Merge pull request #1739 from huggingface/t5

[WIP] Adding Google T5 model

README.md

@@ -144,7 +144,8 @@ At some point in the future, you'll be able to seamlessly move from pre-training
 9. **[CTRL](https://github.com/salesforce/ctrl/)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
 10. **[CamemBERT](https://camembert-model.fr)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 11. **[ALBERT](https://github.com/google-research/ALBERT)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
-11. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
+12. **[T5](https://github.com/google-research/text-to-text-transfer-transformer)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
+13. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 

docs/source/pretrained_models.rst

@@ -217,6 +217,21 @@ Here is the full list of the currently provided pretrained models together with
 |                   |                                                            | | ALBERT xxlarge model with no dropout, additional training data and longer training                                                  |
 |                   |                                                            | (see `details <https://github.com/google-research/ALBERT>`__)                                                                         |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| T5                | ``t5-small``                                               | | ~60M parameters with 6-layers, 512-hidden-state, 2048 feed-forward hidden-state, 8-heads,                                           |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-base``                                                | | ~220M parameters with 12-layers, 768-hidden-state, 3072 feed-forward hidden-state, 12-heads,                                        |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-large``                                               | | ~770M parameters with 24-layers, 1024-hidden-state, 4096 feed-forward hidden-state, 16-heads,                                       |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-3B``                                                  | | ~2.8B parameters with 24-layers, 1024-hidden-state, 16384 feed-forward hidden-state, 32-heads,                                      |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``t5-11B``                                                 | | ~11B parameters with 24-layers, 1024-hidden-state, 65536 feed-forward hidden-state, 128-heads,                                      |
+|                   |                                                            | | Trained on English text: the Colossal Clean Crawled Corpus (C4)                                                                     |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 
 
 .. <https://huggingface.co/transformers/examples.html>`__

templates/adding_a_new_model/configuration_xxx.py

@@ -84,7 +84,7 @@ class XxxConfig(PretrainedConfig):
                  summary_first_dropout=0.1,
                  **kwargs):
         super(XxxConfig, self).__init__(**kwargs)
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, six.string_types) else -1
+        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd

templates/adding_a_new_model/convert_xxx_original_tf_checkpoint_to_pytorch.py

@@ -26,9 +26,9 @@ from transformers import XxxConfig, XxxForPreTraining, load_tf_weights_in_xxx
 import logging
 logging.basicConfig(level=logging.INFO)
 
-def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, xxx_config_file, pytorch_dump_path):
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
     # Initialise PyTorch model
-    config = XxxConfig.from_json_file(xxx_config_file)
+    config = XxxConfig.from_json_file(config_file)
     print("Building PyTorch model from configuration: {}".format(str(config)))
     model = XxxForPreTraining(config)
 
@@ -48,11 +48,11 @@ if __name__ == "__main__":
                         type = str,
                         required = True,
                         help = "Path to the TensorFlow checkpoint path.")
-    parser.add_argument("--xxx_config_file",
+    parser.add_argument("--config_file",
                         default = None,
                         type = str,
                         required = True,
-                        help = "The config json file corresponding to the pre-trained XXX model. \n"
+                        help = "The config json file corresponding to the pre-trained model. \n"
                             "This specifies the model architecture.")
     parser.add_argument("--pytorch_dump_path",
                         default = None,
@@ -61,5 +61,5 @@ if __name__ == "__main__":
                         help = "Path to the output PyTorch model.")
     args = parser.parse_args()
     convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
-                                     args.xxx_config_file,
+                                     args.config_file,
                                      args.pytorch_dump_path)

templates/adding_a_new_model/modeling_tf_xxx.py

@@ -26,6 +26,8 @@ import logging
 import math
 import os
 import sys
+import copy
+import itertools
 from io import open
 
 import numpy as np

templates/adding_a_new_model/modeling_xxx.py

@@ -25,6 +25,8 @@ import logging
 import math
 import os
 import sys
+import copy
+import itertools
 from io import open
 
 import torch

transformers/__init__.py

@@ -48,6 +48,7 @@ from .tokenization_roberta import RobertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer
 from .tokenization_albert import AlbertTokenizer
 from .tokenization_camembert import CamembertTokenizer
+from .tokenization_t5 import T5Tokenizer
 
