(decoder_config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_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.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        encoder_attention_mask = kwargs_encoder.get("attention_mask", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            if encoder_attention_mask is not None:
+                # Apply masking
+                encoder_attention_mask = (encoder_attention_mask != 0).to(hidden_states)
+                hidden_states = hidden_states * encoder_attention_mask.unsqueeze(-1)
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = encoder_attention_mask
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
+    T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class T5WithLMHeadModel(T5PreTrainedModel):
+    r"""
+        **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 ``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 = T5Tokenizer.from_pretrained('t5-small')
+        model = T5WithLMHeadModel.from_pretrained('t5-small')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids=input_ids, lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(T5WithLMHeadModel, self).__init__(config)
+        self.model_dim = config.d_model
+
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = T5Stack(encoder_config)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = T5Stack(decoder_config)
+
+        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def forward(self, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+
+        lm_labels = kwargs.pop('decoder_lm_labels', None)
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        sequence_output = decoder_outputs[0]
+        # Rescale output before projecting on vocab
+        # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+        sequence_output = sequence_output * (self.model_dim ** -0.5)
+        lm_logits = self.lm_head(sequence_output)
+
+        decoder_outputs = (lm_logits,) + decoder_outputs[1:]  # Add hidden states and attention if they are here
+        if lm_labels is not None:
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = lm_labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            decoder_outputs = (loss,) + decoder_outputs  # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
+
+        return decoder_outputs + encoder_outputs

