add example for Model2Model in quickstart

Updated documentation due to typo

README.md

@@ -56,6 +56,7 @@ Choose the right framework for every part of a model's lifetime
 | [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
 | [Quick tour: TF 2.0 and PyTorch ](#Quick-tour-TF-20-training-and-PyTorch-interoperability) | Train a TF 2.0 model in 10 lines of code, load it in PyTorch |
 | [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
+| [Quick tour: Share your models ](#Quick-tour-of-model-sharing) | Upload and share your fine-tuned models with the community |
 | [Migrating from pytorch-transformers to transformers](#Migrating-from-pytorch-transformers-to-transformers) | Migrating your code from pytorch-transformers to transformers |
 | [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-transformers) | Migrating your code from pytorch-pretrained-bert to transformers |
 | [Documentation][(v2.2.0/v2.2.1/v2.2.2)](https://huggingface.co/transformers/v2.2.0) [(v2.1.1)](https://huggingface.co/transformers/v2.1.1) [(v2.0.0)](https://huggingface.co/transformers/v2.0.0) [(v1.2.0)](https://huggingface.co/transformers/v1.2.0) [(v1.1.0)](https://huggingface.co/transformers/v1.1.0) [(v1.0.0)](https://huggingface.co/transformers/v1.0.0) [(master)](https://huggingface.co/transformers) | Full API documentation and more |
@@ -144,7 +145,8 @@ 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).
 
@@ -445,6 +447,46 @@ python ./examples/run_generation.py \
     --repetition_penalty=1.2 \
 ```
 
+## Quick tour of model sharing
+
+New in `v2.2.2`: you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
+
+**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
+
+```shell
+transformers-cli login
+# log in using the same credentials as on huggingface.co
+```
+Upload your model:
+```shell
+transformers-cli upload ./path/to/pretrained_model/
+
+# ^^ Upload folder containing weights/tokenizer/config
+# saved via `.save_pretrained()`
+
+transformers-cli upload ./config.json [--filename folder/foobar.json]
+
+# ^^ Upload a single file
+# (you can optionally override its filename, which can be nested inside a folder)
+```
+
+Your model will then be accessible through its identifier, a concatenation of your username and the folder name above:
+```python
+"username/model_name"
+```
+
+Anyone can load it from code:
+```python
+tokenizer = AutoTokenizer.from_pretrained("username/pretrained_model")
+model = AutoModel.from_pretrained("username/pretrained_model")
+```
+
+Finally, list all your files on S3:
+```shell
+transformers-cli ls
+# List all your S3 objects.
+```
+
 ## Migrating from pytorch-transformers to transformers
 
 Here is a quick summary of what you should take care of when migrating from `pytorch-transformers` to `transformers`.

docs/source/index.rst

@@ -58,6 +58,7 @@ The library currently contains PyTorch and Tensorflow implementations, pre-train
     installation
     quickstart
     pretrained_models
+    model_sharing
     examples
     notebooks
     serialization

docs/source/model_sharing.md

@@ -0,0 +1,40 @@
+# Model upload and sharing
+
+Starting with `v2.2.2`, you can now upload and share your fine-tuned models with the community, using the <abbr title="Command-line interface">CLI</abbr> that's built-in to the library.
+
+**First, create an account on [https://huggingface.co/join](https://huggingface.co/join)**. Then:
+
+```shell
+transformers-cli login
+# log in using the same credentials as on huggingface.co
+```
+Upload your model:
+```shell
+transformers-cli upload ./path/to/pretrained_model/
+
+# ^^ Upload folder containing weights/tokenizer/config
+# saved via `.save_pretrained()`
+
+transformers-cli upload ./config.json [--filename folder/foobar.json]
+
+# ^^ Upload a single file
+# (you can optionally override its filename, which can be nested inside a folder)
+```
+
+Your model will then be accessible through its identifier, a concatenation of your username and the folder name above:
+```python
+"username/pretrained_model"
+```
+
+Anyone can load it from code:
+```python
+tokenizer = AutoTokenizer.from_pretrained("username/pretrained_model")
+model = AutoModel.from_pretrained("username/pretrained_model")
+```
+
+Finally, list all your files on S3:
+```shell
+transformers-cli ls
+# List all your S3 objects.
+```
+

docs/source/pretrained_models.rst

@@ -79,6 +79,14 @@ Here is the full list of the currently provided pretrained models together with
 |                   | ``bert-base-japanese-char-whole-word-masking``             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | Trained on Japanese text using Whole-Word-Masking. Text is tokenized into characters.                                               |
 |                   |                                                            | (see `details on cl-tohoku repository <https://github.com/cl-tohoku/bert-japanese>`__).                                               |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-finnish-cased-v1``                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased Finnish text.                                                                                                      |
+|                   |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-finnish-uncased-v1``                           | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on uncased Finnish text.                                                                                                    |
+|                   |                                                            | (see `details on turkunlp.org <http://turkunlp.org/FinBERT/>`__).                                                                     |
 +-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
 |                   |                                                            | | OpenAI GPT English model                                                                                                            |
@@ -217,6 +225,21 @@ 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>`__

docs/source/quickstart.md

@@ -219,4 +219,97 @@ sequence = tokenizer.decode(generated)
 print(sequence)
 ```
 
-The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
+The model only requires a single token as input as all the previous tokens' key/value pairs are contained in the `past`.
+
+### Model2Model example
+
+Encoder-decoder architectures require two tokenized inputs: one for the encoder and the other one for the decoder. Let's assume that we want to use `Model2Model` for generative question answering, and start by tokenizing the question and answer that will be fed to the model.
+
+```python
+import torch
+from transformers import BertTokenizer, Model2Model
+
+# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# Load pre-trained model tokenizer (vocabulary)
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+# Encode the input to the encoder (the question)
+question = "Who was Jim Henson?"
+encoded_question = tokenizer.encode(question)
+
+# Encode the input to the decoder (the answer)
+answer = "Jim Henson was a puppeteer"
+encoded_answer = tokenizer.encode(answer)
+
+# Convert inputs to PyTorch tensors
+question_tensor = torch.tensor([encoded_question])
+answer_tensor = torch.tensor([encoded_answer])
+```
+
+Let's see how we can use `Model2Model` to get the value of the loss associated with this (question, answer) pair:
+
+```python
+# In order to compute the loss we need to provide language model
+# labels (the token ids that the model should have produced) to
+# the decoder.
+lm_labels =  encoded_answer
+labels_tensor = torch.tensor([lm_labels])
+
+# Load pre-trained model (weights)
+model = Model2Model.from_pretrained('bert-base-uncased')
+
+# Set the model in evaluation mode to deactivate the DropOut modules
+# This is IMPORTANT to have reproducible results during evaluation!
+model.eval()
+
+# If you have a GPU, put everything on cuda
+question_tensor = question_tensor.to('cuda')
+answer_tensor = answer_tensor.to('cuda')
+labels_tensor = labels_tensor.to('cuda')
+model.to('cuda')
+
+# Predict hidden states features for each layer
+with torch.no_grad():
+    # See the models docstrings for the detail of the inputs
+    outputs = model(question_tensor, answer_tensor, decoder_lm_labels=labels_tensor)
+    # Transformers models always output tuples.
+    # See the models docstrings for the detail of all the outputs
+    # In our case, the first element is the value of the LM loss 
+    lm_loss = outputs[0]
+```
+
+This loss can be used to fine-tune `Model2Model` on the question answering task. Assuming that we fine-tuned the model, let us now see how to generate an answer:
+
+```python
+# Let's re-use the previous question
+question = "Who was Jim Henson?"
+encoded_question = tokenizer.encode(question)
+question_tensor = torch.tensor([encoded_question])
+
+# This time we try to generate the answer, so we start with an empty sequence
+answer = "[CLS]"
+encoded_answer = tokenizer.encode(answer, add_special_tokens=False)
+answer_tensor = torch.tensor([encoded_answer])
+
+# Load pre-trained model (weights)
+model = Model2Model.from_pretrained('fine-tuned-weights')
+model.eval()
+
+# If you have a GPU, put everything on cuda
+question_tensor = encoded_question.to('cuda')
+answer_tensor = encoded_answer.to('cuda')
+model.to('cuda')
+
+# Predict all tokens
+with torch.no_grad():
+    outputs = model(question_tensor, answer_tensor)
+    predictions = outputs[0]
+
+# confirm we were able to predict 'jim'
+predicted_index = torch.argmax(predictions[0, -1]).item()
+predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
+assert predicted_token == 'jim'
+```

examples/README.md

@@ -467,7 +467,7 @@ Training with the previously defined hyper-parameters yields the following resul
 ## Named Entity Recognition
 
 Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
-[`run_tf_ner.py`(https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py)] for Tensorflow 2.
+[`run_tf_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py) for Tensorflow 2.
 This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
 Details and results for the fine-tuning provided by @stefan-it.
 

examples/run_generation.py

@@ -247,7 +247,11 @@ def main():
         out = out[:, len(context_tokens):].tolist()
         for o in out:
             text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
-            text = text[: text.find(args.stop_token) if args.stop_token else None]
+            if args.stop_token:
+                index =  text.find(args.stop_token)
+                if index == -1:
+                    index = None
+                text = text[:index]
 
             print(text)
 

examples/run_squad.py

@@ -61,7 +61,6 @@ MODEL_CLASSES = {
     'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
     'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
     'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
-    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer)
 }
 
 def set_seed(args):
@@ -223,7 +222,7 @@ def evaluate(args, model, tokenizer, prefix=""):
     eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
     # multi-gpu evaluate
-    if args.n_gpu > 1:
+    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
         model = torch.nn.DataParallel(model)
 
     # Eval!

examples/summarization/configuration_bertabs.py

@@ -33,6 +33,8 @@ class BertAbsConfig(PretrainedConfig):
     r""" Class to store the configuration of the BertAbs model.
 
     Arguments:
+        vocab_size: int
+            Number of tokens in the vocabulary.
         max_pos: int
             The maximum sequence length that this model will be used with.
         enc_layer: int
@@ -65,7 +67,7 @@ class BertAbsConfig(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=30522,
+        vocab_size=30522,
         max_pos=512,
         enc_layers=6,
         enc_hidden_size=512,
@@ -81,39 +83,17 @@ class BertAbsConfig(PretrainedConfig):
     ):
         super(BertAbsConfig, self).__init__(**kwargs)
 
-        if self._input_is_path_to_json(vocab_size_or_config_json_file):
-            path_to_json = vocab_size_or_config_json_file
-            with open(path_to_json, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.vocab_size = vocab_size_or_config_json_file
-            self.max_pos = max_pos
+        self.vocab_size = vocab_size
+        self.max_pos = max_pos
 
-            self.enc_layers = enc_layers
-            self.enc_hidden_size = enc_hidden_size
-            self.enc_heads = enc_heads
-            self.enc_ff_size = enc_ff_size
-            self.enc_dropout = enc_dropout
+        self.enc_layers = enc_layers
+        self.enc_hidden_size = enc_hidden_size
+        self.enc_heads = enc_heads
+        self.enc_ff_size = enc_ff_size
+        self.enc_dropout = enc_dropout
 
-            self.dec_layers = dec_layers
-            self.dec_hidden_size = dec_hidden_size
-            self.dec_heads = dec_heads
-            self.dec_ff_size = dec_ff_size
-            self.dec_dropout = dec_dropout
-        else:
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
-
-    def _input_is_path_to_json(self, first_argument):
-        """ Checks whether the first argument passed to config
-        is the path to a JSON file that contains the config.
-        """
-        is_python_2 = sys.version_info[0] == 2
-        if is_python_2:
-            return isinstance(first_argument, unicode)
-        else:
-            return isinstance(first_argument, str)
+        self.dec_layers = dec_layers
+        self.dec_hidden_size = dec_hidden_size
+        self.dec_heads = dec_heads
+        self.dec_ff_size = dec_ff_size
+        self.dec_dropout = dec_dropout

requirements.txt

@@ -5,7 +5,7 @@ boto3
 # Used for downloading models over HTTP
 requests
 # For OpenAI GPT
-regex
+regex != 2019.12.17
 # For XLNet
 sentencepiece
 # For XLM

setup.py

@@ -59,7 +59,7 @@ setup(
                       'boto3',
                       'requests',
                       'tqdm',
-                      'regex',
+                      'regex != 2019.12.17',
                       'sentencepiece',
                       'sacremoses'],
     entry_points={

templates/adding_a_new_model/configuration_xxx.py

@@ -39,7 +39,7 @@ class XxxConfig(PretrainedConfig):
 
 
         Arguments:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XxxModel`.
+            vocab_size: Vocabulary size of `inputs_ids` in `XxxModel`.
             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
@@ -64,7 +64,7 @@ class XxxConfig(PretrainedConfig):
     pretrained_config_archive_map = XXX_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=50257,
+                 vocab_size=50257,
                  n_positions=1024,
                  n_ctx=1024,
                  n_embd=768,
@@ -75,8 +75,6 @@ class XxxConfig(PretrainedConfig):
                  attn_pdrop=0.1,
                  layer_norm_epsilon=1e-5,
                  initializer_range=0.02,
-
-                 num_labels=1,
                  summary_type='cls_index',
                  summary_use_proj=True,
                  summary_activation=None,
@@ -84,7 +82,7 @@ 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
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd
@@ -95,23 +93,11 @@ class XxxConfig(PretrainedConfig):
         self.attn_pdrop = attn_pdrop
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
-
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        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):

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

templates/adding_a_new_model/tests/modeling_tf_xxx_test.py

@@ -111,7 +111,7 @@ class TFXxxModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = XxxConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

templates/adding_a_new_model/tests/modeling_xxx_test.py

@@ -109,7 +109,7 @@ class XxxModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = XxxConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

templates/adding_a_new_model/tokenization_xxx.py

@@ -85,7 +85,7 @@ class XxxTokenizer(PreTrainedTokenizer):
 
     Args:
         vocab_file: Path to a one-wordpiece-per-line vocabulary file
-        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

transformers/__init__.py

@@ -19,7 +19,7 @@ logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 # Files and general utilities
 from .file_utils import (TRANSFORMERS_CACHE, PYTORCH_TRANSFORMERS_CACHE, PYTORCH_PRETRAINED_BERT_CACHE,
                          cached_path, add_start_docstrings, add_end_docstrings,
-                         WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME,
+                         WEIGHTS_NAME, TF2_WEIGHTS_NAME, TF_WEIGHTS_NAME, CONFIG_NAME, MODEL_CARD_NAME,
                          is_tf_available, is_torch_available)
 
 from .data import (is_sklearn_available,
@@ -33,6 +33,9 @@ from .data import (is_sklearn_available,
 if is_sklearn_available():
     from .data import glue_compute_metrics, xnli_compute_metrics
 
+# Model Cards
+from .model_card import ModelCard
+
 # Tokenizers
 from .tokenization_utils import (PreTrainedTokenizer)
 from .tokenization_auto import AutoTokenizer
@@ -48,28 +51,29 @@ 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
-from .configuration_auto import AutoConfig
+from .configuration_auto import AutoConfig, ALL_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
 from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
 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():
     from .modeling_utils import (PreTrainedModel, prune_layer, Conv1D)
     from .modeling_auto import (AutoModel, AutoModelForSequenceClassification, AutoModelForQuestionAnswering,
-                                AutoModelWithLMHead)
+                                AutoModelWithLMHead, ALL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_bert import (BertPreTrainedModel, BertModel, BertForPreTraining,
                                 BertForMaskedLM, BertForNextSentencePrediction,
@@ -77,8 +81,8 @@ 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 +114,9 @@ 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,
@@ -124,7 +131,7 @@ if is_torch_available():
 if is_tf_available():
     from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, TFSequenceSummary, shape_list
     from .modeling_tf_auto import (TFAutoModel, TFAutoModelForSequenceClassification, TFAutoModelForQuestionAnswering,
-                                   TFAutoModelWithLMHead)
+                                   TFAutoModelWithLMHead, TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP)
 
     from .modeling_tf_bert import (TFBertPreTrainedModel, TFBertMainLayer, TFBertEmbeddings,
                                    TFBertModel, TFBertForPreTraining,
@@ -178,6 +185,10 @@ 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_albert.py

@@ -37,7 +37,7 @@ class AlbertConfig(PretrainedConfig):
     pretrained_config_archive_map = ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30000,
+                 vocab_size=30000,
                  embedding_size=128,
                  hidden_size=4096,
                  num_hidden_layers=12,
@@ -83,7 +83,7 @@ class AlbertConfig(PretrainedConfig):
         """
         super(AlbertConfig, self).__init__(**kwargs)
 
-        self.vocab_size = vocab_size_or_config_json_file
+        self.vocab_size = vocab_size
         self.embedding_size = embedding_size
         self.hidden_size = hidden_size
         self.num_hidden_layers = num_hidden_layers
@@ -97,4 +97,4 @@ class AlbertConfig(PretrainedConfig):
         self.max_position_embeddings = max_position_embeddings
         self.type_vocab_size = type_vocab_size
         self.initializer_range = initializer_range
-        self.layer_norm_eps = layer_norm_eps
+        self.layer_norm_eps = layer_norm_eps

transformers/configuration_auto.py

@@ -18,21 +18,40 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 
-from .configuration_bert import BertConfig
-from .configuration_openai import OpenAIGPTConfig
-from .configuration_gpt2 import GPT2Config
-from .configuration_transfo_xl import TransfoXLConfig
-from .configuration_xlnet import XLNetConfig
-from .configuration_xlm import XLMConfig
-from .configuration_roberta import RobertaConfig
-from .configuration_distilbert import DistilBertConfig
-from .configuration_ctrl import CTRLConfig
-from .configuration_camembert import CamembertConfig
-from .configuration_albert import AlbertConfig
+from .configuration_bert import BertConfig, BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_openai import OpenAIGPTConfig, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP
+from .configuration_transfo_xl import TransfoXLConfig, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
+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_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
 
 logger = logging.getLogger(__name__)
 
 
+ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict((key, value)
+    for pretrained_map in [
+        BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        T5_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        ]
+    for key, value, in pretrained_map.items())
+
+
 class AutoConfig(object):
     r""":class:`~transformers.AutoConfig` is a generic configuration class
         that will be instantiated as one of the configuration classes of the library
@@ -68,6 +87,7 @@ 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 +144,9 @@ 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_bert.py

@@ -45,7 +45,9 @@ BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-config.json",
     'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-config.json",
     'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-config.json",
-    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-config.json"
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-config.json",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/config.json",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/config.json",
 }
 
 
@@ -56,7 +58,7 @@ class BertConfig(PretrainedConfig):
 
 
         Arguments:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `BertModel`.
+            vocab_size: Vocabulary size of `inputs_ids` in `BertModel`.
             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
@@ -81,7 +83,7 @@ class BertConfig(PretrainedConfig):
     pretrained_config_archive_map = BERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30522,
+                 vocab_size=30522,
                  hidden_size=768,
                  num_hidden_layers=12,
                  num_attention_heads=12,
@@ -95,25 +97,15 @@ class BertConfig(PretrainedConfig):
                  layer_norm_eps=1e-12,
                  **kwargs):
         super(BertConfig, self).__init__(**kwargs)
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.vocab_size = vocab_size_or_config_json_file
-            self.hidden_size = hidden_size
-            self.num_hidden_layers = num_hidden_layers
-            self.num_attention_heads = num_attention_heads
-            self.hidden_act = hidden_act
-            self.intermediate_size = intermediate_size
-            self.hidden_dropout_prob = hidden_dropout_prob
-            self.attention_probs_dropout_prob = attention_probs_dropout_prob
-            self.max_position_embeddings = max_position_embeddings
-            self.type_vocab_size = type_vocab_size
-            self.initializer_range = initializer_range
-            self.layer_norm_eps = layer_norm_eps
-        else:
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps

transformers/configuration_ctrl.py

@@ -31,7 +31,7 @@ class CTRLConfig(PretrainedConfig):
     """Configuration class to store the configuration of a `CTRLModel`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
+        vocab_size: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
         n_positions: Number of positional embeddings.
         n_ctx: Size of the causal mask (usually same as n_positions).
         dff: Size of the inner dimension of the FFN.
@@ -52,7 +52,7 @@ class CTRLConfig(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=246534,
+        vocab_size=246534,
         n_positions=256,
         n_ctx=256,
         n_embd=1280,
@@ -64,8 +64,6 @@ class CTRLConfig(PretrainedConfig):
         attn_pdrop=0.1,
         layer_norm_epsilon=1e-6,
         initializer_range=0.02,
-
-        num_labels=1,
         summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
@@ -76,7 +74,7 @@ class CTRLConfig(PretrainedConfig):
         """Constructs CTRLConfig.
 