 # Configurations
 from .configuration_utils import PretrainedConfig
@@ -58,12 +59,12 @@ from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CON
 from .configuration_gpt2 import GPT2Config, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlnet import XLNetConfig, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
-from .configuration_ctrl import CTRLConfig, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_xlm import XLMConfig, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_roberta import RobertaConfig, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_distilbert import DistilBertConfig, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_albert import AlbertConfig, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_camembert import CamembertConfig, CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_t5 import T5Config, T5_PRETRAINED_CONFIG_ARCHIVE_MAP
 
 # Modeling
 if is_torch_available():
@@ -77,8 +78,8 @@ if is_torch_available():
                                 BertForTokenClassification, BertForQuestionAnswering,
                                 load_tf_weights_in_bert, BERT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_openai import (OpenAIGPTPreTrainedModel, OpenAIGPTModel,
-                                OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
-                                load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
+                                  OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
+                                  load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_transfo_xl import (TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
                                     AdaptiveEmbedding,
                                     load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
@@ -110,6 +111,9 @@ if is_torch_available():
                                 CamembertForTokenClassification,
                                 CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
     from .modeling_encoder_decoder import PreTrainedEncoderDecoder, Model2Model
+    from .modeling_t5 import (T5PreTrainedModel, T5Model, T5WithLMHeadModel,
+                              load_tf_weights_in_t5,
+                              T5_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_albert import (AlbertPreTrainedModel, AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification,
                                 AlbertForQuestionAnswering,
@@ -178,6 +182,10 @@ if is_tf_available():
     from .modeling_tf_albert import (TFAlbertPreTrainedModel, TFAlbertModel, TFAlbertForMaskedLM,
                                      TFAlbertForSequenceClassification,
                                     TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+
+    from .modeling_tf_t5 import (TFT5PreTrainedModel, TFT5Model, TFT5WithLMHeadModel,
+                                 TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP)
+
     # Optimization
     from .optimization_tf import (WarmUp, create_optimizer, AdamWeightDecay, GradientAccumulator)
 

transformers/__main__.py

@@ -6,6 +6,7 @@ def main():
         "This command line utility let you convert original (author released) model checkpoint to pytorch.\n"
         "It should be used as one of: \n"
         ">> transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
+        ">> transformers t5 TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
         ">> transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
         ">> transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
         ">> transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
@@ -21,6 +22,23 @@ def main():
                     "https://www.tensorflow.org/install/ for installation instructions.")
                 raise
 
+            if len(sys.argv) != 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
+            else:
+                PYTORCH_DUMP_OUTPUT = sys.argv.pop()
+                TF_CONFIG = sys.argv.pop()
+                TF_CHECKPOINT = sys.argv.pop()
+                convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "t5":
+            try:
+                from .convert_t5_original_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
+            except ImportError:
+                print("transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
             if len(sys.argv) != 5:
                 # pylint: disable=line-too-long
                 print("Should be used as `transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")

transformers/configuration_auto.py

@@ -29,6 +29,7 @@ from .configuration_distilbert import DistilBertConfig
 from .configuration_ctrl import CTRLConfig
 from .configuration_camembert import CamembertConfig
 from .configuration_albert import AlbertConfig
+from .configuration_t5 import T5Config
 
 logger = logging.getLogger(__name__)
 
@@ -68,6 +69,7 @@ class AutoConfig(object):
 