transformers/modeling_tf_albert.py

+# coding=utf-8
+# Copyright 2018 The OpenAI Team Authors and 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.
+""" TF 2.0 ALBERT model. """
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import logging
+import sys
+
+import tensorflow as tf
+
+from .configuration_albert import AlbertConfig
+from .modeling_tf_utils import TFPreTrainedModel, get_initializer, shape_list
+from .modeling_tf_bert import ACT2FN, TFBertSelfAttention
+from .file_utils import add_start_docstrings
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    'albert-base-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v1-tf_model.h5",
+    'albert-large-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v1-tf_model.h5",
+    'albert-xlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v1-tf_model.h5",
+    'albert-xxlarge-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v1-tf_model.h5",
+    'albert-base-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-base-v2-tf_model.h5",
+    'albert-large-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-large-v2-tf_model.h5",
+    'albert-xlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xlarge-v2-tf_model.h5",
+    'albert-xxlarge-v2': "https://s3.amazonaws.com/models.huggingface.co/bert/albert-xxlarge-v2-tf_model.h5",
+}
+
+
+class TFAlbertEmbeddings(tf.keras.layers.Layer):
+    """Construct the embeddings from word, position and token_type embeddings.
+    """
+
+    def __init__(self, config, **kwargs):
+        super(TFAlbertEmbeddings, self).__init__(**kwargs)
+
+        self.config = config
+        self.position_embeddings = tf.keras.layers.Embedding(config.max_position_embeddings,
+                                                             config.embedding_size,
+                                                             embeddings_initializer=get_initializer(
+                                                                 self.config.initializer_range),
+                                                             name='position_embeddings')
+        self.token_type_embeddings = tf.keras.layers.Embedding(config.type_vocab_size,
+                                                               config.embedding_size,
+                                                               embeddings_initializer=get_initializer(
+                                                                   self.config.initializer_range),
+                                                               name='token_type_embeddings')
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def build(self, input_shape):
+        """Build shared word embedding layer """
+        with tf.name_scope("word_embeddings"):
+            # Create and initialize weights. The random normal initializer was chosen
+            # arbitrarily, and works well.
+            self.word_embeddings = self.add_weight(
+                "weight",
+                shape=[self.config.vocab_size, self.config.embedding_size],
+                initializer=get_initializer(self.config.initializer_range))
+        super(TFAlbertEmbeddings, self).build(input_shape)
+
+    def call(self, inputs, mode="embedding", training=False):
+        """Get token embeddings of inputs.
+        Args:
+            inputs: list of three int64 tensors with shape [batch_size, length]: (input_ids, position_ids, token_type_ids)
+            mode: string, a valid value is one of "embedding" and "linear".
+        Returns:
+            outputs: (1) If mode == "embedding", output embedding tensor, float32 with
+                shape [batch_size, length, embedding_size]; (2) mode == "linear", output
+                linear tensor, float32 with shape [batch_size, length, vocab_size].
+        Raises:
+            ValueError: if mode is not valid.
+
+        Shared weights logic adapted from
+            https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
+        """
+        if mode == "embedding":
+            return self._embedding(inputs, training=training)
+        elif mode == "linear":
+            return self._linear(inputs)
+        else:
+            raise ValueError("mode {} is not valid.".format(mode))
+
+    def _embedding(self, inputs, training=False):
+        """Applies embedding based on inputs tensor."""
+        input_ids, position_ids, token_type_ids, inputs_embeds = inputs
+
+        if input_ids is not None:
+            input_shape = shape_list(input_ids)
+        else:
+            input_shape = shape_list(inputs_embeds)[:-1]
+
+        seq_length = input_shape[1]
+        if position_ids is None:
+            position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
+        if token_type_ids is None:
+            token_type_ids = tf.fill(input_shape, 0)
+
+        if inputs_embeds is None:
+            inputs_embeds = tf.gather(self.word_embeddings, input_ids)
+        position_embeddings = self.position_embeddings(position_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + position_embeddings + token_type_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings, training=training)
+        return embeddings
+
+    def _linear(self, inputs):
+        """Computes logits by running inputs through a linear layer.
+            Args:
+                inputs: A float32 tensor with shape [batch_size, length, embedding_size]
+            Returns:
+                float32 tensor with shape [batch_size, length, vocab_size].
+        """
+        batch_size = shape_list(inputs)[0]
+        length = shape_list(inputs)[1]
+        x = tf.reshape(inputs, [-1, self.config.embedding_size])
+        logits = tf.matmul(x, self.word_embeddings, transpose_b=True)
+        return tf.reshape(logits, [batch_size, length, self.config.vocab_size])
+
+
+class TFAlbertSelfAttention(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertSelfAttention, self).__init__(**kwargs)
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
+        self.output_attentions = config.output_attentions
+
+        self.num_attention_heads = config.num_attention_heads
+        assert config.hidden_size % config.num_attention_heads == 0
+        self.attention_head_size = int(
+            config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = tf.keras.layers.Dense(self.all_head_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='query')
+        self.key = tf.keras.layers.Dense(self.all_head_size,
+                                         kernel_initializer=get_initializer(
+                                             config.initializer_range),
+                                         name='key')
+        self.value = tf.keras.layers.Dense(self.all_head_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='value')
+
+        self.dropout = tf.keras.layers.Dropout(
+            config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, batch_size):
+        x = tf.reshape(
+            x, (batch_size, -1, self.num_attention_heads, self.attention_head_size))
+        return tf.transpose(x, perm=[0, 2, 1, 3])
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        batch_size = shape_list(hidden_states)[0]
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        # scale attention_scores
+        dk = tf.cast(shape_list(key_layer)[-1], tf.float32)
+        attention_scores = attention_scores / tf.math.sqrt(dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFAlbertModel call() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = tf.nn.softmax(attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+        context_layer = tf.reshape(context_layer,
+                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
+
+        outputs = (context_layer, attention_probs) if self.output_attentions else (
+            context_layer,)
+        return outputs
+
+
+class TFAlbertSelfOutput(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertSelfOutput, self).__init__(**kwargs)
+        self.dense = tf.keras.layers.Dense(config.hidden_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def call(self, inputs, training=False):
+        hidden_states, input_tensor = inputs
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class TFAlbertAttention(TFBertSelfAttention):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertAttention, self).__init__(config, **kwargs)
+
+        self.hidden_size = config.hidden_size
+        self.dense = tf.keras.layers.Dense(config.hidden_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        raise NotImplementedError
+
+    def call(self, inputs, training=False):
+        input_tensor, attention_mask, head_mask = inputs
+
+        batch_size = shape_list(input_tensor)[0]
+        mixed_query_layer = self.query(input_tensor)
+        mixed_key_layer = self.key(input_tensor)
+        mixed_value_layer = self.value(input_tensor)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer, batch_size)
+        key_layer = self.transpose_for_scores(mixed_key_layer, batch_size)
+        value_layer = self.transpose_for_scores(mixed_value_layer, batch_size)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        # (batch size, num_heads, seq_len_q, seq_len_k)
+        attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)
+        # scale attention_scores
+        dk = tf.cast(shape_list(key_layer)[-1], tf.float32)
+        attention_scores = attention_scores / tf.math.sqrt(dk)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in TFBertModel call() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = tf.nn.softmax(attention_scores, axis=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs, training=training)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = tf.matmul(attention_probs, value_layer)
+
+        context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
+        context_layer = tf.reshape(context_layer,
+                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
+
+        self_outputs = (context_layer, attention_probs) if self.output_attentions else (
+            context_layer,)
+
+        hidden_states = self_outputs[0]
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        attention_output = self.LayerNorm(hidden_states + input_tensor)
+
+        # add attentions if we output them
+        outputs = (attention_output,) + self_outputs[1:]
+        return outputs
+
+
+class TFAlbertLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertLayer, self).__init__(**kwargs)
+        self.attention = TFAlbertAttention(config, name='attention')
+
+        self.ffn = tf.keras.layers.Dense(config.intermediate_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='ffn')
+
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+        self.ffn_output = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='ffn_output')
+        self.full_layer_layer_norm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='full_layer_layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        attention_outputs = self.attention(
+            [hidden_states, attention_mask, head_mask], training=training)
+        ffn_output = self.ffn(attention_outputs[0])
+        ffn_output = self.activation(ffn_output)
+        ffn_output = self.ffn_output(ffn_output)
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = self.full_layer_layer_norm(
+            ffn_output + attention_outputs[0])
+
+        # add attentions if we output them
+        outputs = (hidden_states,) + attention_outputs[1:]
+        return outputs
+
+
+class TFAlbertLayerGroup(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertLayerGroup, self).__init__(**kwargs)
+
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.albert_layers = [TFAlbertLayer(config, name="albert_layers_._{}".format(
+            i)) for i in range(config.inner_group_num)]
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        layer_hidden_states = ()
+        layer_attentions = ()
+
+        for layer_index, albert_layer in enumerate(self.albert_layers):
+            layer_output = albert_layer(
+                [hidden_states, attention_mask, head_mask[layer_index]], training=training)
+            hidden_states = layer_output[0]
+
+            if self.output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+            if self.output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (layer_attentions,)
+        # last-layer hidden state, (layer hidden states), (layer attentions)
+        return outputs
+
+
+class TFAlbertTransformer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFAlbertTransformer, self).__init__(**kwargs)
+
+        self.config = config
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.embedding_hidden_mapping_in = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), name='embedding_hidden_mapping_in')
+        self.albert_layer_groups = [TFAlbertLayerGroup(
+            config, name="albert_layer_groups_._{}".format(i)) for i in range(config.num_hidden_groups)]
+
+    def call(self, inputs, training=False):
+        hidden_states, attention_mask, head_mask = inputs
+
+        hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+        all_attentions = ()
+
+        if self.output_hidden_states:
+            all_hidden_states = (hidden_states,)
+
+        for i in range(self.config.num_hidden_layers):
+            # Number of layers in a hidden group
+            layers_per_group = int(
+                self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+            # Index of the hidden group
+            group_idx = int(
+                i / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+            layer_group_output = self.albert_layer_groups[group_idx](
+                [hidden_states, attention_mask, head_mask[group_idx*layers_per_group:(group_idx+1)*layers_per_group]], training=training)
+            hidden_states = layer_group_output[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + layer_group_output[-1]
+
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+
+        # last-layer hidden state, (all hidden states), (all attentions)
+        return outputs
+
+
+class TFAlbertPreTrainedModel(TFPreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = AlbertConfig
+    pretrained_model_archive_map = TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "albert"
+
+
+class TFAlbertMLMHead(tf.keras.layers.Layer):
+    def __init__(self, config, input_embeddings, **kwargs):
+        super(TFAlbertMLMHead, self).__init__(**kwargs)
+        self.vocab_size = config.vocab_size
+
+        self.dense = tf.keras.layers.Dense(config.embedding_size,
+                                           kernel_initializer=get_initializer(
+                                               config.initializer_range),
+                                           name='dense')
+        if isinstance(config.hidden_act, str) or (sys.version_info[0] == 2 and isinstance(config.hidden_act, unicode)):
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = config.hidden_act
+
+        self.LayerNorm = tf.keras.layers.LayerNormalization(
+            epsilon=config.layer_norm_eps, name='LayerNorm')
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = input_embeddings
+
+    def build(self, input_shape):
+        self.bias = self.add_weight(shape=(self.vocab_size,),
+                                    initializer='zeros',
+                                    trainable=True,
+                                    name='bias')
+        self.decoder_bias = self.add_weight(shape=(self.vocab_size,),
+                                    initializer='zeros',
+                                    trainable=True,
+                                    name='decoder/bias')
+        super(TFAlbertMLMHead, self).build(input_shape)
+
+    def call(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        hidden_states = self.decoder(hidden_states, mode="linear") + self.decoder_bias
+        hidden_states = hidden_states + self.bias
+        return hidden_states
+
+
+ALBERT_START_DOCSTRING = r"""    The ALBERT model was proposed in
+    `ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`_
+    by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut. It presents
+    two parameter-reduction techniques to lower memory consumption and increase the trainig speed of BERT.
+
+    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
+    refer to the TF 2.0 documentation for all matter related to general usage and behavior.
+
+    .. _`ALBERT: A Lite BERT for Self-supervised Learning of Language Representations`:
+        https://arxiv.org/abs/1909.11942
+
+    .. _`tf.keras.Model`:
+        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
+
+    Note on the model inputs:
+        TF 2.0 models accepts two formats as inputs:
+
+            - having all inputs as keyword arguments (like PyTorch models), or
+            - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
+
+        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
+            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
+
+    Parameters:
+        config (:class:`~transformers.AlbertConfig`): 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:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+ALBERT_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, ALBERT input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+                
+                ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+                
+                ``token_type_ids:   0   0   0   0  0     0   0``
+
+            Albert 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:`transformers.AlbertTokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.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.
+        **token_type_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` 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 `ALBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding`_ for more details).
+        **position_ids**: (`optional`) ``Numpy array`` or ``tf.Tensor`` 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]``.
+        **head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` 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 Albert Model transformer outputing raw hidden-states without any specific head on top.",
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertModel(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **pooler_output**: ``tf.Tensor`` 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 Albert 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 ``tf.Tensor`` (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 ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertModel
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v1')
+        model = TFAlbertModel.from_pretrained('albert-base-v1')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+
+    def __init__(self, config, **kwargs):
+        super(TFAlbertModel, self).__init__(config, **kwargs)
+        self.num_hidden_layers = config.num_hidden_layers
+
+        self.embeddings = TFAlbertEmbeddings(config, name="embeddings")
+        self.encoder = TFAlbertTransformer(config, name="encoder")
+        self.pooler = tf.keras.layers.Dense(config.hidden_size, kernel_initializer=get_initializer(
+            config.initializer_range), activation='tanh', name='pooler')
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        raise NotImplementedError
+
+    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
+        """
+        raise NotImplementedError
+
+    def call(self, inputs, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, training=False):
+        if isinstance(inputs, (tuple, list)):
+            input_ids = inputs[0]
+            attention_mask = inputs[1] if len(inputs) > 1 else attention_mask
+            token_type_ids = inputs[2] if len(inputs) > 2 else token_type_ids
+            position_ids = inputs[3] if len(inputs) > 3 else position_ids
+            head_mask = inputs[4] if len(inputs) > 4 else head_mask
+            inputs_embeds = inputs[5] if len(inputs) > 5 else inputs_embeds
+            assert len(inputs) <= 6, "Too many inputs."
+        elif isinstance(inputs, dict):
+            input_ids = inputs.get('input_ids')
+            attention_mask = inputs.get('attention_mask', attention_mask)
+            token_type_ids = inputs.get('token_type_ids', token_type_ids)
+            position_ids = inputs.get('position_ids', position_ids)
+            head_mask = inputs.get('head_mask', head_mask)
+            inputs_embeds = inputs.get('inputs_embeds', inputs_embeds)
+            assert len(inputs) <= 6, "Too many inputs."
+        else:
+            input_ids = inputs
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = shape_list(input_ids)
+        elif inputs_embeds is not None:
+            input_shape = shape_list(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if attention_mask is None:
+            attention_mask = tf.fill(input_shape, 1)
+        if token_type_ids is None:
+            token_type_ids = tf.fill(input_shape, 0)
+
+        # We create a 3D attention mask from a 2D tensor mask.
+        # Sizes are [batch_size, 1, 1, to_seq_length]
+        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
+        # this attention mask is more simple than the triangular masking of causal attention
+        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
+        extended_attention_mask = attention_mask[:, tf.newaxis, tf.newaxis, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        extended_attention_mask = tf.cast(extended_attention_mask, tf.float32)
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+
+        # 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 not head_mask is None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            [input_ids, position_ids, token_type_ids, inputs_embeds], training=training)
+        encoder_outputs = self.encoder(
+            [embedding_output, extended_attention_mask, head_mask], training=training)
+
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output[:, 0])
+
+        # add hidden_states and attentions if they are here
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[1:]
+        # sequence_output, pooled_output, (hidden_states), (attentions)
+        return outputs
+
+
+@add_start_docstrings("""Albert Model with a `language modeling` head on top. """,
+                      ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertForMaskedLM(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **prediction_scores**: ``Numpy array`` or ``tf.Tensor`` 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 ``Numpy array`` or ``tf.Tensor`` (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 ``Numpy array`` or ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForMaskedLM
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForMaskedLM.from_pretrained('albert-base-v2')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        prediction_scores = outputs[0]
+
+    """
+
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFAlbertForMaskedLM, self).__init__(config, *inputs, **kwargs)
+
+        self.albert = TFAlbertModel(config, name='albert')
+        self.predictions = TFAlbertMLMHead(
+            config, self.albert.embeddings, name='predictions')
+
+    def get_output_embeddings(self):
+        return self.albert.embeddings
+
+    def call(self, inputs, **kwargs):
+        outputs = self.albert(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+        prediction_scores = self.predictions(
+            sequence_output, training=kwargs.get('training', False))
+
+        # Add hidden states and attention if they are here
+        outputs = (prediction_scores,) + outputs[2:]
+
+        return outputs  # prediction_scores, (hidden_states), (attentions)
+
+
+@add_start_docstrings("""Albert Model transformer with a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    ALBERT_START_DOCSTRING, ALBERT_INPUTS_DOCSTRING)
+class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **logits**: ``Numpy array`` or ``tf.Tensor`` 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 ``Numpy array`` or ``tf.Tensor`` (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 ``Numpy array`` or ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import AlbertTokenizer, TFAlbertForSequenceClassification
+
+        tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+        model = TFAlbertForSequenceClassification.from_pretrained('albert-base-v2')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        logits = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFAlbertForSequenceClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.albert = TFAlbertModel(config, name='albert')
+        self.dropout = tf.keras.layers.Dropout(config.hidden_dropout_prob)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.albert(inputs, **kwargs)
+
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output, training=kwargs.get('training', False))
+        logits = self.classifier(pooled_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # logits, (hidden_states), (attentions)
\ No newline at end of file