         Args:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
+            vocab_size: Vocabulary size of `inputs_ids` in `CTRLModel` or a configuration json file.
             n_positions: Number of positional embeddings.
             n_ctx: Size of the causal mask (usually same as n_positions).
             dff: Size of the inner dimension of the FFN.
@@ -94,8 +92,7 @@ class CTRLConfig(PretrainedConfig):
                 initializing all weight matrices.
         """
         super(CTRLConfig, self).__init__(**kwargs)
-
-        self.vocab_size = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.n_ctx = n_ctx
         self.n_positions = n_positions
         self.n_embd = n_embd
@@ -108,23 +105,11 @@ class CTRLConfig(PretrainedConfig):
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
 
-        self.num_labels = num_labels
         self.summary_type = summary_type
         self.summary_use_proj = summary_use_proj
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif 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):

transformers/configuration_distilbert.py

@@ -37,7 +37,7 @@ class DistilBertConfig(PretrainedConfig):
     pretrained_config_archive_map = DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30522,
+                 vocab_size=30522,
                  max_position_embeddings=512,
                  sinusoidal_pos_embds=False,
                  n_layers=6,
@@ -53,31 +53,21 @@ class DistilBertConfig(PretrainedConfig):
                  seq_classif_dropout=0.2,
                  **kwargs):
         super(DistilBertConfig, self).__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.sinusoidal_pos_embds = sinusoidal_pos_embds
+        self.n_layers = n_layers
+        self.n_heads = n_heads
+        self.dim = dim
+        self.hidden_dim = hidden_dim
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation = activation
+        self.initializer_range = initializer_range
+        self.tie_weights_ = tie_weights_
+        self.qa_dropout = qa_dropout
+        self.seq_classif_dropout = seq_classif_dropout
 
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.vocab_size = vocab_size_or_config_json_file
-            self.max_position_embeddings = max_position_embeddings
-            self.sinusoidal_pos_embds = sinusoidal_pos_embds
-            self.n_layers = n_layers
-            self.n_heads = n_heads
-            self.dim = dim
-            self.hidden_dim = hidden_dim
-            self.dropout = dropout
-            self.attention_dropout = attention_dropout
-            self.activation = activation
-            self.initializer_range = initializer_range
-            self.tie_weights_ = tie_weights_
-            self.qa_dropout = qa_dropout
-            self.seq_classif_dropout = seq_classif_dropout
-        else:
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
     @property
     def hidden_size(self):
         return self.dim

transformers/configuration_gpt2.py

@@ -36,7 +36,7 @@ class GPT2Config(PretrainedConfig):
     """Configuration class to store the configuration of a `GPT2Model`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
+        vocab_size: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
         n_positions: Number of positional embeddings.
         n_ctx: Size of the causal mask (usually same as n_positions).
         n_embd: Dimensionality of the embeddings and hidden states.
@@ -56,7 +56,7 @@ class GPT2Config(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=50257,
+        vocab_size=50257,
         n_positions=1024,
         n_ctx=1024,
         n_embd=768,
@@ -67,8 +67,6 @@ class GPT2Config(PretrainedConfig):
         attn_pdrop=0.1,
         layer_norm_epsilon=1e-5,
         initializer_range=0.02,
-
-        num_labels=1,
         summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
@@ -79,7 +77,7 @@ class GPT2Config(PretrainedConfig):
         """Constructs GPT2Config.
 
         Args:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
+            vocab_size: Vocabulary size of `inputs_ids` in `GPT2Model` or a configuration json file.
             n_positions: Number of positional embeddings.
             n_ctx: Size of the causal mask (usually same as n_positions).
             n_embd: Dimensionality of the embeddings and hidden states.
@@ -96,37 +94,22 @@ class GPT2Config(PretrainedConfig):
                 initializing all weight matrices.
         """
         super(GPT2Config, self).__init__(**kwargs)
-
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.vocab_size = vocab_size_or_config_json_file
-            self.n_ctx = n_ctx
-            self.n_positions = n_positions
-            self.n_embd = n_embd
-            self.n_layer = n_layer
-            self.n_head = n_head
-            self.resid_pdrop = resid_pdrop
-            self.embd_pdrop = embd_pdrop
-            self.attn_pdrop = attn_pdrop
-            self.layer_norm_epsilon = layer_norm_epsilon
-            self.initializer_range = initializer_range
-
-            self.num_labels = num_labels
-            self.summary_type = summary_type
-            self.summary_use_proj = summary_use_proj
-            self.summary_activation = summary_activation
-            self.summary_first_dropout = summary_first_dropout
-            self.summary_proj_to_labels = summary_proj_to_labels
-        else:
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
 
     @property
     def max_position_embeddings(self):

transformers/configuration_openai.py

@@ -35,7 +35,7 @@ class OpenAIGPTConfig(PretrainedConfig):
     Configuration class to store the configuration of a `OpenAIGPTModel`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
+        vocab_size: Vocabulary size of `inputs_ids` in `OpenAIGPTModel` or a configuration json file.
         n_positions: Number of positional embeddings.
         n_ctx: Size of the causal mask (usually same as n_positions).
         n_embd: Dimensionality of the embeddings and hidden states.
@@ -58,7 +58,7 @@ class OpenAIGPTConfig(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=40478,
+        vocab_size=40478,
         n_positions=512,
         n_ctx=512,
         n_embd=768,
@@ -71,8 +71,6 @@ class OpenAIGPTConfig(PretrainedConfig):
         layer_norm_epsilon=1e-5,
         initializer_range=0.02,
         predict_special_tokens=True,
-
-        num_labels=1,
         summary_type='cls_index',
         summary_use_proj=True,
         summary_activation=None,
@@ -83,39 +81,24 @@ class OpenAIGPTConfig(PretrainedConfig):
         """Constructs OpenAIGPTConfig.
         """
         super(OpenAIGPTConfig, self).__init__(**kwargs)
-
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.vocab_size = vocab_size_or_config_json_file
-            self.n_ctx = n_ctx
-            self.n_positions = n_positions
-            self.n_embd = n_embd
-            self.n_layer = n_layer
-            self.n_head = n_head
-            self.afn = afn
-            self.resid_pdrop = resid_pdrop
-            self.embd_pdrop = embd_pdrop
-            self.attn_pdrop = attn_pdrop
-            self.layer_norm_epsilon = layer_norm_epsilon
-            self.initializer_range = initializer_range
-            self.predict_special_tokens = predict_special_tokens
-
-            self.num_labels = num_labels
-            self.summary_type = summary_type
-            self.summary_use_proj = summary_use_proj
-            self.summary_activation = summary_activation
-            self.summary_first_dropout = summary_first_dropout
-            self.summary_proj_to_labels = summary_proj_to_labels
-        else:
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
+        self.vocab_size = vocab_size
+        self.n_ctx = n_ctx
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.afn = afn
+        self.resid_pdrop = resid_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+        self.predict_special_tokens = predict_special_tokens
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_first_dropout = summary_first_dropout
+        self.summary_proj_to_labels = summary_proj_to_labels
 
     @property
     def max_position_embeddings(self):

transformers/configuration_t5.py

@@ -0,0 +1,108 @@
+# 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=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
+        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
+
+    @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/configuration_transfo_xl.py

@@ -34,7 +34,7 @@ class TransfoXLConfig(PretrainedConfig):
     """Configuration class to store the configuration of a `TransfoXLModel`.
 
         Args:
-            vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `TransfoXLModel` or a configuration json file.
+            vocab_size: Vocabulary size of `inputs_ids` in `TransfoXLModel` or a configuration json file.
             cutoffs: cutoffs for the adaptive softmax
             d_model: Dimensionality of the model's hidden states.
             d_embed: Dimensionality of the embeddings
@@ -68,7 +68,7 @@ class TransfoXLConfig(PretrainedConfig):
     pretrained_config_archive_map = TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=267735,
+                 vocab_size=267735,
                  cutoffs=[20000, 40000, 200000],
                  d_model=1024,
                  d_embed=1024,
@@ -100,7 +100,7 @@ class TransfoXLConfig(PretrainedConfig):
         """Constructs TransfoXLConfig.
         """
         super(TransfoXLConfig, self).__init__(**kwargs)
-        self.n_token = vocab_size_or_config_json_file if isinstance(vocab_size_or_config_json_file, int) else -1
+        self.vocab_size = vocab_size
         self.cutoffs = []
         self.cutoffs.extend(cutoffs)
         self.tie_weight = tie_weight
@@ -133,27 +133,17 @@ class TransfoXLConfig(PretrainedConfig):
         self.init_std = init_std
         self.layer_norm_epsilon = layer_norm_epsilon
 
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif 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.tgt_len + self.ext_len + self.mem_len
 
     @property
-    def vocab_size(self):
-        return self.n_token
+    def n_token(self):  # Backward compatibility
+        return self.vocab_size
 
-    @vocab_size.setter
-    def vocab_size(self, value):
-        self.n_token = value
+    @n_token.setter
+    def n_token(self, value):  # Backward compatibility
+        self.vocab_size = value
 
     @property
     def hidden_size(self):

transformers/configuration_utils.py

@@ -49,8 +49,7 @@ class PretrainedConfig(object):
     pretrained_config_archive_map = {}
 
     def __init__(self, **kwargs):
-        self.finetuning_task = kwargs.pop('finetuning_task', None)
-        self.num_labels = kwargs.pop('num_labels', 2)
+        # Attributes with defaults
         self.output_attentions = kwargs.pop('output_attentions', False)
         self.output_hidden_states = kwargs.pop('output_hidden_states', False)
         self.output_past = kwargs.pop('output_past', True)  # Not used by all models
@@ -59,6 +58,22 @@ class PretrainedConfig(object):
         self.pruned_heads = kwargs.pop('pruned_heads', {})
         self.is_decoder = kwargs.pop('is_decoder', False)
 
+        # Fine-tuning task arguments
+        self.finetuning_task = kwargs.pop('finetuning_task', None)
+        self.num_labels = kwargs.pop('num_labels', 2)
+        self.id2label = kwargs.pop('id2label', {i: 'LABEL_{}'.format(i) for i in range(self.num_labels)})
+        self.id2label = dict((int(key), value) for key, value in self.id2label.items())
+        self.label2id = kwargs.pop('label2id', dict(zip(self.id2label.values(), self.id2label.keys())))
+        self.label2id = dict((key, int(value)) for key, value in self.label2id.items())
+
+        # Additional attributes without default values
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error("Can't set {} with value {} for {}".format(key, value, self))
+                raise err
+
     def save_pretrained(self, save_directory):
         """ Save a configuration object to the directory `save_directory`, so that it
             can be re-loaded using the :func:`~transformers.PretrainedConfig.from_pretrained` class method.
@@ -136,10 +151,14 @@ class PretrainedConfig(object):
             config_file = pretrained_model_name_or_path
         else:
             config_file = hf_bucket_url(pretrained_model_name_or_path, postfix=CONFIG_NAME)
-        # redirect to the cache, if necessary
+
         try:
+            # Load from URL or cache if already cached
             resolved_config_file = cached_path(config_file, cache_dir=cache_dir, force_download=force_download,
                                                proxies=proxies, resume_download=resume_download)
+            # Load config
+            config = cls.from_json_file(resolved_config_file)
+
         except EnvironmentError:
             if pretrained_model_name_or_path in cls.pretrained_config_archive_map:
                 msg = "Couldn't reach server at '{}' to download pretrained model configuration file.".format(
@@ -153,15 +172,18 @@ class PretrainedConfig(object):
                         config_file, CONFIG_NAME)
             raise EnvironmentError(msg)
 
+        except json.JSONDecodeError:
+            msg = "Couldn't reach server at '{}' to download configuration file or " \
+                  "configuration file is not a valid JSON file. " \
+                  "Please check network or file content here: {}.".format(config_file, resolved_config_file)
+            raise EnvironmentError(msg)
+
         if resolved_config_file == config_file:
             logger.info("loading configuration file {}".format(config_file))
         else:
             logger.info("loading configuration file {} from cache at {}".format(
                 config_file, resolved_config_file))
 
-        # Load config
-        config = cls.from_json_file(resolved_config_file)
-
         if hasattr(config, 'pruned_heads'):
             config.pruned_heads = dict((int(key), value) for key, value in config.pruned_heads.items())
 
@@ -183,17 +205,15 @@ class PretrainedConfig(object):
     @classmethod
     def from_dict(cls, json_object):
         """Constructs a `Config` from a Python dictionary of parameters."""
-        config = cls(vocab_size_or_config_json_file=-1)
-        for key, value in json_object.items():
-            setattr(config, key, value)
-        return config
+        return cls(**json_object)
 
     @classmethod
     def from_json_file(cls, json_file):
         """Constructs a `Config` from a json file of parameters."""
         with open(json_file, "r", encoding='utf-8') as reader:
             text = reader.read()
-        return cls.from_dict(json.loads(text))
+        dict_obj = json.loads(text)
+        return cls(**dict_obj)
 
     def __eq__(self, other):
         return self.__dict__ == other.__dict__

transformers/configuration_xlm.py

@@ -42,7 +42,7 @@ class XLMConfig(PretrainedConfig):
     """Configuration class to store the configuration of a `XLMModel`.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of `inputs_ids` in `XLMModel`.
+        vocab_size: Vocabulary size of `inputs_ids` in `XLMModel`.
         d_model: Size of the encoder layers and the pooler layer.
         n_layer: Number of hidden layers in the Transformer encoder.
         n_head: Number of attention heads for each attention layer in
@@ -81,7 +81,7 @@ class XLMConfig(PretrainedConfig):
     pretrained_config_archive_map = XLM_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=30145,
+                 vocab_size=30145,
                  emb_dim=2048,
                  n_layers=12,
                  n_heads=16,
@@ -103,9 +103,6 @@ class XLMConfig(PretrainedConfig):
                  unk_index=3,
                  mask_index=5,
                  is_encoder=True,
-
-                 finetuning_task=None,
-                 num_labels=2,
                  summary_type='first',
                  summary_use_proj=True,
                  summary_activation=None,
@@ -117,56 +114,46 @@ class XLMConfig(PretrainedConfig):
         """Constructs XLMConfig.
         """
         super(XLMConfig, self).__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.emb_dim = emb_dim
+        self.n_layers = n_layers
+        self.n_heads = n_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.gelu_activation = gelu_activation
+        self.sinusoidal_embeddings = sinusoidal_embeddings
+        self.causal = causal
+        self.asm = asm
+        self.n_langs = n_langs
+        self.use_lang_emb = use_lang_emb
+        self.layer_norm_eps = layer_norm_eps
+        self.bos_index = bos_index
+        self.eos_index = eos_index
+        self.pad_index = pad_index
+        self.unk_index = unk_index
+        self.mask_index = mask_index
+        self.is_encoder = is_encoder
+        self.max_position_embeddings = max_position_embeddings
+        self.embed_init_std = embed_init_std
+        self.init_std = init_std
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_proj_to_labels = summary_proj_to_labels
+        self.summary_first_dropout = summary_first_dropout
+        self.start_n_top = start_n_top
+        self.end_n_top = end_n_top
 
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.n_words = vocab_size_or_config_json_file
-            self.emb_dim = emb_dim
-            self.n_layers = n_layers
-            self.n_heads = n_heads
-            self.dropout = dropout
-            self.attention_dropout = attention_dropout
-            self.gelu_activation = gelu_activation
-            self.sinusoidal_embeddings = sinusoidal_embeddings
-            self.causal = causal
-            self.asm = asm
-            self.n_langs = n_langs
-            self.use_lang_emb = use_lang_emb
-            self.layer_norm_eps = layer_norm_eps
-            self.bos_index = bos_index
-            self.eos_index = eos_index
-            self.pad_index = pad_index
-            self.unk_index = unk_index
-            self.mask_index = mask_index
-            self.is_encoder = is_encoder
-            self.max_position_embeddings = max_position_embeddings
-            self.embed_init_std = embed_init_std
-            self.init_std = init_std
-            self.finetuning_task = finetuning_task
-            self.num_labels = num_labels
-            self.summary_type = summary_type
-            self.summary_use_proj = summary_use_proj
-            self.summary_activation = summary_activation
-            self.summary_proj_to_labels = summary_proj_to_labels
-            self.summary_first_dropout = summary_first_dropout
-            self.start_n_top = start_n_top
-            self.end_n_top = end_n_top
-        else:
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
+        if "n_words" in kwargs:
+            self.n_words = kwargs["n_words"]
 
     @property
-    def vocab_size(self):
-        return self.n_words
+    def n_words(self):  # For backward compatibility
+        return self.vocab_size
 
-    @vocab_size.setter
-    def vocab_size(self, value):
-        self.n_words = value
+    @n_words.setter
+    def n_words(self, value):  # For backward compatibility
+        self.vocab_size = value
 
     @property
     def hidden_size(self):

transformers/configuration_xlnet.py

@@ -35,7 +35,7 @@ class XLNetConfig(PretrainedConfig):
     """Configuration class to store the configuration of a ``XLNetModel``.
 
     Args:
-        vocab_size_or_config_json_file: Vocabulary size of ``inputs_ids`` in ``XLNetModel``.
+        vocab_size: Vocabulary size of ``inputs_ids`` in ``XLNetModel``.
         d_model: Size of the encoder layers and the pooler layer.
         n_layer: Number of hidden layers in the Transformer encoder.
         n_head: Number of attention heads for each attention layer in
@@ -72,28 +72,22 @@ class XLNetConfig(PretrainedConfig):
     pretrained_config_archive_map = XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP
 
     def __init__(self,
-                 vocab_size_or_config_json_file=32000,
+                 vocab_size=32000,
                  d_model=1024,
                  n_layer=24,
                  n_head=16,
                  d_inner=4096,
-                 max_position_embeddings=512,
                  ff_activation="gelu",
                  untie_r=True,
                  attn_type="bi",
-
                  initializer_range=0.02,
                  layer_norm_eps=1e-12,
-
                  dropout=0.1,
                  mem_len=None,
                  reuse_len=None,
                  bi_data=False,
                  clamp_len=-1,
                  same_length=False,
-
-                 finetuning_task=None,
-                 num_labels=2,
                  summary_type='last',
                  summary_use_proj=True,
                  summary_activation='tanh',
@@ -104,58 +98,45 @@ class XLNetConfig(PretrainedConfig):
         """Constructs XLNetConfig.
         """
         super(XLNetConfig, self).__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.n_layer = n_layer
+        self.n_head = n_head
+        assert d_model % n_head == 0
+        self.d_head = d_model // n_head
+        self.ff_activation = ff_activation
+        self.d_inner = d_inner
+        self.untie_r = untie_r
+        self.attn_type = attn_type
 
-        if isinstance(vocab_size_or_config_json_file, str) or (sys.version_info[0] == 2
-                        and isinstance(vocab_size_or_config_json_file, unicode)):
-            with open(vocab_size_or_config_json_file, "r", encoding='utf-8') as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                setattr(config, key, value)
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.n_token = vocab_size_or_config_json_file
-            self.d_model = d_model
-            self.n_layer = n_layer
-            self.n_head = n_head
-            assert d_model % n_head == 0
-            self.d_head = d_model // n_head
-            self.ff_activation = ff_activation
-            self.d_inner = d_inner
-            self.untie_r = untie_r
-            self.attn_type = attn_type
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
 
-            self.initializer_range = initializer_range
-            self.layer_norm_eps = layer_norm_eps
+        self.dropout = dropout
+        self.mem_len = mem_len
+        self.reuse_len = reuse_len
+        self.bi_data = bi_data
+        self.clamp_len = clamp_len
+        self.same_length = same_length
 
-            self.dropout = dropout
-            self.mem_len = mem_len
-            self.reuse_len = reuse_len
-            self.bi_data = bi_data
-            self.clamp_len = clamp_len
-            self.same_length = same_length
-
-            self.finetuning_task = finetuning_task
-            self.num_labels = num_labels
-            self.summary_type = summary_type
-            self.summary_use_proj = summary_use_proj
-            self.summary_activation = summary_activation
-            self.summary_last_dropout = summary_last_dropout
-            self.start_n_top = start_n_top
-            self.end_n_top = end_n_top
-        else:
-            raise ValueError("First argument must be either a vocabulary size (int)"
-                             " or the path to a pretrained model config file (str)")
+        self.summary_type = summary_type
+        self.summary_use_proj = summary_use_proj
+        self.summary_activation = summary_activation
+        self.summary_last_dropout = summary_last_dropout
+        self.start_n_top = start_n_top
+        self.end_n_top = end_n_top
 
     @property
     def max_position_embeddings(self):
         return -1
 
     @property
-    def vocab_size(self):
-        return self.n_token
+    def n_token(self):  # Backward compatibility
+        return self.vocab_size
 
-    @vocab_size.setter
-    def vocab_size(self, value):
-        self.n_token = value
+    @n_token.setter
+    def n_token(self, value):  # Backward compatibility
+        self.vocab_size = value
 
     @property
     def hidden_size(self):

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_roberta_original_pytorch_checkpoint_to_pytorch.py

@@ -20,6 +20,13 @@ import argparse
 import logging
 import numpy as np
 import torch
+import pathlib
+
+import fairseq
+from packaging import version
+
+if version.parse(fairseq.__version__) < version.parse("0.9.0"):
+    raise Exception("requires fairseq >= 0.9.0")
 
 from fairseq.models.roberta import RobertaModel as FairseqRobertaModel
 from fairseq.modules import TransformerSentenceEncoderLayer
@@ -45,8 +52,9 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
     """
     roberta = FairseqRobertaModel.from_pretrained(roberta_checkpoint_path)
     roberta.eval()  # disable dropout
+    roberta_sent_encoder = roberta.model.decoder.sentence_encoder
     config = BertConfig(
-        vocab_size_or_config_json_file=50265,
+        vocab_size=roberta_sent_encoder.embed_tokens.num_embeddings,
         hidden_size=roberta.args.encoder_embed_dim,
         num_hidden_layers=roberta.args.encoder_layers,
         num_attention_heads=roberta.args.encoder_attention_heads,
@@ -64,7 +72,6 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
 