         The configuration class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Config (T5 model)
             - contains `distilbert`: DistilBertConfig (DistilBERT model)
             - contains `albert`: AlbertConfig (ALBERT model)
             - contains `camembert`: CamembertConfig (CamemBERT model)
@@ -124,7 +126,9 @@ class AutoConfig(object):
             assert unused_kwargs == {'foo': False}
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5Config.from_pretrained(pretrained_model_name_or_path, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         elif 'albert' in pretrained_model_name_or_path:
             return AlbertConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)

transformers/configuration_t5.py

+# coding=utf-8
+# Copyright 2010, The T5 Authors and HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" T5 model configuration """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import sys
+import six
+from io import open
+
+from .configuration_utils import PretrainedConfig
+
+logger = logging.getLogger(__name__)
+
+T5_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-config.json",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-config.json",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-config.json",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-config.json",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-config.json",
+}
+
+
+class T5Config(PretrainedConfig):
+    r"""
+        :class:`~transformers.T5Config` is the configuration class to store the configuration of a
+        `T5Model`.
+
+
+        Arguments:
+            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `T5Model`.
+            hidden_size: Size of the encoder layers and the pooler layer.
+            num_hidden_layers: Number of hidden layers in the Transformer encoder.
+            num_attention_heads: Number of attention heads for each attention layer in
+                the Transformer encoder.
+            intermediate_size: The size of the "intermediate" (i.e., feed-forward)
+                layer in the Transformer encoder.
+            hidden_act: The non-linear activation function (function or string) in the
+                encoder and pooler. If string, "gelu", "relu", "swish" and "gelu_new" are supported.
+            hidden_dropout_prob: The dropout probabilitiy for all fully connected
+                layers in the embeddings, encoder, and pooler.
+            attention_probs_dropout_prob: The dropout ratio for the attention
+                probabilities.
+            max_position_embeddings: The maximum sequence length that this model might
+                ever be used with. Typically set this to something large just in case
+                (e.g., 512 or 1024 or 2048).
+            type_vocab_size: The vocabulary size of the `token_type_ids` passed into
+                `T5Model`.
+            initializer_factor: A factor for initializing all weight matrices (should be kept to 1.0, used for initialization testing).
+            layer_norm_eps: The epsilon used by LayerNorm.
+    """
+    pretrained_config_archive_map = T5_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+    def __init__(self,
+                 vocab_size_or_config_json_file=32128,
+                 n_positions=512,
+                 d_model=512,
+                 d_kv=64,
+                 d_ff=2048,
+                 num_layers=6,
+                 num_heads=8,
+                 relative_attention_num_buckets=32,
+                 dropout_rate=0.1,
+                 layer_norm_epsilon=1e-6,
+                 initializer_factor=1.0,
+                 **kwargs):
+        super(T5Config, self).__init__(**kwargs)
+        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.n_positions = n_positions
+        self.d_model = d_model
+        self.d_kv = d_kv
+        self.d_ff = d_ff
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.dropout_rate = dropout_rate
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_factor = initializer_factor
+
+        if isinstance(vocab_size_or_config_json_file, six.string_types):
+            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
+                json_config = json.loads(reader.read())
+            for key, value in json_config.items():
+                self.__dict__[key] = value
+        elif not isinstance(vocab_size_or_config_json_file, int):
+            raise ValueError(
+                "First argument must be either a vocabulary size (int)"
+                "or the path to a pretrained model config file (str)"
+            )
+
+    @property
+    def max_position_embeddings(self):
+        return self.n_positions
+
+    @property
+    def hidden_size(self):
+        return self.d_model
+
+    @property
+    def num_attention_heads(self):
+        return self.num_heads
+
+    @property
+    def num_hidden_layers(self):
+        return self.num_layers

transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -34,7 +34,8 @@ from transformers import (load_pytorch_checkpoint_in_tf2_model,
                                   RobertaConfig, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   DistilBertConfig, TFDistilBertForMaskedLM, TFDistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
                                   CTRLConfig, TFCTRLLMHeadModel, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
-                                  AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
+                                  AlbertConfig, TFAlbertForMaskedLM, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  T5Config, TFT5WithLMHeadModel, T5_PRETRAINED_CONFIG_ARCHIVE_MAP)
 