transformers/modeling_tf_auto.py

@@ -18,21 +18,48 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 
-from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, TFBertForQuestionAnswering
-from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel
-from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel
-from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel
-from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, TFXLNetForQuestionAnsweringSimple
-from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, TFXLMForQuestionAnsweringSimple
-from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification
-from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification
-from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel
+from .configuration_auto import (BertConfig, CTRLConfig, DistilBertConfig,
+                                 GPT2Config, OpenAIGPTConfig, RobertaConfig,
+                                 TransfoXLConfig, XLMConfig, XLNetConfig)
+
+from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, \
+    TFBertForQuestionAnswering, TFBertForTokenClassification, TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel, TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel, TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, \
+    TFXLNetForQuestionAnsweringSimple, TFXLNetForTokenClassification, TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, \
+    TFXLMForQuestionAnsweringSimple, TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, \
+    TFRobertaForTokenClassification, TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification, TFDistilBertForTokenClassification, TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel, TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_albert import TFAlbertModel, TFAlbertForMaskedLM, TFAlbertForSequenceClassification, TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_t5 import TFT5Model, TFT5WithLMHeadModel, TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
 
 from .file_utils import add_start_docstrings
 
 logger = logging.getLogger(__name__)
 
 
+TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
+    for pretrained_map in [
+        TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ]
+    for key, value, in pretrained_map.items())
+
+
 class TFAutoModel(object):
     r"""
         :class:`~transformers.TFAutoModel` is a generic model class
@@ -45,6 +72,7 @@ class TFAutoModel(object):
 
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5Model (T5 model)
             - contains `distilbert`: TFDistilBertModel (DistilBERT model)
             - contains `roberta`: TFRobertaModel (RoBERTa model)
             - contains `bert`: TFBertModel (Bert model)
@@ -59,7 +87,50 @@ class TFAutoModel(object):
     """
     def __init__(self):
         raise EnvironmentError("TFAutoModel is designed to be instantiated "
-            "using the `TFAutoModel.from_pretrained(pretrained_model_name_or_path)` method.")
+            "using the `TFAutoModel.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModel.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: TFDistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: TFRobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: TFBertModel (Bert model)
+                    - isInstance of `openai-gpt` configuration class: TFOpenAIGPTModel (OpenAI GPT model)
+                    - isInstance of `gpt2` configuration class: TFGPT2Model (OpenAI GPT-2 model)
+                    - isInstance of `ctrl` configuration class: TFCTRLModel (Salesforce CTRL  model)
+                    - isInstance of `transfo-xl` configuration class: TFTransfoXLModel (Transformer-XL model)
+                    - isInstance of `xlnet` configuration class: TFXLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: TFXLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = TFAutoModel.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertModel(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaModel(config)
+        elif isinstance(config, BertConfig):
+            return TFBertModel(config)
+        elif isinstance(config, OpenAIGPTConfig):
+            return TFOpenAIGPTModel(config)
+        elif isinstance(config, GPT2Config):
+            return TFGPT2Model(config)
+        elif isinstance(config, TransfoXLConfig):
+            return TFTransfoXLModel(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetModel(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMModel(config)
+        elif isinstance(config, CTRLConfig):
+            return TFCTRLModel(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -68,6 +139,7 @@ class TFAutoModel(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5Model (T5 model)
             - contains `distilbert`: TFDistilBertModel (DistilBERT model)
             - contains `roberta`: TFRobertaModel (RoBERTa model)
             - contains `bert`: TFTFBertModel (Bert model)
@@ -81,6 +153,7 @@ class TFAutoModel(object):
             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 string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -109,6 +182,9 @@ class TFAutoModel(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file 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.
@@ -133,8 +209,12 @@ class TFAutoModel(object):
             model = TFAutoModel.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return TFT5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -153,7 +233,7 @@ class TFAutoModel(object):
             return TFCTRLModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
+                         "'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
                          "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
 
 
@@ -169,6 +249,7 @@ class TFAutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5WithLMHeadModel (T5 model)
             - contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
             - contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
             - contains `bert`: TFBertForMaskedLM (Bert model)
@@ -183,7 +264,50 @@ class TFAutoModelWithLMHead(object):
     """
     def __init__(self):
         raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
-            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelWithLMHead.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `openai-gpt` configuration class: OpenAIGPTModel (OpenAI GPT model)
+                    - isInstance of `gpt2` configuration class: GPT2Model (OpenAI GPT-2 model)
+                    - isInstance of `ctrl` configuration class: CTRLModel (Salesforce CTRL  model)
+                    - isInstance of `transfo-xl` configuration class: TransfoXLModel (Transformer-XL model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelWithLMHead.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertForMaskedLM(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaForMaskedLM(config)
+        elif isinstance(config, BertConfig):
+            return TFBertForMaskedLM(config)
+        elif isinstance(config, OpenAIGPTConfig):
+            return TFOpenAIGPTLMHeadModel(config)
+        elif isinstance(config, GPT2Config):
+            return TFGPT2LMHeadModel(config)
+        elif isinstance(config, TransfoXLConfig):
+            return TFTransfoXLLMHeadModel(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetLMHeadModel(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMWithLMHeadModel(config)
+        elif isinstance(config, CTRLConfig):
+            return TFCTRLLMHeadModel(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -195,6 +319,7 @@ class TFAutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: TFT5WithLMHeadModel (T5 model)
             - contains `distilbert`: TFDistilBertForMaskedLM (DistilBERT model)
             - contains `roberta`: TFRobertaForMaskedLM (RoBERTa model)
             - contains `bert`: TFBertForMaskedLM (Bert model)
@@ -209,6 +334,7 @@ class TFAutoModelWithLMHead(object):
             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 string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -237,6 +363,9 @@ class TFAutoModelWithLMHead(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file 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.
@@ -261,8 +390,12 @@ class TFAutoModelWithLMHead(object):
             model = TFAutoModelWithLMHead.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return TFT5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -281,7 +414,7 @@ class TFAutoModelWithLMHead(object):
             return TFCTRLLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
+                         "'distilbert', 'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', "
                          "'xlm', 'roberta', 'ctrl'".format(pretrained_model_name_or_path))
 