     # Now let's copy all the weights.
     # Embeddings
-    roberta_sent_encoder = roberta.model.decoder.sentence_encoder
     model.roberta.embeddings.word_embeddings.weight = roberta_sent_encoder.embed_tokens.weight
     model.roberta.embeddings.position_embeddings.weight = roberta_sent_encoder.embed_positions.weight
     model.roberta.embeddings.token_type_embeddings.weight.data = torch.zeros_like(model.roberta.embeddings.token_type_embeddings.weight)  # just zero them out b/c RoBERTa doesn't use them.
@@ -79,15 +86,18 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
         ### self attention
         self_attn: BertSelfAttention = layer.attention.self
         assert(
-            roberta_layer.self_attn.in_proj_weight.shape == torch.Size((3 * config.hidden_size, config.hidden_size))
+            roberta_layer.self_attn.k_proj.weight.data.shape == \
+            roberta_layer.self_attn.q_proj.weight.data.shape == \
+            roberta_layer.self_attn.v_proj.weight.data.shape == \
+            torch.Size((config.hidden_size, config.hidden_size))
         )
-        # we use three distinct linear layers so we split the source layer here.
-        self_attn.query.weight.data = roberta_layer.self_attn.in_proj_weight[:config.hidden_size, :]
-        self_attn.query.bias.data = roberta_layer.self_attn.in_proj_bias[:config.hidden_size]
-        self_attn.key.weight.data = roberta_layer.self_attn.in_proj_weight[config.hidden_size:2*config.hidden_size, :]
-        self_attn.key.bias.data = roberta_layer.self_attn.in_proj_bias[config.hidden_size:2*config.hidden_size]
-        self_attn.value.weight.data = roberta_layer.self_attn.in_proj_weight[2*config.hidden_size:, :]
-        self_attn.value.bias.data = roberta_layer.self_attn.in_proj_bias[2*config.hidden_size:]
+
+        self_attn.query.weight.data = roberta_layer.self_attn.q_proj.weight
+        self_attn.query.bias.data = roberta_layer.self_attn.q_proj.bias
+        self_attn.key.weight.data = roberta_layer.self_attn.k_proj.weight
+        self_attn.key.bias.data = roberta_layer.self_attn.k_proj.bias
+        self_attn.value.weight.data = roberta_layer.self_attn.v_proj.weight
+        self_attn.value.bias.data = roberta_layer.self_attn.v_proj.bias
 
         ### self-attention output
         self_output: BertSelfOutput = layer.attention.output
@@ -151,6 +161,7 @@ def convert_roberta_checkpoint_to_pytorch(roberta_checkpoint_path, pytorch_dump_
     if not success:
         raise Exception("Something went wRoNg")
 
+    pathlib.Path(pytorch_dump_folder_path).mkdir(parents=True, exist_ok=True)
     print(f"Saving model to {pytorch_dump_folder_path}")
     model.save_pretrained(pytorch_dump_folder_path)
 

transformers/convert_t5_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,65 @@
+# 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

@@ -26,14 +26,6 @@ from contextlib import contextmanager
 
 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
-try:
-    import tensorflow as tf
-    assert hasattr(tf, '__version__') and int(tf.__version__[0]) >= 2
-    _tf_available = True  # pylint: disable=invalid-name
-    logger.info("TensorFlow version {} available.".format(tf.__version__))
-except (ImportError, AssertionError):
-    _tf_available = False  # pylint: disable=invalid-name
-
 try:
     import torch
     _torch_available = True  # pylint: disable=invalid-name
@@ -41,6 +33,13 @@ try:
 except ImportError:
     _torch_available = False  # pylint: disable=invalid-name
 
+try:
+    import tensorflow as tf
+    assert hasattr(tf, '__version__') and int(tf.__version__[0]) >= 2
+    _tf_available = True  # pylint: disable=invalid-name
+    logger.info("TensorFlow version {} available.".format(tf.__version__))
+except (ImportError, AssertionError):
+    _tf_available = False  # pylint: disable=invalid-name
 
 try:
     from torch.hub import _get_torch_home
@@ -72,9 +71,14 @@ WEIGHTS_NAME = "pytorch_model.bin"
 TF2_WEIGHTS_NAME = 'tf_model.h5'
 TF_WEIGHTS_NAME = 'model.ckpt'
 CONFIG_NAME = "config.json"
+MODEL_CARD_NAME = "model_card.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/hf_api.py

@@ -131,8 +131,9 @@ class HfApi:
         # the client still has to specify it when uploading the file.
         with open(filepath, "rb") as f:
             pf = TqdmProgressFileReader(f)
+            data = f if pf.total_size > 0 else ""
 
-            r = requests.put(urls.write, data=f, headers={
+            r = requests.put(urls.write, data=data, headers={
                 "content-type": urls.type,
             })
             r.raise_for_status()

transformers/model_card.py

@@ -0,0 +1,226 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Configuration base class and utilities."""
+
+from __future__ import (absolute_import, division, print_function,
+                        unicode_literals)
+
+import copy
+import json
+import logging
+import os
+from io import open
+
+from .configuration_auto import ALL_PRETRAINED_CONFIG_ARCHIVE_MAP
+
+from .file_utils import CONFIG_NAME, MODEL_CARD_NAME, cached_path, is_remote_url, hf_bucket_url
+
+
+logger = logging.getLogger(__name__)
+
+
+class ModelCard(object):
+    r""" Model Card class.
+        Store model card as well as methods for loading/downloading/saving model cards.
+
+        Please read the following paper for details and explanation on the sections:
+            "Model Cards for Model Reporting"
+                by Margaret Mitchell, Simone Wu,
+                Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer,
+                Inioluwa Deborah Raji and Timnit Gebru for the proposal behind model cards.
+            Link: https://arxiv.org/abs/1810.03993
+
+        Note:
+            A model card can be loaded and saved to disk.
+
+        Parameters:
+    """
+    def __init__(self, **kwargs):
+        # Recomended attributes from https://arxiv.org/abs/1810.03993 (see papers)
+        self.model_details = kwargs.pop('model_details', {})
+        self.intended_use = kwargs.pop('intended_use', {})
+        self.factors = kwargs.pop('factors', {})
+        self.metrics = kwargs.pop('metrics', {})
+        self.evaluation_data = kwargs.pop('evaluation_data', {})
+        self.training_data = kwargs.pop('training_data', {})
+        self.quantitative_analyses = kwargs.pop('quantitative_analyses', {})
+        self.ethical_considerations = kwargs.pop('ethical_considerations', {})
+        self.caveats_and_recommendations = kwargs.pop('caveats_and_recommendations', {})
+
+        # Open additional attributes
+        for key, value in kwargs.items():
+            try:
+                setattr(self, key, value)
+            except AttributeError as err:
+                logger.error("Can't set {} with value {} for {}".format(key, value, self))
+                raise err
+
+    def save_pretrained(self, save_directory_or_file):
+        """ Save a model card object to the directory or file `save_directory_or_file`.
+        """
+        if os.path.isdir(save_directory_or_file):
+            # If we save using the predefined names, we can load using `from_pretrained`
+            output_model_card_file = os.path.join(save_directory_or_file, MODEL_CARD_NAME)
+        else:
+            output_model_card_file = save_directory_or_file
+
+        self.to_json_file(output_model_card_file)
+        logger.info("Model card saved in {}".format(output_model_card_file))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        r""" Instantiate a :class:`~transformers.ModelCard` from a pre-trained model model card.
+
+        Parameters:
+            pretrained_model_name_or_path: either:
+
+                - a string with the `shortcut name` of a pre-trained model card to load from cache or download, e.g.: ``bert-base-uncased``.
+                - a string with the `identifier name` of a pre-trained model card that was user-uploaded to our S3, e.g.: ``dbmdz/bert-base-german-cased``.
+                - a path to a `directory` containing a mode card file saved using the :func:`~transformers.ModelCard.save_pretrained` method, e.g.: ``./my_model_directory/``.
+                - a path or url to a saved model card JSON `file`, e.g.: ``./my_model_directory/model_card.json``.
+
+            cache_dir: (`optional`) string:
+                Path to a directory in which a downloaded pre-trained model
+                card should be cached if the standard cache should not be used.
+
+            kwargs: (`optional`) dict: key/value pairs with which to update the ModelCard object after loading.
+
+                - The values in kwargs of any keys which are model card attributes will be used to override the loaded values.
+                - Behavior concerning key/value pairs whose keys are *not* model card attributes is controlled by the `return_unused_kwargs` keyword parameter.
+
+            force_download: (`optional`) boolean, default False:
+                Force to (re-)download the model card file and override the cached version if it exists.
+
+            resume_download: (`optional`) boolean, default False:
+                Do not delete incompletely recieved file. Attempt to resume the download if such a file exists.
+
+            proxies: (`optional`) dict, default None:
+                A dictionary of proxy servers to use by protocol or endpoint, e.g.: {'http': 'foo.bar:3128', 'http://hostname': 'foo.bar:4012'}.
+                The proxies are used on each request.
+
+            return_unused_kwargs: (`optional`) bool:
+
+                - If False, then this function returns just the final model card object.
+                - If True, then this functions returns a tuple `(model card, unused_kwargs)` where `unused_kwargs` is a dictionary consisting of the key/value pairs whose keys are not model card attributes: ie the part of kwargs which has not been used to update `ModelCard` and is otherwise ignored.
+
+        Examples::
+
+            model_card = ModelCard.from_pretrained('bert-base-uncased')    # Download model card from S3 and cache.
+            model_card = ModelCard.from_pretrained('./test/saved_model/')  # E.g. model card was saved using `save_pretrained('./test/saved_model/')`
+            model_card = ModelCard.from_pretrained('./test/saved_model/model_card.json')
+            model_card = ModelCard.from_pretrained('bert-base-uncased', output_attention=True, foo=False)
+
+        """
+        cache_dir = kwargs.pop('cache_dir', None)
+        force_download = kwargs.pop('force_download', False)
+        resume_download = kwargs.pop('resume_download', False)
+        proxies = kwargs.pop('proxies', None)
+        return_unused_kwargs = kwargs.pop('return_unused_kwargs', False)
+
+        if pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
+            # For simplicity we use the same pretrained url than the configuration files but with a different suffix (model_card.json)
+            model_card_file = ALL_PRETRAINED_CONFIG_ARCHIVE_MAP[pretrained_model_name_or_path]
+            model_card_file = model_card_file.replace(CONFIG_NAME, MODEL_CARD_NAME)
+        elif os.path.isdir(pretrained_model_name_or_path):
+            model_card_file = os.path.join(pretrained_model_name_or_path, MODEL_CARD_NAME)
+        elif os.path.isfile(pretrained_model_name_or_path) or is_remote_url(pretrained_model_name_or_path):
+            model_card_file = pretrained_model_name_or_path
+        else:
+            model_card_file = hf_bucket_url(pretrained_model_name_or_path, postfix=MODEL_CARD_NAME)
+
+        try:
+            # Load from URL or cache if already cached
+            resolved_model_card_file = cached_path(model_card_file, cache_dir=cache_dir, force_download=force_download,
+                                               proxies=proxies, resume_download=resume_download)
+            if resolved_model_card_file == model_card_file:
+                logger.info("loading model card file {}".format(model_card_file))
+            else:
+                logger.info("loading model card file {} from cache at {}".format(
+                    model_card_file, resolved_model_card_file))
+            # Load model card
+            model_card = cls.from_json_file(resolved_model_card_file)
+
+        except EnvironmentError:
+            if pretrained_model_name_or_path in ALL_PRETRAINED_CONFIG_ARCHIVE_MAP:
+                logger.warning("Couldn't reach server at '{}' to download model card file.".format(
+                        model_card_file))
+            else:
+                logger.warning("Model name '{}' was not found in model name list ({}). " \
+                      "We assumed '{}' was a path or url to a model card file named {} or " \
+                      "a directory containing such a file but couldn't find any such file at this path or url.".format(
+                        pretrained_model_name_or_path,
+                        ', '.join(ALL_PRETRAINED_CONFIG_ARCHIVE_MAP.keys()),
+                        model_card_file, MODEL_CARD_NAME))
+            logger.warning("Creating an empty model card.")
+
+            # We fall back on creating an empty model card
+            model_card = cls()
+
+        except json.JSONDecodeError:
+            logger.warning("Couldn't reach server at '{}' to download model card file or "
+                           "model card file is not a valid JSON file. "
+                           "Please check network or file content here: {}.".format(model_card_file, resolved_model_card_file))
+            logger.warning("Creating an empty model card.")
+
+            # We fall back on creating an empty model card
+            model_card = cls()
+
+        # Update model card with kwargs if needed
+        to_remove = []
+        for key, value in kwargs.items():
+            if hasattr(model_card, key):
+                setattr(model_card, key, value)
+                to_remove.append(key)
+        for key in to_remove:
+            kwargs.pop(key, None)
+
+        logger.info("Model card: %s", str(model_card))
+        if return_unused_kwargs:
+            return model_card, kwargs
+        else:
+            return model_card
+
+    @classmethod
+    def from_dict(cls, json_object):
+        """Constructs a `ModelCard` from a Python dictionary of parameters."""
+        return cls(**json_object)
+
+    @classmethod
+    def from_json_file(cls, json_file):
+        """Constructs a `ModelCard` from a json file of parameters."""
+        with open(json_file, "r", encoding='utf-8') as reader:
+            text = reader.read()
+        dict_obj = json.loads(text)
+        return cls(**dict_obj)
+
+    def __eq__(self, other):
+        return self.__dict__ == other.__dict__
+
+    def __repr__(self):
+        return str(self.to_json_string())
+
+    def to_dict(self):
+        """Serializes this instance to a Python dictionary."""
+        output = copy.deepcopy(self.__dict__)
+        return output
+
+    def to_json_string(self):
+        """Serializes this instance to a JSON string."""
+        return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n"
+
+    def to_json_file(self, json_file_path):
+        """ Save this instance to a json file."""
+        with open(json_file_path, "w", encoding='utf-8') as writer:
+            writer.write(self.to_json_string())

transformers/modeling_auto.py

@@ -18,17 +18,18 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 
-from .modeling_bert import BertModel, BertForMaskedLM, BertForSequenceClassification, BertForQuestionAnswering
-from .modeling_openai import OpenAIGPTModel, OpenAIGPTLMHeadModel
-from .modeling_gpt2 import GPT2Model, GPT2LMHeadModel
-from .modeling_ctrl import CTRLModel, CTRLLMHeadModel
-from .modeling_transfo_xl import TransfoXLModel, TransfoXLLMHeadModel
-from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering
-from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering
-from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification
-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_bert import BertModel, BertForMaskedLM, BertForSequenceClassification, BertForQuestionAnswering, BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_openai import OpenAIGPTModel, OpenAIGPTLMHeadModel, OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_gpt2 import GPT2Model, GPT2LMHeadModel, GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_ctrl import CTRLModel, CTRLLMHeadModel, CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_transfo_xl import TransfoXLModel, TransfoXLLMHeadModel, TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_xlnet import XLNetModel, XLNetLMHeadModel, XLNetForSequenceClassification, XLNetForQuestionAnswering, XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_xlm import XLMModel, XLMWithLMHeadModel, XLMForSequenceClassification, XLMForQuestionAnswering, XLM_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_roberta import RobertaModel, RobertaForMaskedLM, RobertaForSequenceClassification, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_distilbert import DistilBertModel, DistilBertForQuestionAnswering, DistilBertForMaskedLM, DistilBertForSequenceClassification, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_camembert import CamembertModel, CamembertForMaskedLM, CamembertForSequenceClassification, CamembertForMultipleChoice, CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_albert import AlbertModel, AlbertForMaskedLM, AlbertForSequenceClassification, AlbertForQuestionAnswering, ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_t5 import T5Model, T5WithLMHeadModel, T5_PRETRAINED_MODEL_ARCHIVE_MAP
 
 from .modeling_utils import PreTrainedModel, SequenceSummary
 
@@ -37,6 +38,24 @@ from .file_utils import add_start_docstrings
 logger = logging.getLogger(__name__)
 
 
+ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
+    for pretrained_map in [
+        BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ]
+    for key, value, in pretrained_map.items())
+
+
 class AutoModel(object):
     r"""
         :class:`~transformers.AutoModel` is a generic model class
@@ -49,6 +68,7 @@ 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 +94,7 @@ 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 +167,9 @@ 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 +208,7 @@ 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 +237,7 @@ 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 +309,9 @@ 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_bert.py

@@ -51,7 +51,9 @@ BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-pytorch_model.bin",
     'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-pytorch_model.bin",
     'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-pytorch_model.bin",
-    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-pytorch_model.bin"
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-pytorch_model.bin",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/pytorch_model.bin",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/pytorch_model.bin",
 }
 
 

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_gpt2.py

@@ -634,6 +634,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
     """
     def __init__(self, config):
         super(GPT2DoubleHeadsModel, self).__init__(config)
+        config.num_labels = 1
         self.transformer = GPT2Model(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)

transformers/modeling_openai.py

@@ -590,6 +590,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
     def __init__(self, config):
         super(OpenAIGPTDoubleHeadsModel, self).__init__(config)
 
+        config.num_labels = 1
         self.transformer = OpenAIGPTModel(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)