 if is_torch_available():
     import torch
@@ -48,7 +49,8 @@ if is_torch_available():
                                       RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
                                       CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
-                                      AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+                                      AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+                                      T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP)
 else:
     (BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
@@ -59,7 +61,8 @@ else:
     RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
     DistilBertForMaskedLM, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
     CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
-    AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP) = (
+    AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+    T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP) = (
         None, None, None, None,
         None, None,
         None, None,
@@ -69,6 +72,7 @@ else:
         None, None, None,
         None, None, None,
         None, None,
+        None, None,
         None, None)
 
 
@@ -90,7 +94,8 @@ MODEL_CLASSES = {
     'distilbert': (DistilBertConfig, TFDistilBertForMaskedLM, DistilBertForMaskedLM, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'distilbert-base-uncased-distilled-squad': (DistilBertConfig, TFDistilBertForQuestionAnswering, DistilBertForQuestionAnswering, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP, DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
     'ctrl': (CTRLConfig, TFCTRLLMHeadModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP, CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP),
-    'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
+    'albert': (AlbertConfig, TFAlbertForMaskedLM, AlbertForMaskedLM, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP, ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP),
+    't5': (T5Config, TFT5WithLMHeadModel, T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP, T5_PRETRAINED_CONFIG_ARCHIVE_MAP),
 }
 
 def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file, tf_dump_path, compare_with_pt_model=False, use_cached_models=True):
@@ -115,24 +120,21 @@ def convert_pt_checkpoint_to_tf(model_type, pytorch_checkpoint_path, config_file
     tf_model = load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path)
 
     if compare_with_pt_model:
-        inputs_list = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-        tf_inputs = tf.constant(inputs_list)
-        tfo = tf_model(tf_inputs, training=False)  # build the network
+        tfo = tf_model(tf_model.dummy_inputs, training=False)  # build the network
 
         state_dict = torch.load(pytorch_checkpoint_path, map_location='cpu')
         pt_model = pt_model_class.from_pretrained(pretrained_model_name_or_path=None,
                                                   config=config,
                                                   state_dict=state_dict)
 
-        pt_inputs = torch.tensor(inputs_list)
         with torch.no_grad():
-            pto = pt_model(pt_inputs)
+            pto = pt_model(**pt_model.dummy_inputs)
 
-        np_pt = pto[0].detach().numpy()
+        np_pt = pto[0].numpy()
         np_tf = tfo[0].numpy()
         diff = np.amax(np.abs(np_pt - np_tf))
         print("Max absolute difference between models outputs {}".format(diff))
-        assert diff <= 2e-2, "Error, model absolute difference is >2e-2"
+        assert diff <= 2e-2, "Error, model absolute difference is >2e-2: {}".format(diff)
 
     # Save pytorch-model
     print("Save TensorFlow model to {}".format(tf_dump_path))

transformers/convert_t5_original_tf_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2018 The 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.
+"""Convert T5 checkpoint."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import torch
+
+from transformers import T5Config, T5Model, load_tf_weights_in_t5
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = T5Config.from_json_file(config_file)
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    model = T5Model(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_t5(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--tf_checkpoint_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--config_file",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "The config json file corresponding to the pre-trained T5 model. \n"
+                            "This specifies the model architecture.")
+    parser.add_argument("--pytorch_dump_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
+                                     args.config_file,
+                                     args.pytorch_dump_path)

transformers/file_utils.py

@@ -73,8 +73,13 @@ TF2_WEIGHTS_NAME = 'tf_model.h5'
 TF_WEIGHTS_NAME = 'model.ckpt'
 CONFIG_NAME = "config.json"
 
+
+DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
+DUMMY_MASK = [[1, 1, 1, 1, 1], [1, 1, 1, 0, 0], [0, 0, 0, 1, 1]]
+
 S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
 