 
@@ -306,8 +439,39 @@ class TFAutoModelForSequenceClassification(object):
         This class cannot be instantiated using `__init__()` (throws an error).
     """
     def __init__(self):
-        raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
-            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+        raise EnvironmentError("TFAutoModelForSequenceClassification is designed to be instantiated "
+            "using the `TFAutoModelForSequenceClassification.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelForSequenceClassification.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `roberta` configuration class: RobertaModel (RoBERTa model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForSequenceClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertForSequenceClassification(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaForSequenceClassification(config)
+        elif isinstance(config, BertConfig):
+            return TFBertForSequenceClassification(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetForSequenceClassification(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMForSequenceClassification(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -332,6 +496,7 @@ class TFAutoModelForSequenceClassification(object):
             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 string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -360,6 +525,9 @@ class TFAutoModelForSequenceClassification(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file 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.
@@ -386,6 +554,8 @@ class TFAutoModelForSequenceClassification(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -396,7 +566,7 @@ class TFAutoModelForSequenceClassification(object):
             return TFXLMForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
+                         "'distilbert', 'bert', 'xlnet', 'xlm', 'roberta'".format(pretrained_model_name_or_path))
 
 
 class TFAutoModelForQuestionAnswering(object):
@@ -419,8 +589,36 @@ class TFAutoModelForQuestionAnswering(object):
         This class cannot be instantiated using `__init__()` (throws an error).
     """
     def __init__(self):
-        raise EnvironmentError("TFAutoModelWithLMHead is designed to be instantiated "
-            "using the `TFAutoModelWithLMHead.from_pretrained(pretrained_model_name_or_path)` method.")
+        raise EnvironmentError("TFAutoModelForQuestionAnswering is designed to be instantiated "
+            "using the `TFAutoModelForQuestionAnswering.from_pretrained(pretrained_model_name_or_path)` or "
+            "`TFAutoModelForQuestionAnswering.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBERT model)
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `xlm` configuration class: XLMModel (XLM model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = AutoModelForSequenceClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, DistilBertConfig):
+            return TFDistilBertForQuestionAnswering(config)
+        elif isinstance(config, BertConfig):
+            return TFBertForQuestionAnswering(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetForQuestionAnswering(config)
+        elif isinstance(config, XLMConfig):
+            return TFXLMForQuestionAnswering(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
@@ -444,6 +642,7 @@ class TFAutoModelForQuestionAnswering(object):
             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 string with the `identifier name` of a pre-trained model that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
                 - a path to a `directory` containing model weights saved using :func:`~transformers.PreTrainedModel.save_pretrained`, e.g.: ``./my_model_directory/``.
                 - a path or url to a `PyTorch, TF 1.X or TF 2.0 checkpoint file` (e.g. `./tf_model/model.ckpt.index`). In the case of a PyTorch checkpoint, ``from_pt`` should be set to True and a configuration object should be provided as ``config`` argument.
 
@@ -472,6 +671,9 @@ class TFAutoModelForQuestionAnswering(object):
             force_download: (`optional`) boolean, default False:
                 Force to (re-)download the model weights and configuration files and override the cached versions if they exists.
 
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file 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.
@@ -506,4 +708,121 @@ class TFAutoModelForQuestionAnswering(object):
             return TFXLMForQuestionAnsweringSimple.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         raise ValueError("Unrecognized model identifier in {}. Should contains one of "
-                         "'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
+                         "'distilbert', 'bert', 'xlnet', 'xlm'".format(pretrained_model_name_or_path))
+
+
+class TFAutoModelForTokenClassification:
+    def __init__(self):
+        raise EnvironmentError("TFAutoModelForTokenClassification is designed to be instantiated "
+                               "using the `TFAutoModelForTokenClassification.from_pretrained(pretrained_model_name_or_path)` or "
+                               "`AutoModelForTokenClassification.from_config(config)` methods.")
+
+    @classmethod
+    def from_config(cls, config):
+        r""" Instantiates one of the base model classes of the library
+        from a configuration.
+
+            config: (`optional`) instance of a class derived from :class:`~transformers.PretrainedConfig`:
+                The model class to instantiate is selected based on the configuration class:
+                    - isInstance of `bert` configuration class: BertModel (Bert model)
+                    - isInstance of `xlnet` configuration class: XLNetModel (XLNet model)
+                    - isInstance of `distilbert` configuration class: DistilBertModel (DistilBert model)
+                    - isInstance of `roberta` configuration class: RobteraModel (Roberta model)
+
+        Examples::
+
+            config = BertConfig.from_pretrained('bert-base-uncased')    # Download configuration from S3 and cache.
+            model = TFAutoModelForTokenClassification.from_config(config)  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        """
+        if isinstance(config, BertConfig):
+            return TFBertForTokenClassification(config)
+        elif isinstance(config, XLNetConfig):
+            return TFXLNetForTokenClassification(config)
+        elif isinstance(config, DistilBertConfig):
+            return TFDistilBertForTokenClassification(config)
+        elif isinstance(config, RobertaConfig):
+            return TFRobertaForTokenClassification(config)
+        raise ValueError("Unrecognized configuration class {}".format(config))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
+        r""" Instantiates one of the question answering model classes of the library
+        from a pre-trained model configuration.
+
+        The `from_pretrained()` method takes care of returning the correct model class instance
+        using pattern matching on the `pretrained_model_name_or_path` string.
+
+        The model class to instantiate is selected as the first pattern matching
+        in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `bert`: BertForTokenClassification (Bert model)
+            - contains `xlnet`: XLNetForTokenClassification (XLNet model)
+            - contains `distilbert`: DistilBertForTokenClassification (DistilBert model)
+            - contains `roberta`: RobertaForTokenClassification (Roberta model)
+
+        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()`
+
+        Params:
+            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:`~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:`~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:`~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:`~transformers.PreTrainedModel.save_pretrained` and :func:`~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.
+
+            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`) 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:`~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 = TFAutoModelForTokenClassification.from_pretrained('bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = TFAutoModelForTokenClassification.from_pretrained('./test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = TFAutoModelForTokenClassification.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 = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+            model = TFAutoModelForTokenClassification.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+        if 'bert' in pretrained_model_name_or_path:
+            return TFBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'xlnet' in pretrained_model_name_or_path:
+            return TFXLNetForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
+            return TFDistilBertForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'roberta' in pretrained_model_name_or_path:
+            return TFRobertaForTokenClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+
+        raise ValueError("Unrecognized model identifier in {}. Should contains one of "
+                         "'bert', 'xlnet', 'distilbert', 'roberta'".format(pretrained_model_name_or_path))

transformers/modeling_tf_bert.py

@@ -28,7 +28,7 @@ import numpy as np
 import tensorflow as tf
 
 from .configuration_bert import BertConfig
-from .modeling_tf_utils import TFPreTrainedModel, get_initializer
+from .modeling_tf_utils import TFPreTrainedModel, get_initializer, shape_list
 from .file_utils import add_start_docstrings
 
 logger = logging.getLogger(__name__)
@@ -48,6 +48,12 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-large-uncased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-large-cased-whole-word-masking-finetuned-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-whole-word-masking-finetuned-squad-tf_model.h5",
     'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-tf_model.h5",
+    'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-tf_model.h5",
+    'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-tf_model.h5",
+    'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-tf_model.h5",
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-tf_model.h5",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/tf_model.h5",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/tf_model.h5",
 }
 
 
@@ -129,7 +135,7 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
                 linear tensor, float32 with shape [batch_size, length, vocab_size].
         Raises:
             ValueError: if mode is not valid.
-        
+
         Shared weights logic adapted from
             https://github.com/tensorflow/models/blob/a009f4fb9d2fc4949e32192a944688925ef78659/official/transformer/v2/embedding_layer.py#L24
         """
@@ -145,10 +151,10 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
         input_ids, position_ids, token_type_ids, inputs_embeds = inputs
 
         if input_ids is not None:
-            input_shape = tf.shape(input_ids)
+            input_shape = shape_list(input_ids)
         else:
-            input_shape = tf.shape(inputs_embeds)[:-1]
-        
+            input_shape = shape_list(inputs_embeds)[:-1]
+
         seq_length = input_shape[1]
         if position_ids is None:
             position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
@@ -172,8 +178,8 @@ class TFBertEmbeddings(tf.keras.layers.Layer):
             Returns:
                 float32 tensor with shape [batch_size, length, vocab_size].
         """
-        batch_size = tf.shape(inputs)[0]
-        length = tf.shape(inputs)[1]
+        batch_size = shape_list(inputs)[0]
+        length = shape_list(inputs)[1]
 
         x = tf.reshape(inputs, [-1, self.hidden_size])
         logits = tf.matmul(x, self.word_embeddings, transpose_b=True)
@@ -214,7 +220,7 @@ class TFBertSelfAttention(tf.keras.layers.Layer):
     def call(self, inputs, training=False):
         hidden_states, attention_mask, head_mask = inputs
 
-        batch_size = tf.shape(hidden_states)[0]
+        batch_size = shape_list(hidden_states)[0]
         mixed_query_layer = self.query(hidden_states)
         mixed_key_layer = self.key(hidden_states)
         mixed_value_layer = self.value(hidden_states)
@@ -225,7 +231,7 @@ class TFBertSelfAttention(tf.keras.layers.Layer):
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
         attention_scores = tf.matmul(query_layer, key_layer, transpose_b=True)  # (batch size, num_heads, seq_len_q, seq_len_k)
-        dk = tf.cast(tf.shape(key_layer)[-1], tf.float32) # scale attention_scores
+        dk = tf.cast(shape_list(key_layer)[-1], tf.float32) # scale attention_scores
         attention_scores = attention_scores / tf.math.sqrt(dk)
 
         if attention_mask is not None:
@@ -246,7 +252,7 @@ class TFBertSelfAttention(tf.keras.layers.Layer):
         context_layer = tf.matmul(attention_probs, value_layer)
 
         context_layer = tf.transpose(context_layer, perm=[0, 2, 1, 3])
-        context_layer = tf.reshape(context_layer, 
+        context_layer = tf.reshape(context_layer,
                                   (batch_size, -1, self.all_head_size))  # (batch_size, seq_len_q, all_head_size)
 
         outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
@@ -502,9 +508,9 @@ class TFBertMainLayer(tf.keras.layers.Layer):
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
-            input_shape = input_ids.shape
+            input_shape = shape_list(input_ids)
         elif inputs_embeds is not None:
-            input_shape = inputs_embeds.shape[:-1]
+            input_shape = shape_list(inputs_embeds)[:-1]
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
@@ -591,7 +597,7 @@ BERT_START_DOCSTRING = r"""    The BERT model was proposed in
             `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
 