transformers/modeling_t5.py

@@ -0,0 +1,886 @@
+# coding=utf-8
+# Copyright 2018 Mesh TensorFlow authors, T5 Authors and HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch T5 model. """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import json
+import logging
+import math
+import os
+import sys
+import copy
+import itertools
+from io import open
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from .modeling_utils import PreTrainedModel, prune_linear_layer
+from .configuration_t5 import T5Config
+from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
+
+logger = logging.getLogger(__name__)
+
+####################################################
+# This dict contrains shortcut names and associated url
+# for the pretrained weights provided with the models
+####################################################
+T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-pytorch_model.bin",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-pytorch_model.bin",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-pytorch_model.bin",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-pytorch_model.bin",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-pytorch_model.bin",
+}
+
+####################################################
+# This is a conversion method from TF 1.0 to PyTorch
+# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
+####################################################
+def load_tf_weights_in_t5(model, config, tf_checkpoint_path):
+    """ Load tf checkpoints in a pytorch model.
+    """
+    try:
+        import re
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error("Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions.")
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    tf_weights = {}
+    for name, shape in init_vars:
+        logger.info("Loading TF weight {} with shape {}".format(name, shape))
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        tf_weights[name] = array
+
+    for txt_name in names:
+        name = txt_name.split('/')
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(n in ["adam_v", "adam_m", "global_step"] for n in name):
+            logger.info("Skipping {}".format("/".join(name)))
+            tf_weights.pop(txt_name, None)
+            continue
+        if '_slot_' in name[-1]:
+            logger.info("Skipping {}".format("/".join(name)))
+            tf_weights.pop(txt_name, None)
+            continue
+        pointer = model
+        array = tf_weights[txt_name]
+        for m_name in name:
+            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
+                l = re.split(r'_(\d+)', m_name)
+            else:
+                l = [m_name]
+            if l[0] in ['kernel', 'scale', 'embedding']:
+                pointer = getattr(pointer, 'weight')
+            # elif l[0] == 'scale':
+            #     pointer = getattr(pointer, 'weight')
+            # elif l[0] == 'output_bias' or l[0] == 'beta':
+            #     pointer = getattr(pointer, 'bias')
+            # elif l[0] == 'squad':
+            #     pointer = getattr(pointer, 'classifier')
+            else:
+                try:
+                    pointer = getattr(pointer, l[0])
+                except AttributeError:
+                    logger.info("Skipping {}".format("/".join(name)))
+                    continue
+            if len(l) >= 2:
+                num = int(l[1])
+                pointer = pointer[num]
+        if l[0] not in ['kernel', 'scale', 'embedding']:
+            pointer = getattr(pointer, 'weight')
+        if l[0] != 'embedding':
+            logger.info("Transposing numpy weight of shape {} for {}".format(array.shape, name))
+            array = np.transpose(array)
+        try:
+            assert pointer.shape == array.shape
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info("Initialize PyTorch weight {}".format(name))
+        pointer.data = torch.from_numpy(array.astype(np.float32))
+        tf_weights.pop(txt_name, None)
+
+    logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
+    # logger.info("Weights not copied to PyTorch model: {}".format(', '.join(tf_weights.keys())))
+    return model
+
+
+####################################################
+# PyTorch Models are constructed by sub-classing
+# - torch.nn.Module for the layers and
+# - PreTrainedModel for the models (it-self a sub-class of torch.nn.Module)
+####################################################
+
+class T5LayerNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """ Construct a layernorm module in the T5 style
+            No bias and no substraction of mean.
+        """
+        super(T5LayerNorm, self).__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, x):
+        variance = x.pow(2).mean(-1, keepdim=True)
+        x = x / torch.sqrt(variance + self.variance_epsilon)
+        return self.weight * x
+
+
+class T5DenseReluDense(nn.Module):
+    def __init__(self, config):
+        super(T5DenseReluDense, self).__init__()
+        self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states):
+        h = self.wi(hidden_states)
+        h = F.relu(h)
+        h = self.dropout(h)
+        h = self.wo(h)
+        return h
+
+
+class T5LayerFF(nn.Module):
+    def __init__(self, config):
+        super(T5LayerFF, self).__init__()
+        self.DenseReluDense = T5DenseReluDense(config)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states):
+        norm_x = self.layer_norm(hidden_states)
+        y = self.DenseReluDense(norm_x)
+        layer_output = hidden_states + self.dropout(y)
+        return layer_output
+
+
+class T5Attention(nn.Module):
+    NEW_ID = itertools.count()
+
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5Attention, self).__init__()
+        self.layer_id = next(T5Attention.NEW_ID)
+        self.is_decoder = config.is_decoder
+        self.has_relative_attention_bias = has_relative_attention_bias
+
+        self.output_attentions = config.output_attentions
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.d_model = config.d_model
+        self.d_kv = config.d_kv
+        self.n_heads = config.num_heads
+        self.dropout = config.dropout_rate
+        self.inner_dim = self.n_heads * self.d_kv
+
+        # Mesh TensorFlow initialization to avoid scaling before softmax
+        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
+
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.n_heads, self.d_kv)
+        heads = set(heads) - self.pruned_heads
+        for head in heads:
+            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+        # Prune linear layers
+        self.q = prune_linear_layer(self.q, index)
+        self.k = prune_linear_layer(self.k, index)
+        self.v = prune_linear_layer(self.v, index)
+        self.o = prune_linear_layer(self.o, index, dim=1)
+        # Update hyper params
+        self.n_heads = self.n_heads - len(heads)
+        self.inner_dim = self.d_kv * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    @staticmethod
+    def _relative_position_bucket(relative_position,
+                                  bidirectional=True,
+                                  num_buckets=32,
+                                  max_distance=128):
+        """
+        Adapted from Mesh Tensorflow:
+        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+        Translate relative position to a bucket number for relative attention.
+        The relative position is defined as memory_position - query_position, i.e.
+        the distance in tokens from the attending position to the attended-to
+        position.  If bidirectional=False, then positive relative positions are
+        invalid.
+        We use smaller buckets for small absolute relative_position and larger buckets
+        for larger absolute relative_positions.  All relative positions >=max_distance
+        map to the same bucket.  All relative positions <=-max_distance map to the
+        same bucket.  This should allow for more graceful generalization to longer
+        sequences than the model has been trained on.
+        Args:
+            relative_position: an int32 Tensor
+            bidirectional: a boolean - whether the attention is bidirectional
+            num_buckets: an integer
+            max_distance: an integer
+        Returns:
+            a Tensor with the same shape as relative_position, containing int32
+            values in the range [0, num_buckets)
+        """
+        ret = 0
+        n = -relative_position
+        if bidirectional:
+            num_buckets //= 2
+            ret += (n < 0).to(torch.long) * num_buckets  # mtf.to_int32(mtf.less(n, 0)) * num_buckets
+            n = torch.abs(n)
+        else:
+            n = torch.max(n, torch.zeros_like(n))
+        # now n is in the range [0, inf)
+
+        # half of the buckets are for exact increments in positions
+        max_exact = num_buckets // 2
+        is_small = (n < max_exact)
+
+        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
+        val_if_large = max_exact + (
+            torch.log(n.float() / max_exact)
+            / math.log(max_distance / max_exact) * (num_buckets - max_exact)).to(torch.long)
+        val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))
+
+        ret += torch.where(is_small, n, val_if_large)
+        return ret
+
+    def compute_bias(self, qlen, klen):
+        """ Compute binned relative position bias """
+        context_position = torch.arange(qlen, dtype=torch.long)[:, None]
+        memory_position = torch.arange(klen, dtype=torch.long)[None, :]
+        relative_position = memory_position - context_position  # shape (qlen, klen)
+        rp_bucket = self._relative_position_bucket(relative_position,  # shape (qlen, klen)
+                                                   bidirectional=not self.is_decoder,
+                                                   num_buckets=self.relative_attention_num_buckets)
+        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
+        values = values.permute([2, 0, 1]).unsqueeze(0) # shape (1, num_heads, qlen, klen)
+        return values
+
+    def forward(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None):
+        """
+        Self-attention (if kv is None) or attention over source sentence (provided by kv).
+        """
+        # Input is (bs, qlen, dim)
+        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
+        bs, qlen, dim = input.size()
+        if kv is None:
+            klen = qlen if cache is None else cache['slen'] + qlen
+        else:
+            klen = kv.size(1)
+
+        def shape(x):
+            """  projection """
+            return x.view(bs, -1, self.n_heads, self.d_kv).transpose(1, 2)
+
+        def unshape(x):
+            """  compute context """
+            return x.transpose(1, 2).contiguous().view(bs, -1, self.inner_dim)
+
+        q = shape(self.q(input))                                          # (bs, n_heads, qlen, dim_per_head)
+        if kv is None:
+            k = shape(self.k(input))                                      # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(input))                                      # (bs, n_heads, qlen, dim_per_head)
+        elif cache is None or self.layer_id not in cache:
+            k = v = kv
+            k = shape(self.k(k))                                          # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(v))                                          # (bs, n_heads, qlen, dim_per_head)
+
+        if cache is not None:
+            if self.layer_id in cache:
+                if kv is None:
+                    k_, v_ = cache[self.layer_id]
+                    k = torch.cat([k_, k], dim=2)                             # (bs, n_heads, klen, dim_per_head)
+                    v = torch.cat([v_, v], dim=2)                             # (bs, n_heads, klen, dim_per_head)
+                else:
+                    k, v = cache[self.layer_id]
+            cache[self.layer_id] = (k, v)
+
+        # q = q / math.sqrt(dim_per_head)                                     # No scaling in T5
+        scores = torch.einsum('bnqd,bnkd->bnqk', q, k)                        # (bs, n_heads, qlen, klen)
+
+        if position_bias is None:
+            if not self.has_relative_attention_bias:
+                raise ValueError("No position_bias provided and no weights to compute position_bias")
+            position_bias = self.compute_bias(qlen, klen)
+            if mask is not None:
+                position_bias = position_bias + mask                          # (bs, n_heads, qlen, klen)
+
+        scores += position_bias
+        weights = F.softmax(scores.float(), dim=-1).type_as(scores)           # (bs, n_heads, qlen, klen)
+        weights = F.dropout(weights, p=self.dropout, training=self.training)  # (bs, n_heads, qlen, klen)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = torch.matmul(weights, v)                                    # (bs, n_heads, qlen, dim_per_head)
+        context = unshape(context)                                            # (bs, qlen, dim)
+
+        context = self.o(context)
+
+        outputs = (context,)
+        if self.output_attentions:
+            outputs = outputs + (weights,)
+        if self.has_relative_attention_bias:
+            outputs = outputs + (position_bias,)
+        return outputs
+
+
+class T5LayerSelfAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5LayerSelfAttention, self).__init__()
+        self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states, attention_mask=None, position_bias=None, head_mask=None):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.SelfAttention(norm_x,
+                                              mask=attention_mask,
+                                              position_bias=position_bias,
+                                              head_mask=head_mask)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5LayerCrossAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5LayerCrossAttention, self).__init__()
+        self.EncDecAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(self, hidden_states, kv, attention_mask=None, position_bias=None, head_mask=None):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.EncDecAttention(norm_x,
+                                                mask=attention_mask,
+                                                kv=kv,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5Block(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super(T5Block, self).__init__()
+        self.is_decoder = config.is_decoder
+        self.layer = nn.ModuleList()
+        self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+        if self.is_decoder:
+            self.layer.append(T5LayerCrossAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+            self.layer.append(T5LayerFF(config))
+        else:
+            self.layer.append(T5LayerFF(config))
+
+    def forward(self, hidden_states, attention_mask=None, position_bias=None,
+                encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
+                head_mask=None):
+        self_attention_outputs = self.layer[0](hidden_states,
+                                                attention_mask=attention_mask,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask)
+        hidden_states = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # Keep self-attention outputs and relative position weights
+
+        if not self.is_decoder:
+            hidden_states = self.layer[1](hidden_states)
+        else:
+            cross_attention_outputs = self.layer[1](hidden_states,
+                                                    kv=encoder_hidden_states,
+                                                    attention_mask=encoder_attention_mask,
+                                                    position_bias=encoder_decoder_position_bias,
+                                                    head_mask=head_mask)
+            hidden_states = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]  # Keep cross-attention outputs and relative position weights
+            hidden_states = self.layer[2](hidden_states)
+
+        outputs = (hidden_states,) + outputs  # add attentions if we output them
+        return outputs  # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+
+
+class T5PreTrainedModel(PreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = T5Config
+    pretrained_model_archive_map = T5_PRETRAINED_MODEL_ARCHIVE_MAP
+    load_tf_weights = load_tf_weights_in_t5
+    base_model_prefix = "transformer"
+
+    @property
+    def dummy_inputs(self):
+        input_ids = torch.tensor(DUMMY_INPUTS)
+        input_mask = torch.tensor(DUMMY_MASK)
+        dummy_inputs = {'decoder_input_ids': input_ids,
+                        'encoder_input_ids': input_ids,
+                        'decoder_attention_mask': input_mask}
+        return dummy_inputs
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        factor = self.config.initializer_factor  # Used for testing weights initialization
+        if isinstance(module, T5LayerNorm):
+            module.weight.data.fill_(factor*1.0)
+        elif isinstance(module, (T5Model, T5WithLMHeadModel)):
+            # Mesh TensorFlow embeddings initialization
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
+            module.shared.weight.data.normal_(mean=0.0, std=factor*1.0)
+        elif isinstance(module, T5DenseReluDense):
+            # Mesh TensorFlow FF initialization
+            # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
+            # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
+            module.wi.weight.data.normal_(mean=0.0, std=factor*((self.config.d_model) ** -0.5))
+            if hasattr(module.wi, 'bias') and module.wi.bias is not None:
+                module.wi.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor*((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, 'bias') and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5Attention):
+            # Mesh TensorFlow attention initialization to avoid scaling before softmax
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
+            d_model = self.config.d_model
+            d_kv = self.config.d_kv
+            n_heads = self.config.num_heads
+            module.q.weight.data.normal_(mean=0.0, std=factor*((d_model * d_kv) ** -0.5))
+            module.k.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
+            module.v.weight.data.normal_(mean=0.0, std=factor*(d_model ** -0.5))
+            module.o.weight.data.normal_(mean=0.0, std=factor*((n_heads * d_kv) ** -0.5))
+            if module.has_relative_attention_bias:
+                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor*((d_model) ** -0.5))
+
+
+class T5Stack(T5PreTrainedModel):
+    def __init__(self, config):
+        super(T5Stack, self).__init__(config)
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.is_decoder = config.is_decoder
+
+        self.block = nn.ModuleList([T5Block(config, has_relative_attention_bias=bool(i == 0))
+                                    for i in range(config.num_layers)])
+        self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        self.init_weights()
+
+    def forward(self,
+                hidden_states,
+                attention_mask=None,
+                encoder_hidden_states=None,
+                encoder_attention_mask=None,
+                head_mask=None):
+
+        batch_size, seq_length = hidden_states.shape[0], hidden_states.shape[1]
+        if attention_mask is None:
+            attention_mask = torch.ones(batch_size, seq_length).to(hidden_states.device)
+        if self.is_decoder and encoder_attention_mask is None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = torch.ones(batch_size, encoder_seq_length).to(hidden_states.device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+        # Provided a padding mask of dimensions [batch_size, seq_length]
+        # - if the model is a decoder, apply a causal mask in addition to the padding mask
+        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                seq_ids = torch.arange(seq_length, device=hidden_states.device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = causal_mask.to(attention_mask)
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -1e9 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+        # extended_attention_mask = (extended_attention_mask == extended_attention_mask.transpose(-1, -2))
+
+        extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
+
+        if self.is_decoder:
+            # If a 2D ou 3D attention mask is provided for the cross-attention
+            # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+            if encoder_attention_mask.dim() == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            if encoder_attention_mask.dim() == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+            # T5 has a mask that can compare sequence ids, we can simulate this here with this transposition
+            # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+            # encoder_extended_attention_mask = (encoder_extended_attention_mask == encoder_extended_attention_mask.transpose(-1, -2))
+
+            encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if head_mask is not None:
+            if head_mask.dim() == 1:
+                head_mask = head_mask.unsqueeze(0).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
+                head_mask = head_mask.expand(self.config.num_layers, -1, -1, -1, -1)
+            elif head_mask.dim() == 2:
+                head_mask = head_mask.unsqueeze(1).unsqueeze(-1).unsqueeze(-1)  # We can specify head_mask for each layer
+            head_mask = head_mask.to(dtype=next(self.parameters()).dtype) # switch to fload if need + fp16 compatibility
+        else:
+            head_mask = [None] * self.config.num_layers
+
+        all_hidden_states = ()
+        all_attentions = ()
+        position_bias = None
+        encoder_decoder_position_bias = None
+
+        hidden_states = self.dropout(hidden_states)
+        for i, layer_module in enumerate(self.block):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states,
+                                         attention_mask=extended_attention_mask,
+                                         position_bias=position_bias,
+                                         encoder_hidden_states=encoder_hidden_states,
+                                         encoder_attention_mask=encoder_extended_attention_mask,
+                                         encoder_decoder_position_bias=encoder_decoder_position_bias,
+                                         head_mask=head_mask[i])
+            # layer_outputs is a tuple with:
+            # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+            hidden_states = layer_outputs[0]
+            if i == 0:
+                # We share the position biases between the layers - the first layer store them
+                # layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+                position_bias = layer_outputs[2 if self.output_attentions else 1]
+                if self.is_decoder:
+                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)  # We keep only self-attention weights for now
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        layer_output = self.dropout(hidden_states)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+T5_START_DOCSTRING = r"""    The T5 model was proposed in
+    `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
+    by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
+    It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
+
+    This model is a PyTorch `torch.nn.Module`_ sub-class. Use it as a regular PyTorch Module and
+    refer to the PyTorch documentation for all matter related to general usage and behavior.
+
+    .. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
+        https://arxiv.org/abs/1910.10683
+
+    .. _`torch.nn.Module`:
+        https://pytorch.org/docs/stable/nn.html#module
+
+    Parameters:
+        config (:class:`~transformers.T5Config`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+T5_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+            T5 is a model with relative position embeddings so you should be able to pad the inputs on
+            the right or the left.
+
+            Indices can be obtained using :class:`transformers.T5Tokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **head_mask**: (`optional`) ``torch.FloatTensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare T5 Model transformer outputting raw hidden-states"
+                      "without any specific head on top.",
+                      T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class T5Model(T5PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = T5Model.from_pretrained('t5-small')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config):
+        super(T5Model, self).__init__(config)
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = T5Stack(encoder_config)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = T5Stack(decoder_config)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def forward(self, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        encoder_attention_mask = kwargs_encoder.get("attention_mask", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            if encoder_attention_mask is not None:
+                # Apply masking
+                encoder_attention_mask = (encoder_attention_mask != 0).to(hidden_states)
+                hidden_states = hidden_states * encoder_attention_mask.unsqueeze(-1)
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = encoder_attention_mask
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
+    T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class T5WithLMHeadModel(T5PreTrainedModel):
+    r"""
+        **lm_labels**: (`optional`) ``torch.LongTensor`` of shape ``(batch_size, sequence_length)``:
+            Labels for computing the masked language modeling loss.
+            Indices should be in ``[-1, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring)
+            Tokens with indices set to ``-1`` are ignored (masked), the loss is only computed for the tokens with labels
+            in ``[0, ..., config.vocab_size]``
+
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **loss**: (`optional`, returned when ``lm_labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Masked language modeling loss.
+        **prediction_scores**: ``torch.FloatTensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``torch.FloatTensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``torch.FloatTensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = T5WithLMHeadModel.from_pretrained('t5-small')
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids=input_ids, lm_labels=input_ids)
+        loss, prediction_scores = outputs[:2]
+
+    """
+    def __init__(self, config):
+        super(T5WithLMHeadModel, self).__init__(config)
+        self.model_dim = config.d_model
+
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = T5Stack(encoder_config)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = T5Stack(decoder_config)
+
+        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def forward(self, **kwargs):
+        # keyword arguments come in 3 flavors: encoder-specific (prefixed by
+        # `encoder_`), decoder-specific (prefixed by `decoder_`) and those
+        # that apply to the model as whole.
+        # We let the specific kwargs override the common ones in case of conflict.
+
+        lm_labels = kwargs.pop('decoder_lm_labels', None)
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        sequence_output = decoder_outputs[0]
+        # Rescale output before projecting on vocab
+        # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+        sequence_output = sequence_output * (self.model_dim ** -0.5)
+        lm_logits = self.lm_head(sequence_output)
+
+        decoder_outputs = (lm_logits,) + decoder_outputs[1:]  # Add hidden states and attention if they are here
+        if lm_labels is not None:
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = lm_labels[..., 1:].contiguous()
+            loss_fct = CrossEntropyLoss(ignore_index=-1)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)),
+                            shift_labels.view(-1))
+            decoder_outputs = (loss,) + decoder_outputs  # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
+
+        return decoder_outputs + encoder_outputs

transformers/modeling_tf_auto.py

@@ -18,21 +18,40 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import logging
 
-from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, TFBertForQuestionAnswering
-from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel
-from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel
-from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel
-from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, TFXLNetForQuestionAnsweringSimple
-from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, TFXLMForQuestionAnsweringSimple
-from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification
-from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification
-from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel
+from .modeling_tf_bert import TFBertModel, TFBertForMaskedLM, TFBertForSequenceClassification, TFBertForQuestionAnswering, TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_openai import TFOpenAIGPTModel, TFOpenAIGPTLMHeadModel, TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_gpt2 import TFGPT2Model, TFGPT2LMHeadModel, TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_transfo_xl import TFTransfoXLModel, TFTransfoXLLMHeadModel, TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_xlnet import TFXLNetModel, TFXLNetLMHeadModel, TFXLNetForSequenceClassification, TFXLNetForQuestionAnsweringSimple, TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_xlm import TFXLMModel, TFXLMWithLMHeadModel, TFXLMForSequenceClassification, TFXLMForQuestionAnsweringSimple, TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_roberta import TFRobertaModel, TFRobertaForMaskedLM, TFRobertaForSequenceClassification, TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_distilbert import TFDistilBertModel, TFDistilBertForQuestionAnswering, TFDistilBertForMaskedLM, TFDistilBertForSequenceClassification, TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_ctrl import TFCTRLModel, TFCTRLLMHeadModel, TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_albert import TFAlbertModel, TFAlbertForMaskedLM, TFAlbertForSequenceClassification, TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP
+from .modeling_tf_t5 import TFT5Model, TFT5WithLMHeadModel, TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
 
 from .file_utils import add_start_docstrings
 
 logger = logging.getLogger(__name__)
 
 
+TF_ALL_PRETRAINED_MODEL_ARCHIVE_MAP = dict((key, value)
+    for pretrained_map in [
+        TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_GPT2_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_CTRL_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLNET_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_XLM_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_ALBERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+        TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP,
+        ]
+    for key, value, in pretrained_map.items())
+
+
 class TFAutoModel(object):
     r"""
         :class:`~transformers.TFAutoModel` is a generic model class
@@ -45,6 +64,7 @@ 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 +88,7 @@ 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)
@@ -137,8 +158,12 @@ class TFAutoModel(object):
             model = TFAutoModel.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return TFT5Model.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -173,6 +198,7 @@ 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 +225,7 @@ 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)
@@ -269,8 +296,12 @@ class TFAutoModelWithLMHead(object):
             model = TFAutoModelWithLMHead.from_pretrained('./pt_model/bert_pytorch_model.bin', from_pt=True, config=config)
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return TFT5WithLMHeadModel.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForMaskedLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:
@@ -398,6 +429,8 @@ class TFAutoModelForSequenceClassification(object):
         """
         if 'distilbert' in pretrained_model_name_or_path:
             return TFDistilBertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
+        elif 'albert' in pretrained_model_name_or_path:
+            return TFAlbertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'roberta' in pretrained_model_name_or_path:
             return TFRobertaForSequenceClassification.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
         elif 'bert' in pretrained_model_name_or_path:

transformers/modeling_tf_bert.py

@@ -51,7 +51,9 @@ TF_BERT_PRETRAINED_MODEL_ARCHIVE_MAP = {
     'bert-base-japanese': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-tf_model.h5",
     'bert-base-japanese-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-whole-word-masking-tf_model.h5",
     'bert-base-japanese-char': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-tf_model.h5",
-    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-tf_model.h5"
+    'bert-base-japanese-char-whole-word-masking': "https://s3.amazonaws.com/models.huggingface.co/bert/cl-tohoku/bert-base-japanese-char-whole-word-masking-tf_model.h5",
+    'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/tf_model.h5",
+    'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/tf_model.h5",
 }
 