+
 def is_torch_available():
     return _torch_available
 

transformers/modeling_auto.py

@@ -29,6 +29,7 @@ from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequen
 from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, DistilBertForSequenceClassification
 from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice
 from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, AlbertForQuestionAnswering
+from .modeling_t5 import T5Model, T5WithLMHeadModel
 
 from .modeling_utils import PreTrainedModel, SequenceSummary
 
@@ -49,6 +50,7 @@ class AutoModel(object):
 
         The base model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Model (T5 model)
             - contains `distilbert`: DistilBertModel (DistilBERT model)
             - contains `albert`: AlbertModel (ALBERT model)
             - contains `camembert`: CamembertModel (CamemBERT model)
@@ -74,6 +76,7 @@ class AutoModel(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Model (T5 model)
             - contains `distilbert`: DistilBertModel (DistilBERT model)
             - contains `albert`: AlbertModel (ALBERT model)
             - contains `camembert`: CamembertModel (CamemBERT model)
@@ -146,7 +149,9 @@ class AutoModel(object):
             model = AutoModel.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'albert' in pretrained_model_name_or_path:
             return AlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
@@ -185,6 +190,7 @@ class AutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5ModelWithLMHead (T5 model)
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
             - contains `albert`: AlbertForMaskedLM (ALBERT model)
             - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
@@ -213,6 +219,7 @@ class AutoModelWithLMHead(object):
 
         The model class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5ModelWithLMHead (T5 model)
             - contains `distilbert`: DistilBertForMaskedLM (DistilBERT model)
             - contains `albert`: AlbertForMaskedLM (ALBERT model)
             - contains `camembert`: CamembertForMaskedLM (CamemBERT model)
@@ -284,7 +291,9 @@ class AutoModelWithLMHead(object):
             model = AutoModelWithLMHead.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'albert' in pretrained_model_name_or_path:
             return AlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

transformers/modeling_encoder_decoder.py

@@ -219,9 +219,7 @@ 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_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)
+        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, 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_auto.py

@@ -27,6 +27,7 @@ from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceC
 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 .modeling_tf_t5 import TFT5Model, TFT5WithLMHeadModel
 
 from .file_utils import add_start_docstrings
 
@@ -45,6 +46,7 @@ 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)
@@ -68,6 +70,7 @@ 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)
@@ -137,7 +140,9 @@ 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 'roberta' in pretrained_model_name_or_path:
             return TFRobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
@@ -173,6 +178,7 @@ 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)
@@ -199,6 +205,7 @@ 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)
@@ -269,7 +276,9 @@ 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 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

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)
 
@@ -134,7 +135,7 @@ 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:
@@ -159,7 +160,8 @@ 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)
@@ -169,6 +171,8 @@ 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
@@ -272,7 +276,7 @@ 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)

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)
+        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)
+        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_utils.py

@@ -24,14 +24,12 @@ import os
 import tensorflow as tf
 
 from .configuration_utils import PretrainedConfig
-from .file_utils import (TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, WEIGHTS_NAME,
+from .file_utils import (TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, WEIGHTS_NAME, DUMMY_INPUTS,
                          cached_path, hf_bucket_url, is_remote_url)
 from .modeling_tf_pytorch_utils import load_pytorch_checkpoint_in_tf2_model
 
 logger = logging.getLogger(__name__)
 
-DUMMY_INPUTS = [[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]]
-
 class TFPreTrainedModel(tf.keras.Model):
     r""" Base class for all TF models.
 
@@ -60,7 +58,7 @@ class TFPreTrainedModel(tf.keras.Model):
         Returns:
             tf.Tensor with dummy inputs
         """
-        return tf.constant(DUMMY_INPUTS)
+        return {'input_ids': tf.constant(DUMMY_INPUTS)}
 
     def __init__(self, config, *inputs, **kwargs):
         super(TFPreTrainedModel, self).__init__(*inputs, **kwargs)