     Parameters:
-        config (:class:`~transformers.BertConfig`): Model configuration class with all the parameters of the model. 
+        config (:class:`~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:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
 """
@@ -605,13 +611,13 @@ BERT_INPUTS_DOCSTRING = r"""
             (a) For sequence pairs:
 
                 ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
-                
+
                 ``token_type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1``
 
             (b) For single sequences:
 
                 ``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
@@ -671,7 +677,7 @@ class TFBertModel(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertModel.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -710,7 +716,7 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForPreTraining.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, seq_relationship_scores = outputs[:2]
 
@@ -759,7 +765,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForMaskedLM.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores = outputs[0]
 
@@ -806,7 +812,7 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForNextSentencePrediction.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         seq_relationship_scores = outputs[0]
 
@@ -851,7 +857,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 
@@ -939,11 +945,11 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel):
             input_ids = inputs
 
         if input_ids is not None:
-            num_choices = tf.shape(input_ids)[1]
-            seq_length = tf.shape(input_ids)[2]
+            num_choices = shape_list(input_ids)[1]
+            seq_length = shape_list(input_ids)[2]
         else:
-            num_choices = tf.shape(inputs_embeds)[1]
-            seq_length = tf.shape(inputs_embeds)[2]
+            num_choices = shape_list(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[2]
 
         flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
         flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
@@ -988,7 +994,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForTokenClassification.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         scores = outputs[0]
 
@@ -1041,7 +1047,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel):
 
         tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
         model = TFBertForQuestionAnswering.from_pretrained('bert-base-uncased')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         start_scores, end_scores = outputs[:2]
 

transformers/modeling_tf_ctrl.py

@@ -95,7 +95,7 @@ class TFMultiHeadAttention(tf.keras.layers.Layer):
 
     def call(self, inputs, training=False):
         v, k, q, mask, layer_past, attention_mask, head_mask = inputs
-        batch_size = q.shape[0]
+        batch_size = shape_list(q)[0]
 
         q = self.Wq(q)
         k = self.Wk(k)
@@ -400,7 +400,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
         **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -418,7 +418,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
 
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = TFCTRLModel.from_pretrained('ctrl')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -462,7 +462,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
         **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).
         **past**:
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -481,7 +481,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
         tokenizer = CTRLTokenizer.from_pretrained('ctrl')
         model = TFCTRLLMHeadModel.from_pretrained('ctrl')
 
-        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        input_ids = torch.tensor(tokenizer.encode("Links Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
         outputs = model(input_ids, labels=input_ids)
         loss, logits = outputs[:2]
 

transformers/modeling_tf_distilbert.py

@@ -37,7 +37,8 @@ logger = logging.getLogger(__name__)
 
 TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'distilbert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-tf_model.h5",
-    'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-tf_model.h5"
+    'distilbert-base-uncased-distilled-squad': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-uncased-distilled-squad-tf_model.h5",
+    'distilbert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/distilbert-base-multilingual-cased-tf_model.h5",
 }
 
 
@@ -137,9 +138,9 @@ class TFEmbeddings(tf.keras.layers.Layer):
             input_ids, position_ids = inputs
 
         if input_ids is not None:
-            seq_length = tf.shape(input_ids)[1]
+            seq_length = shape_list(input_ids)[1]
         else:
-            seq_length = tf.shape(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[1]
 
         if position_ids is None:
             position_ids = tf.range(seq_length, dtype=tf.int32)[tf.newaxis, :]
@@ -160,8 +161,8 @@ class TFEmbeddings(tf.keras.layers.Layer):
             Returns:
                 float32 tensor with shape [batch_size, length, vocab_size].
         """
-        batch_size = tf.shape(inputs)[0]
-        length = tf.shape(inputs)[1]
+        batch_size = shape_list(inputs)[0]
+        length = shape_list(inputs)[1]
 
         x = tf.reshape(inputs, [-1, self.dim])
         logits = tf.matmul(x, self.word_embeddings, transpose_b=True)
@@ -703,6 +704,53 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel):
         return outputs  # logits, (hidden_states), (attentions)
 
 
+@add_start_docstrings("""DistilBert Model with a token classification head on top (a linear layer on top of
+    the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. """,
+    DISTILBERT_START_DOCSTRING, DISTILBERT_INPUTS_DOCSTRING)
+class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.num_labels)``
+            Classification scores (before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (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 ``Numpy array`` or ``tf.Tensor`` (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::
+        import tensorflow as tf
+        from transformers import DistilBertTokenizer, TFDistilBertForTokenClassification
+        tokenizer = DistilBertTokenizer.from_pretrained('bert-base-uncased')
+        model = TFDistilBertForTokenClassification.from_pretrained('bert-base-uncased')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids)
+        scores = outputs[0]
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFDistilBertForTokenClassification, self).__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
+        self.distilbert = TFDistilBertMainLayer(config, name='distilbert')
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        self.classifier = tf.keras.layers.Dense(config.num_labels,
+                                                kernel_initializer=get_initializer(config.initializer_range),
+                                                name='classifier')
+
+    def call(self, inputs, **kwargs):
+        outputs = self.distilbert(inputs, **kwargs)
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output, training=kwargs.get('training', False))
+        logits = self.classifier(sequence_output)
+
+        outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+
+        return outputs  # scores, (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)

transformers/modeling_tf_gpt2.py

@@ -92,7 +92,7 @@ class TFAttention(tf.keras.layers.Layer):
         # q, k, v have shape [batch, heads, sequence, features]
         w = tf.matmul(q, k, transpose_b=True)
         if self.scale:
-            dk = tf.cast(tf.shape(k)[-1], tf.float32) # scale attention_scores
+            dk = tf.cast(shape_list(k)[-1], tf.float32) # scale attention_scores
             w = w / tf.math.sqrt(dk)
 
         # w has shape [batch, heads, dst_sequence, src_sequence], where information flows from src to dst.
@@ -436,7 +436,7 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
         **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
             Sequence of hidden-states at the last layer of the model.
         **past**:
-            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of ``tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -454,7 +454,7 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
 
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = TFGPT2Model.from_pretrained('gpt2')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -476,7 +476,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
         **prediction_scores**: `tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
             Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
         **past**:
-            list of `tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of `tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -495,7 +495,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
         tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
         model = TFGPT2LMHeadModel.from_pretrained('gpt2')
 
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 
@@ -535,7 +535,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         **mc_prediction_scores**: `tf.Tensor`` of shape ``(batch_size, num_choices)``
             Prediction scores of the multiplechoice classification head (scores for each choice before SoftMax).
         **past**:
-            list of `tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            list of `tf.Tensor`` (one for each layer) of shape ``(2, batch_size, num_heads, sequence_length, embed_size_per_head)``:
             that contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding.
         **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
@@ -574,6 +574,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
     """
     def __init__(self, config, *inputs, **kwargs):
         super(TFGPT2DoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
+        config.num_labels = 1
         self.transformer = TFGPT2MainLayer(config, name='transformer')
         self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')
 

transformers/modeling_tf_openai.py

@@ -98,7 +98,7 @@ class TFAttention(tf.keras.layers.Layer):
         # q, k, v have shape [batch, heads, sequence, features]
         w = tf.matmul(q, k, transpose_b=True)
         if self.scale:
-            dk = tf.cast(tf.shape(k)[-1], tf.float32) # scale attention_scores
+            dk = tf.cast(shape_list(k)[-1], tf.float32) # scale attention_scores
             w = w / tf.math.sqrt(dk)
 
         # w has shape [batch, heads, dst_sequence, src_sequence], where information flows from src to dst.
@@ -431,7 +431,7 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = TFOpenAIGPTModel.from_pretrained('openai-gpt')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -467,7 +467,7 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel):
 
         tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
         model = TFOpenAIGPTLMHeadModel.from_pretrained('openai-gpt')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]
 
@@ -538,6 +538,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
     """
     def __init__(self, config, *inputs, **kwargs):
         super(TFOpenAIGPTDoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
+        config.num_labels = 1
         self.transformer = TFOpenAIGPTMainLayer(config, name='transformer')
         self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')
 

transformers/modeling_tf_pytorch_utils.py

@@ -78,6 +78,7 @@ def load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_i
     logger.info("Loading PyTorch weights from {}".format(pt_path))
 
     pt_state_dict = torch.load(pt_path, map_location='cpu')
+    logger.info("PyTorch checkpoint contains {:,} parameters".format(sum(t.numel() for t in pt_state_dict.values())))
 
     return load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=tf_inputs, allow_missing_keys=allow_missing_keys)
 