 

transformers/modeling_tf_gpt2.py

@@ -574,6 +574,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
     """
     def __init__(self, config, *inputs, **kwargs):
         super(TFGPT2DoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
+        config.num_labels = 1
         self.transformer = TFGPT2MainLayer(config, name='transformer')
         self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')
 

transformers/modeling_tf_openai.py

@@ -538,6 +538,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
     """
     def __init__(self, config, *inputs, **kwargs):
         super(TFOpenAIGPTDoubleHeadsModel, self).__init__(config, *inputs, **kwargs)
+        config.num_labels = 1
         self.transformer = TFOpenAIGPTMainLayer(config, name='transformer')
         self.multiple_choice_head = TFSequenceSummary(config, initializer_range=config.initializer_range, name='multiple_choice_head')
 

transformers/modeling_tf_pytorch_utils.py

@@ -78,6 +78,7 @@ def load_pytorch_checkpoint_in_tf2_model(tf_model, pytorch_checkpoint_path, tf_i
     logger.info("Loading PyTorch weights from {}".format(pt_path))
 
     pt_state_dict = torch.load(pt_path, map_location='cpu')
+    logger.info("PyTorch checkpoint contains {:,} parameters".format(sum(t.numel() for t in pt_state_dict.values())))
 
     return load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=tf_inputs, allow_missing_keys=allow_missing_keys)
 
@@ -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:
@@ -142,7 +143,11 @@ def load_pytorch_weights_in_tf2_model(tf_model, pt_state_dict, tf_inputs=None, a
         name, transpose = convert_tf_weight_name_to_pt_weight_name(sw_name, start_prefix_to_remove=start_prefix_to_remove)
 
         # Find associated numpy array in pytorch model state dict
-        assert name in pt_state_dict, "{} not found in PyTorch model".format(name)
+        if name not in pt_state_dict:
+            if allow_missing_keys:
+                continue
+            raise AttributeError("{} not found in PyTorch model".format(name))
+
         array = pt_state_dict[name].numpy()
 
         if transpose:
@@ -159,7 +164,8 @@ 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 +175,8 @@ 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
@@ -246,6 +254,7 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
 
     all_tf_weights = set(list(tf_weights_map.keys()))
     loaded_pt_weights_data_ptr = {}
+    missing_keys_pt = []
     for pt_weight_name, pt_weight in current_pt_params_dict.items():
         # Handle PyTorch shared weight ()not duplicated in TF 2.0
         if pt_weight.data_ptr() in loaded_pt_weights_data_ptr:
@@ -254,7 +263,10 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
 
         # Find associated numpy array in pytorch model state dict
         if pt_weight_name not in tf_weights_map:
-            raise ValueError("{} not found in TF 2.0 model".format(pt_weight_name))
+            if allow_missing_keys:
+                missing_keys_pt.append(pt_weight_name)
+                continue
+            raise AttributeError("{} not found in TF 2.0 model".format(pt_weight_name))
 
         array, transpose = tf_weights_map[pt_weight_name]
 
@@ -272,13 +284,14 @@ def load_tf2_weights_in_pytorch_model(pt_model, tf_weights, allow_missing_keys=F
             e.args += (pt_weight.shape, array.shape)
             raise e
 
-        logger.info("Initialize PyTorch weight {}".format(pt_weight_name))
+        # logger.warning("Initialize PyTorch weight {}".format(pt_weight_name))
 
         new_pt_params_dict[pt_weight_name] = torch.from_numpy(array)
         loaded_pt_weights_data_ptr[pt_weight.data_ptr()] = torch.from_numpy(array)
         all_tf_weights.discard(pt_weight_name)
 
     missing_keys, unexpected_keys = pt_model.load_state_dict(new_pt_params_dict, strict=False)
+    missing_keys += missing_keys_pt
 
     if len(missing_keys) > 0:
         logger.info("Weights of {} not initialized from TF 2.0 model: {}".format(

transformers/modeling_tf_t5.py

@@ -0,0 +1,775 @@
+# coding=utf-8
+# Copyright 2018 T5 Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TF 2.0 T5 model. """
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import logging
+import math
+import copy
+import itertools
+
+import tensorflow as tf
+
+from .configuration_t5 import T5Config
+from .modeling_tf_utils import TFPreTrainedModel, TFSharedEmbeddings, shape_list
+from .file_utils import add_start_docstrings, DUMMY_INPUTS, DUMMY_MASK
+
+logger = logging.getLogger(__name__)
+
+TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP = {
+    't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-small-tf_model.h5",
+    't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-base-tf_model.h5",
+    't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-large-tf_model.h5",
+    't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-3b-tf_model.h5",
+    't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-11b-tf_model.h5",
+}
+
+####################################################
+# TF 2.0 Models are constructed using Keras imperative API by sub-classing
+# - tf.keras.layers.Layer for the layers and
+# - TFPreTrainedModel for the models (it-self a sub-class of tf.keras.Model)
+####################################################
+
+class TFT5LayerNorm(tf.keras.layers.Layer):
+    def __init__(self, epsilon=1e-6, **kwargs):
+        """ Construct a layernorm module in the T5 style
+            No bias and no substraction of mean.
+        """
+        super(TFT5LayerNorm, self).__init__(**kwargs)
+        self.variance_epsilon = epsilon
+
+    def build(self, input_shape):
+        """Build shared word embedding layer """
+        self.weight = self.add_weight(
+            "weight",
+            shape=(input_shape[-1],),
+            initializer='ones')
+        super(TFT5LayerNorm, self).build(input_shape)
+
+    def call(self, x):
+        variance = tf.math.reduce_mean(tf.math.square(x), axis=-1, keepdims=True)
+        x = x * tf.math.rsqrt(variance + self.variance_epsilon)
+        return self.weight * x
+
+
+class TFT5DenseReluDense(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFT5DenseReluDense, self).__init__(**kwargs)
+        self.wi = tf.keras.layers.Dense(config.d_ff, use_bias=False, name='wi')
+        self.wo = tf.keras.layers.Dense(config.d_model, use_bias=False, name='wo')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+        self.act = tf.keras.activations.relu
+
+    def call(self, hidden_states, training=False):
+        h = self.wi(hidden_states)
+        h = self.act(h)
+        h = self.dropout(h, training=training)
+        h = self.wo(h)
+        return h
+
+
+class TFT5LayerFF(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFT5LayerFF, self).__init__(**kwargs)
+        self.DenseReluDense = TFT5DenseReluDense(config, name='DenseReluDense')
+        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                        name='layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def call(self, hidden_states, training=False):
+        norm_x = self.layer_norm(hidden_states)
+        y = self.DenseReluDense(norm_x, training=training)
+        layer_output = hidden_states + self.dropout(y, training=training)
+        return layer_output
+
+
+class TFT5Attention(tf.keras.layers.Layer):
+    NEW_ID = itertools.count()
+
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5Attention, self).__init__(**kwargs)
+        self.layer_id = next(TFT5Attention.NEW_ID)
+        self.is_decoder = config.is_decoder
+        self.has_relative_attention_bias = has_relative_attention_bias
+
+        self.output_attentions = config.output_attentions
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.d_model = config.d_model
+        self.d_kv = config.d_kv
+        self.n_heads = config.num_heads
+        self.inner_dim = self.n_heads * self.d_kv
+
+        # Mesh TensorFlow initialization to avoid scaling before softmax
+        self.q = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='q')
+        self.k = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='k')
+        self.v = tf.keras.layers.Dense(self.inner_dim, use_bias=False, name='v')
+        self.o = tf.keras.layers.Dense(self.d_model, use_bias=False, name='o')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = tf.keras.layers.Embedding(self.relative_attention_num_buckets,
+                                                                     self.n_heads,
+                                                                     name='relative_attention_bias')
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        raise NotImplementedError
+
+    @staticmethod
+    def _relative_position_bucket(relative_position,
+                                  bidirectional=True,
+                                  num_buckets=32,
+                                  max_distance=128):
+        """
+        Adapted from Mesh Tensorflow:
+        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+        Translate relative position to a bucket number for relative attention.
+        The relative position is defined as memory_position - query_position, i.e.
+        the distance in tokens from the attending position to the attended-to
+        position.  If bidirectional=False, then positive relative positions are
+        invalid.
+        We use smaller buckets for small absolute relative_position and larger buckets
+        for larger absolute relative_positions.  All relative positions >=max_distance
+        map to the same bucket.  All relative positions <=-max_distance map to the
+        same bucket.  This should allow for more graceful generalization to longer
+        sequences than the model has been trained on.
+        Args:
+            relative_position: an int32 Tensor
+            bidirectional: a boolean - whether the attention is bidirectional
+            num_buckets: an integer
+            max_distance: an integer
+        Returns:
+            a Tensor with the same shape as relative_position, containing int32
+            values in the range [0, num_buckets)
+        """
+        ret = 0
+        n = -relative_position
+        if bidirectional:
+            num_buckets //= 2
+            ret += tf.dtypes.cast(tf.math.less(n, 0), tf.int32) * num_buckets
+            n = tf.math.abs(n)
+        else:
+            n = tf.math.maximum(n, 0)
+        # now n is in the range [0, inf)
+        max_exact = num_buckets // 2
+        is_small = tf.math.less(n, max_exact)
+        val_if_large = max_exact + tf.dtypes.cast(
+            tf.math.log(tf.dtypes.cast(n, tf.float32) / max_exact)
+            / math.log(max_distance / max_exact) * (num_buckets - max_exact), tf.int32)
+        val_if_large = tf.math.minimum(val_if_large, num_buckets - 1)
+        ret += tf.where(is_small, n, val_if_large)
+        return ret
+
+    def compute_bias(self, qlen, klen):
+        """ Compute binned relative position bias """
+        context_position = tf.range(qlen)[:, None]
+        memory_position = tf.range(klen)[None, :]
+        relative_position = memory_position - context_position  # shape (qlen, klen)
+        rp_bucket = self._relative_position_bucket(relative_position,
+                                                   bidirectional=not self.is_decoder,
+                                                   num_buckets=self.relative_attention_num_buckets)
+        values = self.relative_attention_bias(rp_bucket)  # shape (qlen, klen, num_heads)
+        values = tf.expand_dims(tf.transpose(values, [2, 0, 1]), axis=0) # shape (1, num_heads, qlen, klen)
+        return values
+
+    def call(self, input, mask=None, kv=None, position_bias=None, cache=None, head_mask=None, training=False):
+        """
+        Self-attention (if kv is None) or attention over source sentence (provided by kv).
+        """
+        # Input is (bs, qlen, dim)
+        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
+        bs, qlen, dim = shape_list(input)
+        if kv is None:
+            klen = qlen if cache is None else cache['slen'] + qlen
+        else:
+            klen = shape_list(kv)[1]
+
+        def shape(x):
+            """  projection """
+            return tf.transpose(tf.reshape(x, (bs, -1, self.n_heads, self.d_kv)), perm=(0, 2, 1, 3))
+
+        def unshape(x):
+            """  compute context """
+            return tf.reshape(tf.transpose(x, perm=(0, 2, 1, 3)), (bs, -1, self.inner_dim))
+
+        q = shape(self.q(input))                                          # (bs, n_heads, qlen, dim_per_head)
+        if kv is None:
+            k = shape(self.k(input))                                      # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(input))                                      # (bs, n_heads, qlen, dim_per_head)
+        elif cache is None or self.layer_id not in cache:
+            k = v = kv
+            k = shape(self.k(k))                                          # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v(v))                                          # (bs, n_heads, qlen, dim_per_head)
+
+        if cache is not None:
+            if self.layer_id in cache:
+                if kv is None:
+                    k_, v_ = cache[self.layer_id]
+                    k = tf.concat([k_, k], axis=2)                             # (bs, n_heads, klen, dim_per_head)
+                    v = tf.concat([v_, v], axis=2)                             # (bs, n_heads, klen, dim_per_head)
+                else:
+                    k, v = cache[self.layer_id]
+            cache[self.layer_id] = (k, v)
+
+        # q = q / math.sqrt(dim_per_head)                                     # No scaling in T5
+        # scores = tf.matmul(q, k, transpose_b=True)                            # (bs, n_heads, qlen, klen)
+        scores = tf.einsum('bnqd,bnkd->bnqk', q, k)                        # (bs, n_heads, qlen, klen)
+
+        if position_bias is None:
+            if not self.has_relative_attention_bias:
+                raise ValueError("No position_bias provided and no weights to compute position_bias")
+            position_bias = self.compute_bias(qlen, klen)
+            if mask is not None:
+                position_bias = position_bias + mask
+                # mask = (mask == 0).expand_as(scores)                              # (bs, n_heads, qlen, klen)
+                # scores.masked_fill_(mask, -float('inf'))                          # (bs, n_heads, qlen, klen)
+
+        scores += position_bias
+        weights = tf.nn.softmax(scores, axis=-1)                              # (bs, n_heads, qlen, klen)
+        weights = self.dropout(weights, training=training)                    # (bs, n_heads, qlen, klen)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = tf.matmul(weights, v)                                    # (bs, n_heads, qlen, dim_per_head)
+        context = unshape(context)                                            # (bs, qlen, dim)
+
+        context = self.o(context)
+
+        outputs = (context,)
+        if self.output_attentions:
+            outputs = outputs + (weights,)
+        if self.has_relative_attention_bias:
+            outputs = outputs + (position_bias,)
+        return outputs
+
+
+class TFT5LayerSelfAttention(tf.keras.layers.Layer):
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5LayerSelfAttention, self).__init__(**kwargs)
+        self.SelfAttention = TFT5Attention(config,
+                                           has_relative_attention_bias=has_relative_attention_bias,
+                                           name='SelfAttention')
+        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                        name='layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def call(self, hidden_states, attention_mask=None, position_bias=None,
+             head_mask=None, training=False):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.SelfAttention(norm_x,
+                                              mask=attention_mask,
+                                              position_bias=position_bias,
+                                              head_mask=head_mask,
+                                              training=training)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y, training=training)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class TFT5LayerCrossAttention(tf.keras.layers.Layer):
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5LayerCrossAttention, self).__init__(**kwargs)
+        self.EncDecAttention = TFT5Attention(config,
+                                           has_relative_attention_bias=has_relative_attention_bias,
+                                           name='EncDecAttention')
+        self.layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                        name='layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def call(self, hidden_states, kv, attention_mask=None, position_bias=None,
+             head_mask=None, training=False):
+        norm_x = self.layer_norm(hidden_states)
+        attention_output = self.EncDecAttention(norm_x,
+                                                mask=attention_mask,
+                                                kv=kv,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask,
+                                                training=training)
+        y = attention_output[0]
+        layer_output = hidden_states + self.dropout(y, training=training)
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class TFT5Block(tf.keras.layers.Layer):
+    def __init__(self, config, has_relative_attention_bias=False, **kwargs):
+        super(TFT5Block, self).__init__(**kwargs)
+        self.is_decoder = config.is_decoder
+        self.layer = []
+        self.layer.append(TFT5LayerSelfAttention(config,
+                                                 has_relative_attention_bias=has_relative_attention_bias,
+                                                 name='layer_._0'))
+        if self.is_decoder:
+            self.layer.append(TFT5LayerCrossAttention(config,
+                                                      has_relative_attention_bias=has_relative_attention_bias,
+                                                      name='layer_._1'))
+            self.layer.append(TFT5LayerFF(config, name='layer_._2'))
+        else:
+            self.layer.append(TFT5LayerFF(config, name='layer_._1'))
+
+    def call(self, hidden_states, attention_mask=None, position_bias=None,
+             encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None,
+             head_mask=None, training=False):
+        self_attention_outputs = self.layer[0](hidden_states,
+                                                attention_mask=attention_mask,
+                                                position_bias=position_bias,
+                                                head_mask=head_mask,
+                                                training=training)
+        hidden_states = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]
+
+        if not self.is_decoder:
+            hidden_states = self.layer[1](hidden_states, training=training)
+        else:
+            cross_attention_outputs = self.layer[1](hidden_states,
+                                                    kv=encoder_hidden_states,
+                                                    attention_mask=encoder_attention_mask,
+                                                    position_bias=encoder_decoder_position_bias,
+                                                    head_mask=head_mask,
+                                                    training=training)
+            hidden_states = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:]
+            hidden_states = self.layer[2](hidden_states, training=training)
+
+        outputs = (hidden_states,) + outputs  # add attentions if we output them
+        return outputs # hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+
+
+####################################################
+# The full model without a specific pretrained or finetuning head is
+# provided as a tf.keras.layers.Layer usually called "TFT5MainLayer"
+####################################################
+class TFT5MainLayer(tf.keras.layers.Layer):
+    def __init__(self, config, **kwargs):
+        super(TFT5MainLayer, self).__init__(**kwargs)
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.is_decoder = config.is_decoder
+        self.config = config
+        self.num_hidden_layers = config.num_layers
+
+        self.block = [TFT5Block(config,
+                                has_relative_attention_bias=bool(i == 0),
+                                name='block_._{}'.format(i))
+                        for i in range(config.num_layers)]
+        self.final_layer_norm = TFT5LayerNorm(epsilon=config.layer_norm_epsilon,
+                                              name='final_layer_norm')
+        self.dropout = tf.keras.layers.Dropout(config.dropout_rate)
+
+    def _resize_token_embeddings(self, new_num_tokens):
+        raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
+
+    def _prune_heads(self, heads_to_prune):
+        raise NotImplementedError  # Not implemented yet in the library fr TF 2.0 models
+
+    def call(self, hidden_states, attention_mask=None, encoder_hidden_states=None,
+             encoder_attention_mask=None, head_mask=None, training=False):
+
+        batch_size, seq_length = shape_list(hidden_states)[:2]
+        if attention_mask is None:
+            attention_mask = tf.fill((batch_size, seq_length), 1)
+        if self.is_decoder and encoder_attention_mask is None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = tf.fill((batch_size, encoder_seq_length), 1)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        attention_mask = tf.cast(attention_mask, dtype=tf.float32)
+        num_dims_attention_mask = len(shape_list(attention_mask))
+        if num_dims_attention_mask == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif num_dims_attention_mask == 2:
+        # Provided a padding mask of dimensions [batch_size, seq_length]
+        # - if the model is a decoder, apply a causal mask in addition to the padding mask
+        # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if self.config.is_decoder:
+                seq_ids = tf.range(seq_length)
+                causal_mask = tf.less_equal(tf.tile(seq_ids[None, None, :], (batch_size, seq_length, 1)),
+                                            seq_ids[None, :, None])
+                causal_mask = tf.cast(causal_mask, dtype=tf.float32)
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+
+        # T5 has a mask that can compare sequence ids, we can simulate this here with this transposistion
+        # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+        # extended_attention_mask = tf.math.equal(extended_attention_mask,
+        #                                         tf.transpose(extended_attention_mask, perm=(-1, -2)))
+
+        extended_attention_mask = (1.0 - extended_attention_mask) * -1e9
+
+        if self.is_decoder:
+            # If a 2D ou 3D attention mask is provided for the cross-attention
+            # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+            encoder_attention_mask = tf.cast(encoder_attention_mask, dtype=tf.float32)
+            num_dims_encoder_attention_mask = len(shape_list(encoder_attention_mask))
+            if num_dims_encoder_attention_mask == 3:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+            if num_dims_encoder_attention_mask == 2:
+                encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+            # T5 has a mask that can compare sequence ids, we can simulate this here with this transposistion
+            # Cf. https://github.com/tensorflow/mesh/blob/8d2465e9bc93129b913b5ccc6a59aa97abd96ec6/mesh_tensorflow/transformer/transformer_layers.py#L270
+            # encoder_extended_attention_mask = tf.math.equal(encoder_extended_attention_mask,
+            #                                         tf.transpose(encoder_extended_attention_mask, perm=(-1, -2)))
+
+            encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -1e9
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        if not head_mask is None:
+            raise NotImplementedError
+        else:
+            head_mask = [None] * self.num_hidden_layers
+            # head_mask = tf.constant([0] * self.num_hidden_layers)
+
+        all_hidden_states = ()
+        all_attentions = ()
+        position_bias = None
+        encoder_decoder_position_bias = None
+        for i, layer_module in enumerate(self.block):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states,
+                                         attention_mask=extended_attention_mask,
+                                         position_bias=position_bias,
+                                         encoder_hidden_states=encoder_hidden_states,
+                                         encoder_attention_mask=encoder_extended_attention_mask,
+                                         encoder_decoder_position_bias=encoder_decoder_position_bias,
+                                         head_mask=head_mask[i],
+                                         training=training)
+            hidden_states = layer_outputs[0]
+            if i == 0:
+                # We share the position biases between the layers - the first layer store them
+                # layer_outputs = hidden-states, (self-attention weights), (self-attention position bias), (cross-attention weights), (cross-attention position bias)
+                position_bias = layer_outputs[2 if self.output_attentions else 1]
+                if self.is_decoder:
+                    encoder_decoder_position_bias = layer_outputs[4 if self.output_attentions else 2]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        layer_output = self.dropout(hidden_states, training=training)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
+
+
+####################################################
+# TFT5PreTrainedModel is a sub-class of tf.keras.Model
+# which take care of loading and saving pretrained weights
+# and various common utilities.
+# Here you just need to specify a few (self-explanatory)
+# pointers for your model.
+####################################################
+class TFT5PreTrainedModel(TFPreTrainedModel):
+    """ An abstract class to handle weights initialization and
+        a simple interface for dowloading and loading pretrained models.
+    """
+    config_class = T5Config
+    pretrained_model_archive_map = TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP
+    base_model_prefix = "transformer"
+
+    @property
+    def dummy_inputs(self):
+        input_ids = tf.constant(DUMMY_INPUTS)
+        input_mask = tf.constant(DUMMY_MASK)
+        dummy_inputs = {'decoder_input_ids': input_ids,
+                        'encoder_input_ids': input_ids,
+                        'decoder_attention_mask': input_mask}
+        return dummy_inputs
+
+
+T5_START_DOCSTRING = r"""    The T5 model was proposed in
+    `Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`_
+    by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu.
+    It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting.
+
+    This model is a tf.keras.Model `tf.keras.Model`_ sub-class. Use it as a regular TF 2.0 Keras Model and
+    refer to the TF 2.0 documentation for all matter related to general usage and behavior.
+
+    .. _`Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer`:
+        https://arxiv.org/abs/1910.10683
+
+    .. _`tf.keras.Model`:
+        https://www.tensorflow.org/versions/r2.0/api_docs/python/tf/keras/Model
+
+    Note on the model inputs:
+        TF 2.0 models accepts two formats as inputs:
+
+            - having all inputs as keyword arguments (like PyTorch models), or
+            - having all inputs as a list, tuple or dict in the first positional arguments.
+
+        This second option is usefull when using `tf.keras.Model.fit()` method which currently requires having all the tensors in the first argument of the model call function: `model(inputs)`.
+
+        If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :
+
+        - a single Tensor with input_ids only and nothing else: `model(inputs_ids)
+        - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring:
+            `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])`
+        - a dictionary with one or several input Tensors associaed to the input names given in the docstring:
+            `model({'input_ids': input_ids, 'token_type_ids': token_type_ids})`
+
+    Parameters:
+        config (:class:`~transformers.T5Config`): Model configuration class with all the parameters of the model. 
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+"""
+
+T5_INPUTS_DOCSTRING = r"""
+    Inputs:
+        **input_ids**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Indices of input sequence tokens in the vocabulary.
+            To match pre-training, T5 input sequence should be formatted with [CLS] and [SEP] tokens as follows:
+
+            (a) For sequence pairs:
+
+                ``tokens:         [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]``
+
+            (b) For single sequences:
+
+                ``tokens:         [CLS] the dog is hairy . [SEP]``
+
+
+            T5 is a model with relative position embeddings so you should be able to pad the inputs on
+            the right or the left.
+
+            Indices can be obtained using :class:`transformers.T5Tokenizer`.
+            See :func:`transformers.PreTrainedTokenizer.encode` and
+            :func:`transformers.PreTrainedTokenizer.convert_tokens_to_ids` for details.
+        **attention_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length)``:
+            Mask to avoid performing attention on padding token indices.
+            Mask values selected in ``[0, 1]``:
+            ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        **head_mask**: (`optional`) ``Numpy array`` or ``tf.Tensor`` of shape ``(num_heads,)`` or ``(num_layers, num_heads)``:
+            Mask to nullify selected heads of the self-attention modules.
+            Mask values selected in ``[0, 1]``:
+            ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+"""
+
+@add_start_docstrings("The bare T5 Model transformer outputting raw hidden-states"
+                      "without any specific head on top.",
+                      T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class TFT5Model(TFT5PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **last_hidden_state**: ``tf.Tensor`` of shape ``(batch_size, sequence_length, hidden_size)``
+            Sequence of hidden-states at the output of the last layer of the model.
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import T5Tokenizer, TFT5Model
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5Model.from_pretrained('t5-small')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFT5Model, self).__init__(config, *inputs, **kwargs)
+        self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model,
+                                         name='shared')
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = TFT5MainLayer(encoder_config, name='encoder')
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = TFT5MainLayer(decoder_config, name='decoder')
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def get_output_embeddings(self):
+        return self.shared
+
+    def call(self, decoder_input_ids, **kwargs):
+        # We allow two types of multi-inputs:
+        # - traditional keyword arguments in the call method
+        # - all the arguments provided as a dict in the first positional argument of call
+        # The last option is useful to use the tf.keras fit() method.
+
+        if isinstance(decoder_input_ids, dict):
+            kwargs.update(decoder_input_ids)
+        else:
+            kwargs['decoder_input_ids'] = decoder_input_ids
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        return decoder_outputs + encoder_outputs
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top. """,
+    T5_START_DOCSTRING, T5_INPUTS_DOCSTRING)
+class TFT5WithLMHeadModel(TFT5PreTrainedModel):
+    r"""
+    Outputs: `Tuple` comprising various elements depending on the configuration (config) and inputs:
+        **prediction_scores**: ``Numpy array`` or ``tf.Tensor`` of shape ``(batch_size, sequence_length, config.vocab_size)``
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        **hidden_states**: (`optional`, returned when ``config.output_hidden_states=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for the output of each layer + the output of the embeddings)
+            of shape ``(batch_size, sequence_length, hidden_size)``:
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        **attentions**: (`optional`, returned when ``config.output_attentions=True``)
+            list of ``Numpy array`` or ``tf.Tensor`` (one for each layer) of shape ``(batch_size, num_heads, sequence_length, sequence_length)``:
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
+
+    Examples::
+
+        import tensorflow as tf
+        from transformers import T5Tokenizer, TFT5WithLMHeadModel
+
+        tokenizer = T5Tokenizer.from_pretrained('t5-small')
+        model = TFT5WithLMHeadModel.from_pretrained('t5-small')
+        input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute"))[None, :]  # Batch size 1
+        outputs = model(input_ids=input_ids)
+        prediction_scores = outputs[0]
+
+    """
+    def __init__(self, config, *inputs, **kwargs):
+        super(TFT5WithLMHeadModel, self).__init__(config, *inputs, **kwargs)
+        self.model_dim = config.d_model
+
+        self.shared = TFSharedEmbeddings(config.vocab_size, config.d_model,
+                                         name='shared')
+
+        encoder_config = copy.deepcopy(config)
+        self.encoder = TFT5MainLayer(encoder_config, name='encoder')
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        self.decoder = TFT5MainLayer(decoder_config, name='decoder')
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def get_output_embeddings(self):
+        return self.shared
+
+    def call(self, decoder_input_ids, **kwargs):
+        # We allow two types of multi-inputs:
+        # - traditional keyword arguments in the call method
+        # - all the arguments provided as a dict in the first positional argument of call
+        # The last option is useful to use the tf.keras fit() method.
+
+        if isinstance(decoder_input_ids, dict):
+            kwargs.update(decoder_input_ids)
+        else:
+            kwargs['decoder_input_ids'] = decoder_input_ids
+
+        kwargs_common = dict((k, v) for k, v in kwargs.items()
+                             if not k.startswith("encoder_") and not k.startswith("decoder_"))
+        kwargs_encoder = kwargs_common.copy()
+        kwargs_decoder = kwargs_common.copy()
+        kwargs_encoder.update(dict((k[len("encoder_"):], v) for k, v in kwargs.items() if k.startswith("encoder_")))
+        kwargs_decoder.update(dict((k[len("decoder_"):], v) for k, v in kwargs.items() if k.startswith("decoder_")))
+
+        # Encode if needed (training, first prediction pass)
+        encoder_hidden_states = kwargs_encoder.pop("hidden_states", None)
+        if encoder_hidden_states is None:
+            # Convert encoder inputs in embeddings if needed
+            hidden_states = kwargs_encoder.pop("inputs_embeds", None)
+            if hidden_states is None:
+                encoder_inputs_ids = kwargs_encoder.pop("input_ids")
+                hidden_states = self.shared(encoder_inputs_ids)  # Convert inputs in embeddings
+
+            encoder_outputs = self.encoder(hidden_states, **kwargs_encoder)
+            encoder_hidden_states = encoder_outputs[0]
+        else:
+            encoder_outputs = ()
+
+        # Decode
+        # Convert decoder inputs in embeddings if needed
+        hidden_states = kwargs_decoder.pop("inputs_embeds", None)
+        if hidden_states is None:
+            decoder_inputs_ids = kwargs_decoder.pop("input_ids")
+            hidden_states = self.shared(decoder_inputs_ids)
+
+        kwargs_decoder["encoder_hidden_states"] = encoder_hidden_states
+        kwargs_decoder["encoder_attention_mask"] = kwargs_encoder.get("attention_mask", None)
+        decoder_outputs = self.decoder(hidden_states, **kwargs_decoder)
+
+        sequence_output = decoder_outputs[0] * (self.model_dim ** -0.5)
+        lm_logits = self.shared(sequence_output, mode="linear")
+        decoder_outputs = (lm_logits,) + decoder_outputs[1:]
+
+        return decoder_outputs + encoder_outputs

transformers/modeling_tf_transfo_xl.py

@@ -353,7 +353,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
 
-        self.n_token = config.n_token
+        self.n_token = config.vocab_size
 
         self.d_embed = config.d_embed
         self.d_model = config.d_model
@@ -361,7 +361,7 @@ class TFTransfoXLMainLayer(tf.keras.layers.Layer):
         self.d_head = config.d_head
         self.untie_r = config.untie_r
 
-        self.word_emb = TFAdaptiveEmbedding(config.n_token, config.d_embed, config.d_model, config.cutoffs, 
+        self.word_emb = TFAdaptiveEmbedding(config.vocab_size, config.d_embed, config.d_model, config.cutoffs, 
                                             div_val=config.div_val, init_std=config.init_std, name='word_emb')
 
         self.drop = tf.keras.layers.Dropout(config.dropout)
@@ -729,7 +729,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
             raise NotImplementedError
         # use adaptive softmax (including standard softmax)
         else:
-            self.crit = TFAdaptiveSoftmaxMask(config.n_token, config.d_embed, config.d_model, 
+            self.crit = TFAdaptiveSoftmaxMask(config.vocab_size, config.d_embed, config.d_model, 
                                               config.cutoffs, div_val=config.div_val, name='crit')
 
     def reset_length(self, tgt_len, ext_len, mem_len):

transformers/modeling_tf_transfo_xl_utilities.py

@@ -25,15 +25,15 @@ import tensorflow as tf
 from .modeling_tf_utils import shape_list
 
 class TFAdaptiveSoftmaxMask(tf.keras.layers.Layer):
-    def __init__(self, n_token, d_embed, d_proj, cutoffs, div_val=1,
+    def __init__(self, vocab_size, d_embed, d_proj, cutoffs, div_val=1,
                  keep_order=False, **kwargs):
         super(TFAdaptiveSoftmaxMask, self).__init__(**kwargs)
 
-        self.n_token = n_token
+        self.vocab_size = vocab_size
         self.d_embed = d_embed
         self.d_proj = d_proj
 
-        self.cutoffs = cutoffs + [n_token]
+        self.cutoffs = cutoffs + [vocab_size]
         self.cutoff_ends = [0] + self.cutoffs
         self.div_val = div_val
 
@@ -66,11 +66,11 @@ class TFAdaptiveSoftmaxMask(tf.keras.layers.Layer):
                     self.out_projs.append(weight)
                 else:
                     self.out_projs.append(None)
-                weight = self.add_weight(shape=(self.n_token, self.d_embed,),
+                weight = self.add_weight(shape=(self.vocab_size, self.d_embed,),
                                          initializer='zeros',
                                          trainable=True,
                                          name='out_layers_._{}_._weight'.format(i))
-                bias = self.add_weight(shape=(self.n_token,),
+                bias = self.add_weight(shape=(self.vocab_size,),
                                          initializer='zeros',
                                          trainable=True,
                                          name='out_layers_._{}_._bias'.format(i))
@@ -114,7 +114,7 @@ class TFAdaptiveSoftmaxMask(tf.keras.layers.Layer):
         hidden, target = inputs
         head_logprob = 0
         if self.n_clusters == 0:
-            softmax_b = tf.get_variable('bias', [n_token], initializer=tf.zeros_initializer())
+            softmax_b = tf.get_variable('bias', [self.config.vocab_size], initializer=tf.zeros_initializer())
             output = self._logit(hidden, self.out_layers[0][0], self.out_layers[0][1], self.out_projs[0])
             if target is not None:
                 loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=target, logits=output)

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)
@@ -299,7 +297,7 @@ class TFPreTrainedModel(tf.keras.Model):
 
         if from_pt:
             # Load from a PyTorch checkpoint
-            return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file)
+            return load_pytorch_checkpoint_in_tf2_model(model, resolved_archive_file, allow_missing_keys=True)
 
         ret = model(model.dummy_inputs, training=False)  # build the network with dummy inputs
 

transformers/modeling_tf_xlm.py

@@ -460,7 +460,7 @@ class TFXLMPreTrainedModel(TFPreTrainedModel):
             langs_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
         else:
             langs_list = None
-        return [inputs_list, attns_list, langs_list]
+        return {'input_ids': inputs_list, 'attention_mask': attns_list, 'langs': langs_list}
 
 
 XLM_START_DOCSTRING = r"""    The XLM model was proposed in

transformers/modeling_tf_xlnet.py

@@ -366,7 +366,7 @@ class TFXLNetMainLayer(tf.keras.layers.Layer):
         self.use_bfloat16 = config.use_bfloat16
         self.initializer_range = config.initializer_range
 
-        self.word_embedding = TFSharedEmbeddings(config.n_token, config.d_model, initializer_range=config.initializer_range, name='word_embedding')
+        self.word_embedding = TFSharedEmbeddings(config.vocab_size, config.d_model, initializer_range=config.initializer_range, name='word_embedding')
         self.layer = [TFXLNetLayer(config, name='layer_._{}'.format(i)) for i in range(config.n_layer)]
         self.dropout = tf.keras.layers.Dropout(config.dropout)
 

transformers/modeling_transfo_xl.py

@@ -592,14 +592,14 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
         self.output_attentions = config.output_attentions
         self.output_hidden_states = config.output_hidden_states
 
-        self.n_token = config.n_token
+        self.n_token = config.vocab_size
 
         self.d_embed = config.d_embed
         self.d_model = config.d_model
         self.n_head = config.n_head
         self.d_head = config.d_head
 
-        self.word_emb = AdaptiveEmbedding(config.n_token, config.d_embed, config.d_model, config.cutoffs,
+        self.word_emb = AdaptiveEmbedding(config.vocab_size, config.d_embed, config.d_model, config.cutoffs,
                                           div_val=config.div_val)
 
         self.drop = nn.Dropout(config.dropout)
@@ -836,11 +836,11 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
         self.sample_softmax = config.sample_softmax
         # use sampled softmax
         if config.sample_softmax > 0:
-            self.out_layer = nn.Linear(config.d_model, config.n_token)
-            self.sampler = LogUniformSampler(config.n_token, config.sample_softmax)
+            self.out_layer = nn.Linear(config.d_model, config.vocab_size)
+            self.sampler = LogUniformSampler(config.vocab_size, config.sample_softmax)
         # use adaptive softmax (including standard softmax)
         else:
-            self.crit = ProjectedAdaptiveLogSoftmax(config.n_token, config.d_embed, config.d_model,
+            self.crit = ProjectedAdaptiveLogSoftmax(config.vocab_size, config.d_embed, config.d_model,
                                                     config.cutoffs, div_val=config.div_val)
         self.init_weights()
 

transformers/modeling_utils.py

@@ -31,12 +31,11 @@ from torch.nn import CrossEntropyLoss
 from torch.nn import functional as F
 
 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)
 
 logger = logging.getLogger(__name__)
 
-
 try:
     from torch.nn import Identity
 except ImportError:
@@ -72,6 +71,15 @@ class PreTrainedModel(nn.Module):
     load_tf_weights = lambda model, config, path: None
     base_model_prefix = ""
 
+    @property
+    def dummy_inputs(self):
+        """ Dummy inputs to do a forward pass in the network.
+
+        Returns:
+            torch.Tensor with dummy inputs
+        """
+        return {'input_ids': torch.tensor(DUMMY_INPUTS)}
+
     def __init__(self, config, *inputs, **kwargs):
         super(PreTrainedModel, self).__init__()
         if not isinstance(config, PretrainedConfig):
@@ -161,8 +169,7 @@ class PreTrainedModel(nn.Module):
         base_model.vocab_size = new_num_tokens
 
         # Tie weights again if needed
-        if hasattr(self, 'tie_weights'):
-            self.tie_weights()
+        self.tie_weights()
 
         return model_embeds
 
@@ -478,8 +485,7 @@ class PreTrainedModel(nn.Module):
                 raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
                                 model.__class__.__name__, "\n\t".join(error_msgs)))
 
-        if hasattr(model, 'tie_weights'):
-            model.tie_weights()  # make sure word embedding weights are still tied
+        model.tie_weights()  # make sure word embedding weights are still tied if needed
 
         # Set model in evaluation mode to desactivate DropOut modules by default
         model.eval()

transformers/modeling_xlm.py

@@ -227,6 +227,16 @@ class XLMPreTrainedModel(PreTrainedModel):
     def __init__(self, *inputs, **kwargs):
         super(XLMPreTrainedModel, self).__init__(*inputs, **kwargs)
 
+    @property
+    def dummy_inputs(self):
+        inputs_list = torch.tensor([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
+        attns_list = torch.tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        if self.config.use_lang_emb and self.config.n_langs > 1:
+            langs_list = torch.tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        else:
+            langs_list = None
+        return {'input_ids': inputs_list, 'attention_mask': attns_list, 'langs': langs_list}
+
     def _init_weights(self, module):
         """ Initialize the weights. """
         if isinstance(module, nn.Embedding):
@@ -646,7 +656,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
                                                langs=langs,
                                                token_type_ids=token_type_ids,
                                                position_ids=position_ids,
-                                               lengths=lengths, 
+                                               lengths=lengths,
                                                cache=cache,
                                                head_mask=head_mask,
                                                inputs_embeds=inputs_embeds)

transformers/modeling_xlnet.py

@@ -609,7 +609,7 @@ class XLNetModel(XLNetPreTrainedModel):
         self.clamp_len = config.clamp_len
         self.n_layer = config.n_layer
 
-        self.word_embedding = nn.Embedding(config.n_token, config.d_model)
+        self.word_embedding = nn.Embedding(config.vocab_size, config.d_model)
         self.mask_emb = nn.Parameter(torch.FloatTensor(1, 1, config.d_model))
         self.layer = nn.ModuleList([XLNetLayer(config) for _ in range(config.n_layer)])
         self.dropout = nn.Dropout(config.dropout)
@@ -940,7 +940,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         self.same_length = config.same_length
 
         self.transformer = XLNetModel(config)
-        self.lm_loss = nn.Linear(config.d_model, config.n_token, bias=True)
+        self.lm_loss = nn.Linear(config.d_model, config.vocab_size, bias=True)
 
         self.init_weights()
 

transformers/tests/configuration_common_test.py

@@ -16,15 +16,12 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import copy
 import os
-import shutil
 import json
-import random
-import uuid
+import tempfile
 
 import unittest
-import logging
+from .tokenization_tests_commons import TemporaryDirectory
 
 
 class ConfigTester(object):
@@ -48,16 +45,28 @@ class ConfigTester(object):
 
     def create_and_test_config_to_json_file(self):
         config_first = self.config_class(**self.inputs_dict)
-        json_file_path = os.path.join(os.getcwd(), "config_" + str(uuid.uuid4()) + ".json")
-        config_first.to_json_file(json_file_path)
-        config_second = self.config_class.from_json_file(json_file_path)
-        os.remove(json_file_path)
+
+        with TemporaryDirectory() as tmpdirname:
+            json_file_path = os.path.join(tmpdirname, "config.json")
+            config_first.to_json_file(json_file_path)
+            config_second = self.config_class.from_json_file(json_file_path)
+
+        self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
+
+    def create_and_test_config_from_and_save_pretrained(self):
+        config_first = self.config_class(**self.inputs_dict)
+
+        with TemporaryDirectory() as tmpdirname:
+            config_first.save_pretrained(tmpdirname)
+            config_second = self.config_class.from_pretrained(tmpdirname)
+
         self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
 
     def run_common_tests(self):
         self.create_and_test_config_common_properties()
         self.create_and_test_config_to_json_string()
         self.create_and_test_config_to_json_file()
+        self.create_and_test_config_from_and_save_pretrained()
 
 if __name__ == "__main__":
     unittest.main()

transformers/tests/hf_api_test.py

@@ -15,18 +15,30 @@
 from __future__ import absolute_import, division, print_function
 
 import os
-import six
 import time
 import unittest
 
-from transformers.hf_api import HfApi, S3Obj, PresignedUrl, HfFolder, HTTPError
+import requests
+import six
+
+from transformers.hf_api import HfApi, HfFolder, HTTPError, PresignedUrl, S3Obj
 
 USER = "__DUMMY_TRANSFORMERS_USER__"
 PASS = "__DUMMY_TRANSFORMERS_PASS__"
-FILE_KEY = "Test-{}.txt".format(int(time.time()))
-FILE_PATH = os.path.join(
-    os.path.dirname(os.path.abspath(__file__)), "fixtures/input.txt"
-)
+FILES = [
+    (
+        "Test-{}.txt".format(int(time.time())),
+        os.path.join(
+            os.path.dirname(os.path.abspath(__file__)), "fixtures/input.txt"
+        )
+    ),
+    (
+        "yoyo {}.txt".format(int(time.time())), # space is intentional
+        os.path.join(
+            os.path.dirname(os.path.abspath(__file__)), "fixtures/empty.txt"
+        )
+    ),
+]
 
 
 