@@ -131,7 +132,7 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
         start_prefix_to_remove = tf_model.base_model_prefix + '.'
 
     symbolic_weights = tf_model.trainable_weights + tf_model.non_trainable_weights
-
+    tf_loaded_numel = 0
     weight_value_tuples = []
     all_pytorch_weights = set(list(pt_state_dict.keys()))
     for symbolic_weight in symbolic_weights:
@@ -139,7 +140,11 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
         name, transpose = convert_tf_weight_name_to_pt_weight_name(sw_name, start_prefix_to_remove=start_prefix_to_remove)
 
         # Find associated numpy array in pytorch model state dict
-        assert name in pt_state_dict, "{} not found in PyTorch model".format(name)
+        if name not in pt_state_dict:
+            if allow_missing_keys:
+                continue
+            raise AttributeError("{} not found in PyTorch model".format(name))
+
         array = pt_state_dict[name].numpy()
 
         if transpose:
@@ -156,7 +161,8 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
             e.args += (symbolic_weight.shape, array.shape)
             raise e
 
-        logger.info("Initialize TF weight {}".format(symbolic_weight.name))
+        tf_loaded_numel += array.size
+        # logger.warning("Initialize TF weight {}".format(symbolic_weight.name))
 
         weight_value_tuples.append((symbolic_weight, array))
         all_pytorch_weights.discard(name)
@@ -166,6 +172,8 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
     if tf_inputs is not None:
         tfo = tf_model(tf_inputs, training=False)  # Make sure restore ops are run
 
+    logger.info("Loaded {:,} parameters in the TF 2.0 model.".format(tf_loaded_numel))
+
     logger.info("Weights or buffers not loaded from PyTorch model: {}".format(all_pytorch_weights))
 
     return tf_model
@@ -243,6 +251,7 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
 
     all_tf_weights = set(list(tf_weights_map.keys()))
     loaded_pt_weights_data_ptr = {}
+    missing_keys_pt = []
     for pt_weight_name, pt_weight in current_pt_params_dict.items():
         # Handle PyTorch shared weight ()not duplicated in TF 2.0
         if pt_weight.data_ptr() in loaded_pt_weights_data_ptr:
@@ -251,7 +260,10 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
 
         # Find associated numpy array in pytorch model state dict
         if pt_weight_name not in tf_weights_map:
-            raise ValueError("{} not found in TF 2.0 model".format(pt_weight_name))
+            if allow_missing_keys:
+                missing_keys_pt.append(pt_weight_name)
+                continue
+            raise AttributeError("{} not found in TF 2.0 model".format(pt_weight_name))
 
         array, transpose = tf_weights_map[pt_weight_name]
 
@@ -269,13 +281,14 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
             e.args += (pt_weight.shape, array.shape)
             raise e
 
-        logger.info("Initialize PyTorch weight {}".format(pt_weight_name))
+        # logger.warning("Initialize PyTorch weight {}".format(pt_weight_name))
 
         new_pt_params_dict[pt_weight_name] = torch.from_numpy(array)
         loaded_pt_weights_data_ptr[pt_weight.data_ptr()] = torch.from_numpy(array)
         all_tf_weights.discard(pt_weight_name)
 
     missing_keys, unexpected_keys = pt_model.load_state_dict(new_pt_params_dict, strict=False)
+    missing_keys += missing_keys_pt
 
     if len(missing_keys) > 0:
         logger.info("Weights of {} not initialized from TF 2.0 model: {}".format(

transformers/modeling_tf_roberta.py

@@ -24,7 +24,7 @@ import numpy as np
 import tensorflow as tf
 
 from .configuration_roberta import RobertaConfig
-from .modeling_tf_utils import TFPreTrainedModel, get_initializer
+from .modeling_tf_utils import TFPreTrainedModel, get_initializer, shape_list
 from .file_utils import add_start_docstrings
 
 from .modeling_tf_bert import TFBertEmbeddings, TFBertMainLayer, gelu, gelu_new
@@ -51,9 +51,9 @@ class TFRobertaEmbeddings(TFBertEmbeddings):
         input_ids, position_ids, token_type_ids, inputs_embeds = inputs
 
         if input_ids is not None:
-            seq_length = tf.shape(input_ids)[1]
+            seq_length = shape_list(input_ids)[1]
         else:
-            seq_length = tf.shape(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[1]
 
         if position_ids is None:
             position_ids = tf.range(self.padding_idx+1, seq_length+self.padding_idx+1, dtype=tf.int32)[tf.newaxis, :]
@@ -199,7 +199,7 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = TFRobertaModel.from_pretrained('roberta-base')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
 
@@ -276,7 +276,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel):
 
         tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
         model = TFRobertaForMaskedLM.from_pretrained('roberta-base')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids, masked_lm_labels=input_ids)
         prediction_scores = outputs[0]
 
@@ -347,7 +347,7 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel):
 
         tokenizer = RoertaTokenizer.from_pretrained('roberta-base')
         model = TFRobertaForSequenceClassification.from_pretrained('roberta-base')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         labels = tf.constant([1])[None, :]  # Batch size 1
         outputs = model(input_ids)
         logits = outputs[0]