@@ -57,15 +69,21 @@ class HfApiEndpointsTest(HfApiCommonTest):
         self.assertEqual(user, USER)
 
     def test_presign(self):
-        urls = self._api.presign(token=self._token, filename=FILE_KEY)
-        self.assertIsInstance(urls, PresignedUrl)
-        self.assertEqual(urls.type, "text/plain")
+        for FILE_KEY, FILE_PATH in FILES:
+            urls = self._api.presign(token=self._token, filename=FILE_KEY)
+            self.assertIsInstance(urls, PresignedUrl)
+            self.assertEqual(urls.type, "text/plain")
 
     def test_presign_and_upload(self):
-        access_url = self._api.presign_and_upload(
-            token=self._token, filename=FILE_KEY, filepath=FILE_PATH
-        )
-        self.assertIsInstance(access_url, six.string_types)
+        for FILE_KEY, FILE_PATH in FILES:
+            access_url = self._api.presign_and_upload(
+                token=self._token, filename=FILE_KEY, filepath=FILE_PATH
+            )
+            self.assertIsInstance(access_url, six.string_types)
+            with open(FILE_PATH, 'r') as f:
+                body = f.read()
+            r = requests.get(access_url)
+            self.assertEqual(r.text, body)
 
     def test_list_objs(self):
         objs = self._api.list_objs(token=self._token)

transformers/tests/model_card_test.py

@@ -0,0 +1,89 @@
+# coding=utf-8
+# Copyright 2019 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.
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+import json
+import unittest
+
+from transformers.model_card import ModelCard
+from .tokenization_tests_commons import TemporaryDirectory
+
+class ModelCardTester(unittest.TestCase):
+
+    def setUp(self):
+        self.inputs_dict = {'model_details': {
+                                'Organization': 'testing',
+                                'Model date': 'today',
+                                'Model version': 'v2.1, Developed by Test Corp in 2019.',
+                                'Architecture': 'Convolutional Neural Network.',
+                                },
+                            'metrics': 'BLEU and ROUGE-1',
+                            'evaluation_data':{
+                                'Datasets':{
+                                    'BLEU': 'My-great-dataset-v1',
+                                    'ROUGE-1': 'My-short-dataset-v2.1',
+                                },
+                                'Preprocessing': 'See details on https://arxiv.org/pdf/1810.03993.pdf'
+                            },
+                            'training_data':{
+                                'Dataset': 'English Wikipedia dump dated 2018-12-01',
+                                'Preprocessing': 'Using SentencePiece vocabulary of size 52k tokens. See details on https://arxiv.org/pdf/1810.03993.pdf'
+                            },
+                            'quantitative_analyses': {
+                                'BLEU': 55.1,
+                                'ROUGE-1': 76,
+                            },
+                            }
+
+    def test_model_card_common_properties(self):
+        model_card = ModelCard.from_dict(self.inputs_dict)
+        self.assertTrue(hasattr(model_card, 'model_details'))
+        self.assertTrue(hasattr(model_card, 'intended_use'))
+        self.assertTrue(hasattr(model_card, 'factors'))
+        self.assertTrue(hasattr(model_card, 'metrics'))
+        self.assertTrue(hasattr(model_card, 'evaluation_data'))
+        self.assertTrue(hasattr(model_card, 'training_data'))
+        self.assertTrue(hasattr(model_card, 'quantitative_analyses'))
+        self.assertTrue(hasattr(model_card, 'ethical_considerations'))
+        self.assertTrue(hasattr(model_card, 'caveats_and_recommendations'))
+
+    def test_model_card_to_json_string(self):
+        model_card = ModelCard.from_dict(self.inputs_dict)
+        obj = json.loads(model_card.to_json_string())
+        for key, value in self.inputs_dict.items():
+            self.assertEqual(obj[key], value)
+
+    def test_model_card_to_json_file(self):
+        model_card_first = ModelCard.from_dict(self.inputs_dict)
+
+        with TemporaryDirectory() as tmpdirname:
+            filename = os.path.join(tmpdirname, u"model_card.json")
+            model_card_first.to_json_file(filename)
+            model_card_second = ModelCard.from_json_file(filename)
+
+        self.assertEqual(model_card_second.to_dict(), model_card_first.to_dict())
+
+    def test_model_card_from_and_save_pretrained(self):
+        model_card_first = ModelCard.from_dict(self.inputs_dict)
+
+        with TemporaryDirectory() as tmpdirname:
+            model_card_first.save_pretrained(tmpdirname)
+            model_card_second = ModelCard.from_pretrained(tmpdirname)
+
+        self.assertEqual(model_card_second.to_dict(), model_card_first.to_dict())
+
+if __name__ == "__main__":
+    unittest.main()

transformers/tests/modeling_albert_test.py

@@ -110,7 +110,7 @@ class AlbertModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = AlbertConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

transformers/tests/modeling_bert_test.py

@@ -109,7 +109,7 @@ class BertModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = BertConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

transformers/tests/modeling_common_test.py

@@ -58,7 +58,7 @@ else:
 def _config_zero_init(config):
     configs_no_init = copy.deepcopy(config)
     for key in configs_no_init.__dict__.keys():
-        if '_range' in key or '_std' in key:
+        if '_range' in key or '_std' in key or 'initializer_factor' in key:
             setattr(configs_no_init, key, 0.0)
     return configs_no_init
 
@@ -73,6 +73,7 @@ class CommonTestCases:
         test_pruning = True
         test_resize_embeddings = True
         test_head_masking = True
+        is_encoder_decoder = False
 
         def test_save_load(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -83,6 +84,8 @@ class CommonTestCases:
                 model.eval()
                 with torch.no_grad():
                     outputs = model(**inputs_dict)
+                out_2 = outputs[0].numpy()
+                out_2[np.isnan(out_2)] = 0
 
                 with TemporaryDirectory() as tmpdirname:
                     model.save_pretrained(tmpdirname)
@@ -93,9 +96,7 @@ class CommonTestCases:
 
                     # Make sure we don't have nans
                     out_1 = after_outputs[0].cpu().numpy()
-                    out_2 = outputs[0].cpu().numpy()
-                    out_1 = out_1[~np.isnan(out_1)]
-                    out_2 = out_2[~np.isnan(out_2)]
+                    out_1[np.isnan(out_1)] = 0
                     max_diff = np.amax(np.abs(out_1 - out_2))
                     self.assertLessEqual(max_diff, 1e-5)
 
@@ -117,20 +118,32 @@ class CommonTestCases:
                 model = model_class(config)
                 model.to(torch_device)
                 model.eval()
-                first, second = model(inputs_dict["input_ids"])[0], model(inputs_dict["input_ids"])[0]
-                self.assertEqual(first.ne(second).sum().item(), 0)
-
+                with torch.no_grad():
+                    first = model(**inputs_dict)[0]
+                    second = model(**inputs_dict)[0]
+                out_1 = first.cpu().numpy()
+                out_2 = second.cpu().numpy()
+                out_1 = out_1[~np.isnan(out_1)]
+                out_2 = out_2[~np.isnan(out_2)]
+                max_diff = np.amax(np.abs(out_1 - out_2))
+                self.assertLessEqual(max_diff, 1e-5)
 
         def test_attention_outputs(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
+            decoder_seq_length = self.model_tester.decoder_seq_length if hasattr(self.model_tester, 'decoder_seq_length') else self.model_tester.seq_length
+            encoder_seq_length = self.model_tester.encoder_seq_length if hasattr(self.model_tester, 'encoder_seq_length') else self.model_tester.seq_length
+            decoder_key_length = self.model_tester.key_length if hasattr(self.model_tester, 'key_length') else decoder_seq_length
+            encoder_key_length = self.model_tester.key_length if hasattr(self.model_tester, 'key_length') else encoder_seq_length
+
             for model_class in self.all_model_classes:
                 config.output_attentions = True
                 config.output_hidden_states = False
                 model = model_class(config)
                 model.to(torch_device)
                 model.eval()
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 attentions = outputs[-1]
                 self.assertEqual(model.config.output_attentions, True)
                 self.assertEqual(model.config.output_hidden_states, False)
@@ -138,28 +151,42 @@ class CommonTestCases:
                 self.assertListEqual(
                     list(attentions[0].shape[-3:]),
                     [self.model_tester.num_attention_heads,
-                    self.model_tester.seq_length,
-                    self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
+                    encoder_seq_length ,
+                    encoder_key_length])
                 out_len = len(outputs)
 
+                if self.is_encoder_decoder:
+                    self.assertEqual(out_len % 2, 0)
+                    decoder_attentions = outputs[(out_len // 2)-1]
+                    self.assertEqual(model.config.output_attentions, True)
+                    self.assertEqual(model.config.output_hidden_states, False)
+                    self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
+                    self.assertListEqual(
+                        list(decoder_attentions[0].shape[-3:]),
+                        [self.model_tester.num_attention_heads,
+                         decoder_seq_length,
+                         decoder_key_length
+                         ])
+
                 # Check attention is always last and order is fine
                 config.output_attentions = True
                 config.output_hidden_states = True
                 model = model_class(config)
                 model.to(torch_device)
                 model.eval()
-                outputs = model(**inputs_dict)
-                self.assertEqual(out_len+1, len(outputs))
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
+                self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
                 self.assertEqual(model.config.output_attentions, True)
                 self.assertEqual(model.config.output_hidden_states, True)
 
-                attentions = outputs[-1]
-                self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+                self_attentions = outputs[-1]
+                self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
                 self.assertListEqual(
-                    list(attentions[0].shape[-3:]),
+                    list(self_attentions[0].shape[-3:]),
                     [self.model_tester.num_attention_heads,
-                    self.model_tester.seq_length,
-                    self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
+                    encoder_seq_length,
+                    encoder_key_length])
 
         def test_torchscript(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -223,7 +250,6 @@ class CommonTestCases:
 
                 self.assertTrue(models_equal)
 
-
         def test_headmasking(self):
             if not self.test_head_masking:
                 return
@@ -278,7 +304,6 @@ class CommonTestCases:
                 self.assertNotEqual(
                     attentions[-1][..., -1, :, :].flatten().sum().item(), 0.0)
 
-
         def test_head_pruning(self):
             if not self.test_pruning:
                 return
@@ -297,7 +322,8 @@ class CommonTestCases:
                 heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
                                 -1: [0]}
                 model.prune_heads(heads_to_prune)
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
 
                 attentions = outputs[-1]
 
@@ -333,7 +359,8 @@ class CommonTestCases:
                 model = model_class.from_pretrained(directory)
                 model.to(torch_device)
 
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 attentions = outputs[-1]
                 self.assertEqual(attentions[0].shape[-3], 1)
                 self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
@@ -362,7 +389,8 @@ class CommonTestCases:
                 model.to(torch_device)
                 model.eval()
 
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 attentions = outputs[-1]
 
                 self.assertEqual(attentions[0].shape[-3], 1)
@@ -389,7 +417,8 @@ class CommonTestCases:
                 model.to(torch_device)
                 model.eval()
 
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 attentions = outputs[-1]
 
                 self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
@@ -406,7 +435,8 @@ class CommonTestCases:
                 model.to(torch_device)
                 shutil.rmtree(directory)
 
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 attentions = outputs[-1]
 
                 self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
@@ -417,7 +447,8 @@ class CommonTestCases:
                 heads_to_prune = {0: [0], 2: [1, 2]}
                 model.prune_heads(heads_to_prune)
 
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 attentions = outputs[-1]
 
                 self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads -1)
@@ -427,7 +458,6 @@ class CommonTestCases:
 
                 self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
 
-
         def test_hidden_states_output(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
@@ -437,14 +467,16 @@ class CommonTestCases:
                 model = model_class(config)
                 model.to(torch_device)
                 model.eval()
-                outputs = model(**inputs_dict)
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
                 hidden_states = outputs[-1]
                 self.assertEqual(model.config.output_attentions, False)
                 self.assertEqual(model.config.output_hidden_states, True)
                 self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1)
                 self.assertListEqual(
                     list(hidden_states[0].shape[-2:]),
-                    [self.model_tester.seq_length, self.model_tester.hidden_size])
+                    [self.model_tester.encoder_seq_length if hasattr(self.model_tester, 'encoder_seq_length') else self.model_tester.seq_length,
+                     self.model_tester.hidden_size])
 
         def test_resize_tokens_embeddings(self):
             original_config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -550,8 +582,14 @@ class CommonTestCases:
 
         def test_inputs_embeds(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-            input_ids = inputs_dict["input_ids"]
-            del inputs_dict["input_ids"]
+            if not self.is_encoder_decoder:
+                input_ids = inputs_dict["input_ids"]
+                del inputs_dict["input_ids"]
+            else:
+                encoder_input_ids = inputs_dict["encoder_input_ids"]
+                decoder_input_ids = inputs_dict["decoder_input_ids"]
+                del inputs_dict["encoder_input_ids"]
+                del inputs_dict["decoder_input_ids"]
 
             for model_class in self.all_model_classes:
                 model = model_class(config)
@@ -559,9 +597,14 @@ class CommonTestCases:
                 model.eval()
 
                 wte = model.get_input_embeddings()
-                inputs_dict["inputs_embeds"] = wte(input_ids)
-                outputs = model(**inputs_dict)
+                if not self.is_encoder_decoder:
+                    inputs_dict["inputs_embeds"] = wte(input_ids)
+                else:
+                    inputs_dict["encoder_inputs_embeds"] = wte(encoder_input_ids)
+                    inputs_dict["decoder_inputs_embeds"] = wte(decoder_input_ids)
 
+                with torch.no_grad():
+                    outputs = model(**inputs_dict)
 
     class GPTModelTester(CommonModelTester):
 
@@ -633,7 +676,7 @@ class CommonTestCases:
                 mc_token_ids = ids_tensor([self.batch_size, self.n_choices], self.seq_length)
 
             config = self.config_class(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_positions=self.n_positions,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
@@ -649,9 +692,10 @@ class CommonTestCases:
             model.to(torch_device)
             model.eval()
 
-            outputs = model(input_ids, position_ids, token_type_ids)
-            outputs = model(input_ids, position_ids)
-            outputs = model(input_ids)
+            with torch.no_grad():
+                outputs = model(input_ids, position_ids, token_type_ids)
+                outputs = model(input_ids, position_ids)
+                outputs = model(input_ids)
 
             hidden_state = outputs[0]
             self.parent.assertListEqual(
@@ -664,7 +708,8 @@ class CommonTestCases:
             model = self.lm_head_model_class(config)
             model.to(torch_device)
             model.eval()
-            outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
+            with torch.no_grad():
+                outputs = model(input_ids, position_ids, token_type_ids, lm_labels)
             loss, lm_logits = outputs[:2]
 
             total_voc = self.vocab_size
@@ -681,7 +726,8 @@ class CommonTestCases:
                 model = model_class(config)
                 model.to(torch_device)
                 model.eval()
-                outputs = model(input_ids)
+                with torch.no_grad():
+                    outputs = model(input_ids)
                 presents = outputs[-1]
                 self.parent.assertEqual(self.num_hidden_layers, len(presents))
                 self.parent.assertListEqual(
@@ -694,7 +740,8 @@ class CommonTestCases:
             model = self.double_head_model_class(config)
             model.to(torch_device)
             model.eval()
-            outputs = model(input_ids, mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels,
+            with torch.no_grad():
+                outputs = model(input_ids, mc_token_ids, lm_labels=lm_labels, mc_labels=mc_labels,
                             token_type_ids=token_type_ids, position_ids=position_ids)
             lm_loss, mc_loss, lm_logits, mc_logits = outputs[:4]
             loss = [lm_loss, mc_loss]

transformers/tests/modeling_ctrl_test.py

@@ -114,7 +114,7 @@ class CTRLModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = CTRLConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
                 n_head=self.num_attention_heads,

transformers/tests/modeling_distilbert_test.py

@@ -105,7 +105,7 @@ class DistilBertModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = DistilBertConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 dim=self.hidden_size,
                 n_layers=self.num_hidden_layers,
                 n_heads=self.num_attention_heads,

transformers/tests/modeling_gpt2_test.py

@@ -110,7 +110,7 @@ class GPT2ModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = GPT2Config(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
                 n_head=self.num_attention_heads,

transformers/tests/modeling_openai_test.py

@@ -98,7 +98,7 @@ class OpenAIGPTModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = OpenAIGPTConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
                 n_head=self.num_attention_heads,

transformers/tests/modeling_roberta_test.py

@@ -106,7 +106,7 @@ class RobertaModelTest(CommonTestCases.CommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = RobertaConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

transformers/tests/modeling_t5_test.py

@@ -0,0 +1,185 @@
+# coding=utf-8
+# Copyright 2018 Google 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.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+
+from transformers import is_torch_available
+
+from .modeling_common_test import (CommonTestCases, ids_tensor, floats_tensor)
+from .configuration_common_test import ConfigTester
+from .utils import require_torch, slow, torch_device
+
+if is_torch_available():
+    from transformers import (T5Config, T5Model, T5WithLMHeadModel)
+    from transformers.modeling_t5 import T5_PRETRAINED_MODEL_ARCHIVE_MAP
+
+
+@require_torch
+class T5ModelTest(CommonTestCases.CommonModelTester):
+
+    all_model_classes = (T5Model, T5WithLMHeadModel) if is_torch_available() else ()
+    test_pruning = False
+    test_torchscript = False
+    test_resize_embeddings = False
+    is_encoder_decoder = True
+
+    class T5ModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     encoder_seq_length=7,
+                     decoder_seq_length=9,
+                     is_training=True,
+                     use_attention_mask=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     n_positions=14,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     d_ff=37,
+                     relative_attention_num_buckets=8,
+                     dropout_rate=0.1,
+                     initializer_factor=0.002,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.encoder_seq_length = encoder_seq_length
+            self.decoder_seq_length = decoder_seq_length
+            self.is_training = is_training
+            self.use_attention_mask = use_attention_mask
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.n_positions = n_positions
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.d_ff = d_ff
+            self.relative_attention_num_buckets = relative_attention_num_buckets
+            self.dropout_rate = dropout_rate
+            self.initializer_factor = initializer_factor
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            encoder_input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size)
+            decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+            encoder_attention_mask = None
+            decoder_attention_mask = None
+            if self.use_attention_mask:
+                encoder_attention_mask = ids_tensor([self.batch_size, self.encoder_seq_length], vocab_size=2)
+                decoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+            decoder_lm_labels = None
+            if self.use_labels:
+                decoder_lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+            config = T5Config(
+                vocab_size=self.vocab_size,
+                n_positions=self.n_positions,
+                d_model=self.hidden_size,
+                d_ff=self.d_ff,
+                d_kv=self.hidden_size // self.num_attention_heads,
+                num_layers=self.num_hidden_layers,
+                num_heads=self.num_attention_heads,
+                relative_attention_num_buckets=self.relative_attention_num_buckets,
+                dropout_rate=self.dropout_rate,
+                initializer_factor=self.initializer_factor)
+
+            return (config, encoder_input_ids, decoder_input_ids, encoder_attention_mask, decoder_attention_mask, decoder_lm_labels)
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(
+                list(result["loss"].size()),
+                [])
+
+        def create_and_check_t5_model(self, config, encoder_input_ids, decoder_input_ids, encoder_attention_mask, decoder_attention_mask, decoder_lm_labels):
+            model = T5Model(config=config)
+            model.eval()
+            decoder_output, encoder_output = model(encoder_input_ids=encoder_input_ids,
+                                                   decoder_input_ids=decoder_input_ids,
+                                                   encoder_attention_mask=encoder_attention_mask,
+                                                   decoder_attention_mask=decoder_attention_mask)
+            decoder_output, encoder_output = model(encoder_input_ids=encoder_input_ids,
+                                                   decoder_input_ids=decoder_input_ids)
+
+            result = {
+                "encoder_output": encoder_output,
+                "decoder_output": decoder_output,
+            }
+            self.parent.assertListEqual(
+                list(result["encoder_output"].size()),
+                [self.batch_size, self.encoder_seq_length, self.hidden_size])
+            self.parent.assertListEqual(
+                list(result["decoder_output"].size()),
+                [self.batch_size, self.decoder_seq_length, self.hidden_size])
+
+
+        def create_and_check_t5_with_lm_head(self, config, encoder_input_ids, decoder_input_ids, encoder_attention_mask, decoder_attention_mask, decoder_lm_labels):
+            model = T5WithLMHeadModel(config=config)
+            model.eval()
+            outputs = model(encoder_input_ids=encoder_input_ids, decoder_input_ids=decoder_input_ids,
+                            decoder_attention_mask=decoder_attention_mask, decoder_lm_labels=decoder_lm_labels)
+            loss, prediction_scores = outputs[0], outputs[1]
+            result = {
+                "loss": loss,
+                "prediction_scores": prediction_scores,
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].size()),
+                [self.batch_size, self.decoder_seq_length, self.vocab_size])
+            self.check_loss_output(result)
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, encoder_input_ids, decoder_input_ids, encoder_attention_mask,
+             decoder_attention_mask, decoder_lm_labels) = config_and_inputs
+            inputs_dict = {'encoder_input_ids': encoder_input_ids,
+                           'decoder_input_ids': decoder_input_ids,
+                           'decoder_attention_mask': decoder_attention_mask,
+                           'encoder_attention_mask': encoder_attention_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = T5ModelTest.T5ModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=T5Config, d_model=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_t5_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_t5_model(*config_and_inputs)
+
+    def test_with_lm_head(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_t5_with_lm_head(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in list(T5_PRETRAINED_MODEL_ARCHIVE_MAP.keys())[:1]:
+            model = T5Model.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()

transformers/tests/modeling_tf_albert_test.py

@@ -118,7 +118,7 @@ class TFAlbertModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = AlbertConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

transformers/tests/modeling_tf_bert_test.py

@@ -114,7 +114,7 @@ class TFBertModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = BertConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

transformers/tests/modeling_tf_common_test.py

@@ -69,6 +69,7 @@ class TFCommonTestCases:
         test_torchscript = True
         test_pruning = True
         test_resize_embeddings = True
+        is_encoder_decoder = False
 
         def test_initialization(self):
             pass
@@ -129,8 +130,12 @@ class TFCommonTestCases:
                                       for name, key in inputs_dict.items())
                 with torch.no_grad():
                     pto = pt_model(**pt_inputs_dict)
-                tfo = tf_model(inputs_dict)
-                max_diff = np.amax(np.abs(tfo[0].numpy() - pto[0].numpy()))
+                tfo = tf_model(inputs_dict, training=False)
+                tf_hidden_states = tfo[0].numpy()
+                pt_hidden_states = pto[0].numpy()
+                tf_hidden_states[np.isnan(tf_hidden_states)] = 0
+                pt_hidden_states[np.isnan(pt_hidden_states)] = 0
+                max_diff = np.amax(np.abs(tf_hidden_states - pt_hidden_states))
                 self.assertLessEqual(max_diff, 2e-2)
 