transformers/modeling_tf_t5.py

+# coding=utf-8
+# Copyright 2018 T5 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.
+""" TF 2.0 T5 model. """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import logging
+import math
+import copy
+import itertools
+
+import tensorflow as tf
+
+from .configuration_t5 import T5Config
+from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, shape_list
+from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
+
+logger = logging.getLogger(__name__)
+
+TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-tf_model.h5",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-tf_model.h5",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-tf_model.h5",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-tf_model.h5",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-tf_model.h5",
+}
+
+####################################################
+# TF 2.0 Models are constructed using Keras imperative API by sub-classing
+# - tf.keras.layers.Layer for the layers and
+# - TFPreTrainedModel for the models (it-self a sub-class of tf.keras.Model)
+####################################################
+
+class TFT5LayerNorm(tf.keras.layers.Layer):
+    def __init__(self, epsilon=1e-6, **kwargs):
+        """ Construct a layernorm module in the T5 style
+            No bias and no substraction of mean.
+        """
+        super(TFT5LayerNorm, self).__init__(**kwargs)
+        self.variance_epsilon = epsilon
+
+    def build(self, input_shape):
+        """Build shared word embedding layer """
+        self.weight = self.add_weight(
+            "weight",
+            shape=(input_shape[-1],),
+            initializer='ones')
+        super(TFT5LayerNorm, self).build(input_shape)
+
+    def call(self, x):
+        variance = tf.math.reduce_mean(tf.math.square(x), axis=-1, keepdims=True)
+        x = x * tf.math.rsqrt(variance + self.variance_epsilon)
+        return self.weight * x
+
+
+class TFT5DenseReluDense(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFT5DenseReluDense, self).__init__(**kwargs)
+        self.wi = tf.keras.layers.Dense(config.d_ff, use_bias=False, name='wi')
+        self.wo = tf.keras.layers.Dense(config.d_model, use_bias=False, name='wo')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+        self.act = tf.keras.activations.relu
+
+    def call(self, hidden_states, training=False):
+        h = self.wi(hidden_states)
+        h = self.act(h)
+        h = self.dropout(h, training=training)
+        h = self.wo(h)
+        return h
+
+
+class TFT5LayerFF(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFT5LayerFF, self).__init__(**kwargs)
+        self.DenseReluDense = TFT5DenseReluDense(config, name='DenseReluDense')
+        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                        name='layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def call(self, hidden_states, training=False):
+        norm_x = self.layer_norm(hidden_states)
+        y = self.DenseReluDense(norm_x, training=training)
+        layer_output = hidden_states + self.dropout(y, training=training)
+        return layer_output
+
+
+class TFT5Attention(tf.keras.layers.Layer):
+    NEW_ID = itertools.count()
+
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5Attention, self).__init__(**kwargs)
+        self.layer_id = next(TFT5Attention.NEW_ID)
+        self.is_decoder = config.is_decoder
+        self.has_relative_attention_bias = has_relative_attention_bias
+
+        self.output_attentions = config.output_attentions
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.d_model = config.d_model
+        self.d_kv = config.d_kv
+        self.n_heads = config.num_heads
+        self.inner_dim = self.n_heads * self.d_kv
+
+        # Mesh TensorFlow initialization to avoid scaling before softmax
+        self.q = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='q')
+        self.k = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='k')
+        self.v = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='v')
+        self.o = tf.keras.layers.Dense(self.d_model, use_bias=False, name='o')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = tf.keras.layers.Embedding(self.relative_attention_num_buckets,
+                                                                     self.n_heads,
+                                                                     name='relative_attention_bias')
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        raise NotImplementedError
+
+    @staticmethod
+    def _relative_position_bucket(relative_position,
+                                  bidirectional=True,
+                                  num_buckets=32,
+                                  max_distance=128):
+        """
+        Adapted from Mesh Tensorflow:
+        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+        Translate relative position to a bucket number for relative attention.
+        The relative position is defined as memory_position - query_position, i.e.
+        the distance in tokens from the attending position to the attended-to
+        position.  If bidirectional=False, then positive relative positions are
+        invalid.
+        We use smaller buckets for small absolute relative_position and larger buckets
+        for larger absolute relative_positions.  All relative positions >=max_distance
+        map to the same bucket.  All relative positions <=-max_distance map to the
+        same bucket.  This should allow for more graceful generalization to longer
+        sequences than the model has been trained on.
+        Args:
+            relative_position: an int32 Tensor
+            bidirectional: a boolean - whether the attention is bidirectional
+            num_buckets: an integer
+            max_distance: an integer
+        Returns:
+            a Tensor with the same shape as relative_position, containing int32
+            values in the range [0, num_buckets)
+        """
+        ret = 0
+        n = -relative_position
+        if bidirectional:
+            num_buckets //= 2
+            ret += tf.dtypes.cast(tf.math.less(n, 0), tf.int32) * num_buckets
+            n = tf.math.abs(n)
+        else:
+            n = tf.math.maximum(n, 0)
+        # now n is in the range [0, inf)
+        max_exact = num_buckets // 2
+        is_small = tf.math.less(n, max_exact)
+        val_if_large = max_exact + tf.dtypes.cast(
+            tf.math.log(tf.dtypes.cast(n, tf.float32) / max_exact)
+            / math.log(max_distance / max_exact) * (num_buckets - max_exact), tf.int32)
+        val_if_large = tf.math.minimum(val_if_large, num_buckets - 1)
+        ret += tf.where(is_small, n, val_if_large)
+        return ret
+
+    def compute_bias(self, qlen, klen):
+        """ Compute binned relative position bias """
+        context_position = tf.range(qlen)[:, None]
+        memory_position = tf.range(klen)[None, :]
+        relative_position = memory_position - context_position  # shape (qlen, klen)
+        rp_bucket = self._relative_position_bucket(relative_position,
+                                                   bidirectional=not self.is_decoder,
+                                                   num_buckets=self.relative_attention_num_buckets)
+        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
+        values = tf.expand_dims(tf.transpose(values, [2, 0, 1]), axis=0) # shape (1, num_heads, qlen, klen)
+        return values
+
+    def call(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None, training=False):
+        """
+        Self-attention (if kv is None) or attention over source sentence (provided by kv).
+        """
+        # Input is (bs, qlen, dim)
+        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
+        bs, qlen, dim = shape_list(input)
+        if kv is None:
+            klen = qlen if cache is None else cache['slen'] + qlen
+        else:
+            klen = shape_list(kv)[1]
+
+        def shape(x):
+            """  projection """
+            return tf.transpose(tf.reshape(x, (bs, -1, self.n_heads, self.d_kv)), perm=(0, 2, 1, 3))
+
+        def unshape(x):
+            """  compute context """
+            return tf.reshape(tf.transpose(x, perm=(0, 2, 1, 3)), (bs, -1, self.inner_dim))
+
+        q = shape(self.q(input))                                          # (bs, n_heads, qlen, dim_per_head)
+        if kv is None:
+            k = shape(self.k(input))                                      # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(input))                                      # (bs, n_heads, qlen, dim_per_head)
+        elif cache is None or self.layer_id not in cache:
+            k = v = kv
+            k = shape(self.k(k))                                          # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(v))                                          # (bs, n_heads, qlen, dim_per_head)
+
+        if cache is not None:
+            if self.layer_id in cache:
+                if kv is None:
+                    k_, v_ = cache[self.layer_id]
+                    k = tf.concat([k_, k], axis=2)                             # (bs, n_heads, klen, dim_per_head)
+                    v = tf.concat([v_, v], axis=2)                             # (bs, n_heads, klen, dim_per_head)
+                else:
+                    k, v = cache[self.layer_id]
+            cache[self.layer_id] = (k, v)
+
+        # q = q / math.sqrt(dim_per_head)                                     # No scaling in T5
+        # scores = tf.matmul(q, k, transpose_b=True)                            # (bs, n_heads, qlen, klen)
+        scores = tf.einsum('bnqd,bnkd->bnqk', q, k)                        # (bs, n_heads, qlen, klen)
+
+        if position_bias is None:
+            if not self.has_relative_attention_bias:
+                raise ValueError("No position_bias provided and no weights to compute position_bias")
+            position_bias = self.compute_bias(qlen, klen)
+            if mask is not None:
+                position_bias = position_bias + mask
+                # mask = (mask == 0).expand_as(scores)                              # (bs, n_heads, qlen, klen)
+                # scores.masked_fill_(mask, -float('inf'))                          # (bs, n_heads, qlen, klen)
+
+        scores += position_bias
+        weights = tf.nn.softmax(scores, axis=-1)                              # (bs, n_heads, qlen, klen)
+        weights = self.dropout(weights, training=training)                    # (bs, n_heads, qlen, klen)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = tf.matmul(weights, v)                                    # (bs, n_heads, qlen, dim_per_head)
+        context = unshape(context)                                            # (bs, qlen, dim)
+
+        context = self.o(context)
+
+        outputs = (context,)
+        if self.output_attentions:
+            outputs = outputs + (weights,)
+        if self.has_relative_attention_bias:
+            outputs = outputs + (position_bias,)
+        return outputs
+
+
+class TFT5LayerSelfAttention(tf.keras.layers.Layer):
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5LayerSelfAttention, self).__init__(**kwargs)
+        self.SelfAttention = TFT5Attention(config,
+                                           has_relative_attention_bias=has_relative_attention_bias,
+                                           name='SelfAttention')
+        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                        name='layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def call(self, hidden_states, attention_mask=None, position_bias=None,
+             head_mask=None, training=False):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.SelfAttention(norm_x,
+                                              mask=attention_mask,
+                                              position_bias=position_bias,
+                                              head_mask=head_mask,
+                                              training=training)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y, training=training)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class TFT5LayerCrossAttention(tf.keras.layers.Layer):
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5LayerCrossAttention, self).__init__(**kwargs)
+        self.EncDecAttention = TFT5Attention(config,
+                                           has_relative_attention_bias=has_relative_attention_bias,
+                                           name='EncDecAttention')
+        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                        name='layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def call(self, hidden_states, kv, attention_mask=None, position_bias=None,
+             head_mask=None, training=False):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.EncDecAttention(norm_x,
+                                                mask=attention_mask,
+                                                kv=kv,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask,
+                                                training=training)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y, training=training)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class TFT5Block(tf.keras.layers.Layer):
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5Block, self).__init__(**kwargs)
+        self.is_decoder = config.is_decoder
+        self.layer = []
+        self.layer.append(TFT5LayerSelfAttention(config,
+                                                 has_relative_attention_bias=has_relative_attention_bias,
+                                                 name='layer_._0'))
+        if self.is_decoder:
+            self.layer.append(TFT5LayerCrossAttention(config,
+                                                      has_relative_attention_bias=has_relative_attention_bias,
+                                                      name='layer_._1'))
+            self.layer.append(TFT5LayerFF(config, name='layer_._2'))
+        else:
+            self.layer.append(TFT5LayerFF(config, name='layer_._1'))
+
+    def call(self, hidden_states, attention_mask=None, position_bias=None,
+             encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
+             head_mask=None, training=False):
+        self_attention_outputs = self.layer[0](hidden_states,
+                                                attention_mask=attention_mask,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask,
+                                                training=training)
+        hidden_states = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]
+
+        if not self.is_decoder:
+            hidden_states = self.layer[1](hidden_states, training=training)
+        else:
+            cross_attention_outputs = self.layer[1](hidden_states,
+                                                    kv=encoder_hidden_states,
+                                                    attention_mask=encoder_attention_mask,
+                                                    position_bias=encoder_decoder_position_bias,
+                                                    head_mask=head_mask,
+                                                    training=training)
+            hidden_states = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]
+            hidden_states = self.layer[2](hidden_states, training=training)
+
+        outputs = (hidden_states,) + outputs  # add attentions if we output them
+        return outputs # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+
+
+####################################################
+# The full model without a specific pretrained or finetuning head is
+# provided as a tf.keras.layers.Layer usually called "TFT5MainLayer"
+####################################################
+class TFT5MainLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFT5MainLayer, self).__init__(**kwargs)
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.is_decoder = config.is_decoder
+        self.config = config
+        self.num_hidden_layers = config.num_layers
+
+        self.block = [TFT5Block(config,
+                                has_relative_attention_bias=bool(i == 0),
+                                name='block_._{}'.format(i))
+                        for i in range(config.num_layers)]
+        self.final_layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                              name='final_layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
+
+    def _prune_heads(self, heads_to_prune):
+        raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
+
+    def call(self, hidden_states, attention_mask=None, encoder_hidden_states=None,
+             encoder_attention_mask=None, head_mask=None, training=False):
+
+        batch_size, seq_length = shape_list(hidden_states)[:2]
+        if attention_mask is None:
+            attention_mask = tf.fill((batch_size, seq_length), 1)
+        if self.is_decoder and encoder_attention_mask is None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = tf.fill((batch_size, encoder_seq_length), 1)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        attention_mask = tf.cast(attention_mask, dtype=tf.float32)
+        num_dims_attention_mask = len(shape_list(attention_mask))
+        if num_dims_attention_mask == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif num_dims_attention_mask == 2:
+        # Provided a padding mask of dimensions [batch_size, seq_length]
+        # - if the model is a decoder, apply a causal mask in addition to the padding mask
+        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                seq_ids = tf.range(seq_length)
+                causal_mask = tf.less_equal(tf.tile(seq_ids[None, None, :], (batch_size, seq_length, 1)),
+                                            seq_ids[None, :, None])
+                causal_mask = tf.cast(causal_mask, dtype=tf.float32)
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposistion
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+        # extended_attention_mask = tf.math.equal(extended_attention_mask,
+        #                                         tf.transpose(extended_attention_mask, perm=(-1, -2)))
+
+        extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
+
+        if self.is_decoder:
+            # If a 2D ou 3D attention mask is provided for the cross-attention
+            # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+            encoder_attention_mask = tf.cast(encoder_attention_mask, dtype=tf.float32)
+            num_dims_encoder_attention_mask = len(shape_list(encoder_attention_mask))
+            if num_dims_encoder_attention_mask == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            if num_dims_encoder_attention_mask == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+            # T5 has a mask that can compare sequence ids, we can simulate this here with this transposistion
+            # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+            # encoder_extended_attention_mask = tf.math.equal(encoder_extended_attention_mask,
+            #                                         tf.transpose(encoder_extended_attention_mask, perm=(-1, -2)))
+
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            encoder_extended_attention_mask = None
+
+        # 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 not head_mask is None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        all_hidden_states = ()
+        all_attentions = ()
+        position_bias = None
+        encoder_decoder_position_bias = None
+        for i, layer_module in enumerate(self.block):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states,
+                                         attention_mask=extended_attention_mask,
+                                         position_bias=position_bias,
+                                         encoder_hidden_states=encoder_hidden_states,
+                                         encoder_attention_mask=encoder_extended_attention_mask,
+                                         encoder_decoder_position_bias=encoder_decoder_position_bias,
+                                         head_mask=head_mask[i],
+                                         training=training)
+            hidden_states = layer_outputs[0]
+            if i == 0:
+                # We share the position biases between the layers - the first layer store them
+                # layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+                position_bias = layer_outputs[2 if self.output_attentions else 1]
+                if self.is_decoder:
+                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        layer_output = self.dropout(hidden_states, training=training)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        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)
+
+
+####################################################
+# TFT5PreTrainedModel is a sub-class of tf.keras.Model
+# which take care of loading and saving pretrained weights
+# and various common utilities.
+# Here you just need to specify a few (self-explanatory)
+# pointers for your model.
+####################################################
+class TFT5PreTrainedModel(TFPreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = T5Config
+    pretrained_model_archive_map = TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "transformer"
+
+    @property
+    def dummy_inputs(self):
+        input_ids = tf.constant(DUMMY_INPUTS)
+        input_mask = tf.constant(DUMMY_MASK)
+        dummy_inputs = {'decoder_input_ids': input_ids,
+                        'encoder_input_ids': input_ids,
+                        'decoder_attention_mask': input_mask}
+        return dummy_inputs
+
+
+T5_START_DOCSTRING = r"""    The T5 model was proposed in
+    `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
+    by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
+    It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
+
+    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
+    refer to the TF 2.0 documentation for all matter related to general usage and behavior.
+
+    .. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
+        https://arxiv.org/abs/1910.10683
+
+    .. _`tf.keras.Model`:
+        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
+
+    Note on the model inputs:
+        TF 2.0 models accepts two formats as inputs:
+
+            - having all inputs as keyword arguments (like PyTorch models), or
+            - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
+
+        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
+            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
+
+    Parameters:
+        config (:class:`~transformers.T5Config`): 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:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+T5_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+
+            T5 is a model with relative position embeddings so you should be able to pad the inputs on
+            the right or the left.
+
+            Indices can be obtained using :class:`transformers.T5Tokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.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`) ``Numpy array`` or ``tf.Tensor`` 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 T5 Model transformer outputting raw hidden-states"
+                      "without any specific head on top.",
+                      T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class TFT5Model(TFT5PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``tf.Tensor`` 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 ``tf.Tensor`` (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 ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import T5Tokenizer, TFT5Model
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5Model.from_pretrained('t5-small')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFT5Model, self).__init__(config, *inputs, **kwargs)
+        self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model,
+                                         name='shared')
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = TFT5MainLayer(encoder_config, name='encoder')
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = TFT5MainLayer(decoder_config, name='decoder')
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def get_output_embeddings(self):
+        return self.shared
+
+    def call(self, decoder_input_ids, **kwargs):
+        # We allow two types of multi-inputs:
+        # - traditional keyword arguments in the call method
+        # - all the arguments provided as a dict in the first positional argument of call
+        # The last option is useful to use the tf.keras fit() method.
+
+        if isinstance(decoder_input_ids, dict):
+            kwargs.update(decoder_input_ids)
+        else:
+            kwargs['decoder_input_ids'] = decoder_input_ids
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
+    T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class TFT5WithLMHeadModel(TFT5PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **prediction_scores**: ``Numpy array`` or ``tf.Tensor`` 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 ``Numpy array`` or ``tf.Tensor`` (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 ``Numpy array`` or ``tf.Tensor`` (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::
+
+        import tensorflow as tf
+        from transformers import T5Tokenizer, TFT5WithLMHeadModel
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5WithLMHeadModel.from_pretrained('t5-small')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        prediction_scores = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFT5WithLMHeadModel, self).__init__(config, *inputs, **kwargs)
+        self.model_dim = config.d_model
+
+        self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model,
+                                         name='shared')
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = TFT5MainLayer(encoder_config, name='encoder')
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = TFT5MainLayer(decoder_config, name='decoder')
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def get_output_embeddings(self):
+        return self.shared
+
+    def call(self, decoder_input_ids, **kwargs):
+        # We allow two types of multi-inputs:
+        # - traditional keyword arguments in the call method
+        # - all the arguments provided as a dict in the first positional argument of call
+        # The last option is useful to use the tf.keras fit() method.
+
+        if isinstance(decoder_input_ids, dict):
+            kwargs.update(decoder_input_ids)
+        else:
+            kwargs['decoder_input_ids'] = decoder_input_ids
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        sequence_output = decoder_outputs[0] * (self.model_dim ** -0.5)
+        lm_logits = self.shared(sequence_output, mode="linear")
+        decoder_outputs = (lm_logits,) + decoder_outputs[1:]
+
+        return decoder_outputs + encoder_outputs

transformers/modeling_tf_transfo_xl.py

@@ -337,7 +337,7 @@ class TFAdaptiveEmbedding(tf.keras.layers.Layer):
                 emb_i = tf.einsum('id,de->ie', emb_i, self.emb_projs[i])
 
                 mask_idx = tf.cast(tf.where(mask_i), dtype=tf.int64)
-                emb_flat += tf.scatter_nd(mask_idx, emb_i, tf.cast(tf.shape(emb_flat), dtype=tf.int64))
+                emb_flat += tf.scatter_nd(mask_idx, emb_i, tf.cast(shape_list(emb_flat), dtype=tf.int64))
 
             embed_shape = shape_list(inp) + [self.d_proj]
             embed = tf.reshape(emb_flat, embed_shape)
@@ -353,7 +353,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
 
-        self.n_token = config.n_token
+        self.n_token = config.vocab_size
 
         self.d_embed = config.d_embed
         self.d_model = config.d_model
@@ -361,7 +361,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         self.d_head = config.d_head
         self.untie_r = config.untie_r
 
-        self.word_emb = TFAdaptiveEmbedding(config.n_token, config.d_embed, config.d_model, config.cutoffs, 
+        self.word_emb = TFAdaptiveEmbedding(config.vocab_size, config.d_embed, config.d_model, config.cutoffs, 
                                             div_val=config.div_val, init_std=config.init_std, name='word_emb')
 
         self.drop = tf.keras.layers.Dropout(config.dropout)
@@ -673,7 +673,7 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
 
         tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
         model = TFTransfoXLModel.from_pretrained('transfo-xl-wt103')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         last_hidden_states, mems = outputs[:2]
 
@@ -715,7 +715,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
 
         tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
         model = TFTransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
-        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
         outputs = model(input_ids)
         prediction_scores, mems = outputs[:2]
 
@@ -729,7 +729,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
             raise NotImplementedError
         # use adaptive softmax (including standard softmax)
         else:
-            self.crit = TFAdaptiveSoftmaxMask(config.n_token, config.d_embed, config.d_model, 
+            self.crit = TFAdaptiveSoftmaxMask(config.vocab_size, config.d_embed, config.d_model, 
                                               config.cutoffs, div_val=config.div_val, name='crit')
 
     def reset_length(self, tgt_len, ext_len, mem_len):