                 # Check we can load pt model in tf and vice-versa with checkpoint => model functions
@@ -150,13 +155,21 @@ class TFCommonTestCases:
                 with torch.no_grad():
                     pto = pt_model(**pt_inputs_dict)
                 tfo = tf_model(inputs_dict)
-                max_diff = np.amax(np.abs(tfo[0].numpy() - pto[0].numpy()))
+                tfo = tfo[0].numpy()
+                pto = pto[0].numpy()
+                tfo[np.isnan(tfo)] = 0
+                pto[np.isnan(pto)] = 0
+                max_diff = np.amax(np.abs(tfo - pto))
                 self.assertLessEqual(max_diff, 2e-2)
 
         def test_compile_tf_model(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
-            input_ids = tf.keras.Input(batch_shape=(2, 2000), name='input_ids', dtype='int32')
+            if self.is_encoder_decoder:
+                input_ids = {'decoder_input_ids': tf.keras.Input(batch_shape=(2, 2000), name='decoder_input_ids', dtype='int32'),
+                             'encoder_input_ids': tf.keras.Input(batch_shape=(2, 2000), name='encoder_input_ids', dtype='int32')}
+            else:
+                input_ids = tf.keras.Input(batch_shape=(2, 2000), name='input_ids', dtype='int32')
             optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
             loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
             metric = tf.keras.metrics.SparseCategoricalAccuracy('accuracy')
@@ -189,7 +202,7 @@ class TFCommonTestCases:
                 outputs_dict = model(inputs_dict)
 
                 inputs_keywords = copy.deepcopy(inputs_dict)
-                input_ids = inputs_keywords.pop('input_ids')
+                input_ids = inputs_keywords.pop('input_ids' if not self.is_encoder_decoder else 'decoder_input_ids', None)
                 outputs_keywords = model(input_ids, **inputs_keywords)
 
                 output_dict = outputs_dict[0].numpy()
@@ -200,6 +213,11 @@ class TFCommonTestCases:
         def test_attention_outputs(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
+            decoder_seq_length = self.model_tester.decoder_seq_length if hasattr(self.model_tester, 'decoder_seq_length') else self.model_tester.seq_length
+            encoder_seq_length = self.model_tester.encoder_seq_length if hasattr(self.model_tester, 'encoder_seq_length') else self.model_tester.seq_length
+            decoder_key_length = self.model_tester.key_length if hasattr(self.model_tester, 'key_length') else decoder_seq_length
+            encoder_key_length = self.model_tester.key_length if hasattr(self.model_tester, 'key_length') else encoder_seq_length
+
             for model_class in self.all_model_classes:
                 config.output_attentions = True
                 config.output_hidden_states = False
@@ -212,16 +230,28 @@ class TFCommonTestCases:
                 self.assertListEqual(
                     list(attentions[0].shape[-3:]),
                     [self.model_tester.num_attention_heads,
-                    self.model_tester.seq_length,
-                    self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
+                    encoder_seq_length,
+                    encoder_key_length])
                 out_len = len(outputs)
 
+                if self.is_encoder_decoder:
+                    self.assertEqual(out_len % 2, 0)
+                    decoder_attentions = outputs[(out_len // 2)-1]
+                    self.assertEqual(model.config.output_attentions, True)
+                    self.assertEqual(model.config.output_hidden_states, False)
+                    self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
+                    self.assertListEqual(
+                        list(decoder_attentions[0].shape[-3:]),
+                        [self.model_tester.num_attention_heads,
+                         decoder_seq_length,
+                         decoder_key_length])
+
                 # Check attention is always last and order is fine
                 config.output_attentions = True
                 config.output_hidden_states = True
                 model = model_class(config)
                 outputs = model(inputs_dict)
-                self.assertEqual(out_len+1, len(outputs))
+                self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
                 self.assertEqual(model.config.output_attentions, True)
                 self.assertEqual(model.config.output_hidden_states, True)
 
@@ -230,8 +260,8 @@ class TFCommonTestCases:
                 self.assertListEqual(
                     list(attentions[0].shape[-3:]),
                     [self.model_tester.num_attention_heads,
-                    self.model_tester.seq_length,
-                    self.model_tester.key_len if hasattr(self.model_tester, 'key_len') else self.model_tester.seq_length])
+                    encoder_seq_length,
+                    encoder_key_length])
 
         def test_hidden_states_output(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -264,35 +294,53 @@ class TFCommonTestCases:
             for model_class in self.all_model_classes:
                 model = model_class(config)
                 first, second = model(inputs_dict, training=False)[0], model(inputs_dict, training=False)[0]
-                self.assertTrue(tf.math.equal(first, second).numpy().all())
+                out_1 = first.numpy()
+                out_2 = second.numpy()
+                out_1 = out_1[~np.isnan(out_1)]
+                out_2 = out_2[~np.isnan(out_2)]
+                max_diff = np.amax(np.abs(out_1 - out_2))
+                self.assertLessEqual(max_diff, 1e-5)
+
+        def _get_embeds(self, wte, input_ids):
+            # ^^ In our TF models, the input_embeddings can take slightly different forms,
+            # so we try a few of them.
+            # We used to fall back to just synthetically creating a dummy tensor of ones:
+            try:
+                x = wte(input_ids, mode="embedding")
+            except:
+                try:
+                    x = wte([input_ids], mode="embedding")
+                except:
+                    try:
+                        x = wte([input_ids, None, None, None], mode="embedding")
+                    except:
+                        if hasattr(self.model_tester, "embedding_size"):
+                            x = tf.ones(input_ids.shape + [self.model_tester.embedding_size], dtype=tf.dtypes.float32)
+                        else:
+                            x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32)
+            return x
 
         def test_inputs_embeds(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-            input_ids = inputs_dict["input_ids"]
-            del inputs_dict["input_ids"]
+            if not self.is_encoder_decoder:
+                input_ids = inputs_dict["input_ids"]
+                del inputs_dict["input_ids"]
+            else:
+                encoder_input_ids = inputs_dict["encoder_input_ids"]
+                decoder_input_ids = inputs_dict["decoder_input_ids"]
+                del inputs_dict["encoder_input_ids"]
+                del inputs_dict["decoder_input_ids"]
 
             for model_class in self.all_model_classes:
                 model = model_class(config)
 
                 wte = model.get_input_embeddings()
-                try:
-                    x = wte(input_ids, mode="embedding")
-                except:
-                    try:
-                        x = wte([input_ids], mode="embedding")
-                    except:
-                        try:
-                            x = wte([input_ids, None, None, None], mode="embedding")
-                        except:
-                            if hasattr(self.model_tester, "embedding_size"):
-                                x = tf.ones(input_ids.shape + [self.model_tester.embedding_size], dtype=tf.dtypes.float32)
-                            else:
-                                x = tf.ones(input_ids.shape + [self.model_tester.hidden_size], dtype=tf.dtypes.float32)
-                # ^^ In our TF models, the input_embeddings can take slightly different forms,
-                # so we try a few of them.
-                # We used to fall back to just synthetically creating a dummy tensor of ones:
-                #
-                inputs_dict["inputs_embeds"] = x
+                if not self.is_encoder_decoder:
+                    inputs_dict["inputs_embeds"] = self._get_embeds(wte, input_ids)
+                else:
+                    inputs_dict["encoder_inputs_embeds"] = self._get_embeds(wte, encoder_input_ids)
+                    inputs_dict["decoder_inputs_embeds"] = self._get_embeds(wte, decoder_input_ids)
+
                 outputs = model(inputs_dict)
 
 

transformers/tests/modeling_tf_ctrl_test.py

@@ -112,7 +112,7 @@ class TFCTRLModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = CTRLConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
                 n_head=self.num_attention_heads,

transformers/tests/modeling_tf_distilbert_test.py

@@ -107,7 +107,7 @@ class TFDistilBertModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = DistilBertConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 dim=self.hidden_size,
                 n_layers=self.num_hidden_layers,
                 n_heads=self.num_attention_heads,

transformers/tests/modeling_tf_gpt2_test.py

@@ -115,7 +115,7 @@ class TFGPT2ModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = GPT2Config(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
                 n_head=self.num_attention_heads,

transformers/tests/modeling_tf_openai_gpt_test.py

@@ -114,7 +114,7 @@ class TFOpenAIGPTModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = OpenAIGPTConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 n_embd=self.hidden_size,
                 n_layer=self.num_hidden_layers,
                 n_head=self.num_attention_heads,

transformers/tests/modeling_tf_roberta_test.py

@@ -109,7 +109,7 @@ class TFRobertaModelTest(TFCommonTestCases.TFCommonModelTester):
                 choice_labels = ids_tensor([self.batch_size], self.num_choices)
 
             config = RobertaConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 hidden_size=self.hidden_size,
                 num_hidden_layers=self.num_hidden_layers,
                 num_attention_heads=self.num_attention_heads,

transformers/tests/modeling_tf_t5_test.py

@@ -0,0 +1,172 @@
+# coding=utf-8
+# Copyright 2018 Google 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.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import unittest
+import shutil
+import sys
+
+from .modeling_tf_common_test import (TFCommonTestCases, ids_tensor)
+from .configuration_common_test import ConfigTester
+from .utils import require_tf, slow
+
+from transformers import T5Config, is_tf_available
+
+if is_tf_available():
+    import tensorflow as tf
+    from transformers.modeling_tf_t5 import (TFT5Model, TFT5WithLMHeadModel,
+                                             TF_T5_PRETRAINED_MODEL_ARCHIVE_MAP)
+
+
+@require_tf
+class TFT5ModelTest(TFCommonTestCases.TFCommonModelTester):
+
+    is_encoder_decoder = True
+    all_model_classes = (TFT5Model, TFT5WithLMHeadModel) if is_tf_available() else ()
+
+    class TFT5ModelTester(object):
+
+        def __init__(self,
+                     parent,
+                     batch_size=13,
+                     seq_length=7,
+                     is_training=True,
+                     use_input_mask=True,
+                     use_labels=True,
+                     vocab_size=99,
+                     n_positions=14,
+                     hidden_size=32,
+                     num_hidden_layers=5,
+                     num_attention_heads=4,
+                     d_ff=37,
+                     relative_attention_num_buckets=8,
+                     dropout_rate=0.1,
+                     initializer_factor=0.002,
+                     scope=None,
+                    ):
+            self.parent = parent
+            self.batch_size = batch_size
+            self.seq_length = seq_length
+            self.is_training = is_training
+            self.use_input_mask = use_input_mask
+            self.use_labels = use_labels
+            self.vocab_size = vocab_size
+            self.n_positions = n_positions
+            self.hidden_size = hidden_size
+            self.num_hidden_layers = num_hidden_layers
+            self.num_attention_heads = num_attention_heads
+            self.d_ff = d_ff
+            self.relative_attention_num_buckets = relative_attention_num_buckets
+            self.dropout_rate = dropout_rate
+            self.initializer_factor = initializer_factor
+            self.scope = scope
+
+        def prepare_config_and_inputs(self):
+            input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            input_mask = None
+            if self.use_input_mask:
+                input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            token_labels = None
+            if self.use_labels:
+                token_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+            config = T5Config(
+                vocab_size=self.vocab_size,
+                n_positions=self.n_positions,
+                d_model=self.hidden_size,
+                d_ff=self.d_ff,
+                d_kv=self.hidden_size // self.num_attention_heads,
+                num_layers=self.num_hidden_layers,
+                num_heads=self.num_attention_heads,
+                relative_attention_num_buckets=self.relative_attention_num_buckets,
+                dropout_rate=self.dropout_rate,
+                initializer_factor=self.initializer_factor)
+
+            return (config, input_ids, input_mask, token_labels)
+
+        def create_and_check_t5_model(self, config, input_ids, input_mask, token_labels):
+            model = TFT5Model(config=config)
+            inputs = {'encoder_input_ids': input_ids,
+                      'decoder_input_ids': input_ids,
+                      'decoder_attention_mask': input_mask}
+            encoder_output, decoder_output = model(inputs)
+
+            encoder_output, decoder_output = model(input_ids,
+                                                   decoder_attention_mask=input_mask,
+                                                   encoder_input_ids=input_ids)
+
+            result = {
+                "encoder_output": encoder_output.numpy(),
+                "decoder_output": decoder_output.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["encoder_output"].shape),
+                [self.batch_size, self.seq_length, self.hidden_size])
+            self.parent.assertListEqual(
+                list(result["decoder_output"].shape),
+                [self.batch_size, self.seq_length, self.hidden_size])
+
+
+        def create_and_check_t5_with_lm_head(self, config, input_ids, input_mask, token_labels):
+            model = TFT5WithLMHeadModel(config=config)
+            inputs = {'encoder_input_ids': input_ids,
+                      'decoder_input_ids': input_ids,
+                      'decoder_attention_mask': input_mask}
+            prediction_scores, decoder_output = model(inputs)
+            result = {
+                "prediction_scores": prediction_scores.numpy(),
+            }
+            self.parent.assertListEqual(
+                list(result["prediction_scores"].shape),
+                [self.batch_size, self.seq_length, self.vocab_size])
+
+
+        def prepare_config_and_inputs_for_common(self):
+            config_and_inputs = self.prepare_config_and_inputs()
+            (config, input_ids, input_mask, token_labels) = config_and_inputs
+            inputs_dict = {'encoder_input_ids': input_ids,
+                           'decoder_input_ids': input_ids,
+                           'decoder_attention_mask': input_mask}
+            return config, inputs_dict
+
+    def setUp(self):
+        self.model_tester = TFT5ModelTest.TFT5ModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=T5Config, d_model=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_t5_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_t5_model(*config_and_inputs)
+
+    def test_with_lm_head(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_t5_with_lm_head(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        cache_dir = "/tmp/transformers_test/"
+        for model_name in ['t5-small']:
+            model = TFT5Model.from_pretrained(model_name, cache_dir=cache_dir)
+            shutil.rmtree(cache_dir)
+            self.assertIsNotNone(model)
+
+if __name__ == "__main__":
+    unittest.main()

transformers/tests/modeling_tf_transfo_xl_test.py

@@ -67,7 +67,7 @@ class TFTransfoXLModelTest(TFCommonTestCases.TFCommonModelTester):
             self.batch_size = batch_size
             self.seq_length = seq_length
             self.mem_len = mem_len
-            self.key_len = seq_length + mem_len
+            self.key_length = seq_length + mem_len
             self.clamp_len = clamp_len
             self.is_training = is_training
             self.use_labels = use_labels
@@ -92,7 +92,7 @@ class TFTransfoXLModelTest(TFCommonTestCases.TFCommonModelTester):
                 lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
             config = TransfoXLConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 cutoffs=self.cutoffs,

transformers/tests/modeling_tf_xlm_test.py

@@ -125,7 +125,7 @@ class TFXLMModelTest(TFCommonTestCases.TFCommonModelTester):
                 is_impossible_labels = ids_tensor([self.batch_size], 2, dtype=tf.float32)
 
             config = XLMConfig(
-                 vocab_size_or_config_json_file=self.vocab_size,
+                 vocab_size=self.vocab_size,
                  n_special=self.n_special,
                  emb_dim=self.hidden_size,
                  n_layers=self.num_hidden_layers,

transformers/tests/modeling_tf_xlnet_test.py

@@ -64,7 +64,6 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                      num_attention_heads=4,
                      d_inner=128,
                      num_hidden_layers=5,
-                     max_position_embeddings=10,
                      type_sequence_label_size=2,
                      untie_r=True,
                      bi_data=False,
@@ -88,7 +87,6 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
             self.num_attention_heads = num_attention_heads
             self.d_inner = d_inner
             self.num_hidden_layers = num_hidden_layers
-            self.max_position_embeddings = max_position_embeddings
             self.bi_data = bi_data
             self.untie_r = untie_r
             self.same_length = same_length
@@ -122,13 +120,12 @@ class TFXLNetModelTest(TFCommonTestCases.TFCommonModelTester):
                 is_impossible_labels = ids_tensor([self.batch_size], 2, dtype=tf.float32)
 
             config = XLNetConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 d_model=self.hidden_size,
                 n_head=self.num_attention_heads,
                 d_inner=self.d_inner,
                 n_layer=self.num_hidden_layers,
                 untie_r=self.untie_r,
-                max_position_embeddings=self.max_position_embeddings,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 same_length=self.same_length,

transformers/tests/modeling_transfo_xl_test.py

@@ -66,7 +66,7 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
             self.batch_size = batch_size
             self.seq_length = seq_length
             self.mem_len = mem_len
-            self.key_len = seq_length + mem_len
+            self.key_length = seq_length + mem_len
             self.clamp_len = clamp_len
             self.is_training = is_training
             self.use_labels = use_labels
@@ -91,7 +91,7 @@ class TransfoXLModelTest(CommonTestCases.CommonModelTester):
                 lm_labels = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
             config = TransfoXLConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 cutoffs=self.cutoffs,

transformers/tests/modeling_xlm_test.py

@@ -121,7 +121,7 @@ class XLMModelTest(CommonTestCases.CommonModelTester):
                 is_impossible_labels = ids_tensor([self.batch_size], 2).float()
 
             config = XLMConfig(
-                 vocab_size_or_config_json_file=self.vocab_size,
+                 vocab_size=self.vocab_size,
                  n_special=self.n_special,
                  emb_dim=self.hidden_size,
                  n_layers=self.num_hidden_layers,

transformers/tests/modeling_xlnet_test.py

@@ -60,7 +60,6 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                      num_attention_heads=4,
                      d_inner=128,
                      num_hidden_layers=5,
-                     max_position_embeddings=10,
                      type_sequence_label_size=2,
                      untie_r=True,
                      bi_data=False,
@@ -84,7 +83,6 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
             self.num_attention_heads = num_attention_heads
             self.d_inner = d_inner
             self.num_hidden_layers = num_hidden_layers
-            self.max_position_embeddings = max_position_embeddings
             self.bi_data = bi_data
             self.untie_r = untie_r
             self.same_length = same_length
@@ -116,13 +114,12 @@ class XLNetModelTest(CommonTestCases.CommonModelTester):
                 token_labels = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
 
             config = XLNetConfig(
-                vocab_size_or_config_json_file=self.vocab_size,
+                vocab_size=self.vocab_size,
                 d_model=self.hidden_size,
                 n_head=self.num_attention_heads,
                 d_inner=self.d_inner,
                 n_layer=self.num_hidden_layers,
                 untie_r=self.untie_r,
-                max_position_embeddings=self.max_position_embeddings,
                 mem_len=self.mem_len,
                 clamp_len=self.clamp_len,
                 same_length=self.same_length,

transformers/tests/tokenization_t5_test.py

@@ -0,0 +1,77 @@
+# coding=utf-8
+# Copyright 2018 Google 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.
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import os
+import unittest
+
+from transformers.tokenization_t5 import (T5Tokenizer)
+from transformers.tokenization_xlnet import SPIECE_UNDERLINE
+
+from .tokenization_tests_commons import CommonTestCases
+
+SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)),
+                    'fixtures/test_sentencepiece.model')
+
+class T5TokenizationTest(CommonTestCases.CommonTokenizerTester):
+
+    tokenizer_class = T5Tokenizer
+
+    def setUp(self):
+        super(T5TokenizationTest, self).setUp()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = T5Tokenizer(SAMPLE_VOCAB)
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return T5Tokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_input_output_texts(self):
+        input_text = u"This is a test"
+        output_text = u"This is a test"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = T5Tokenizer(SAMPLE_VOCAB)
+
+        tokens = tokenizer.tokenize(u'This is a test')
+        self.assertListEqual(tokens, [u'▁This', u'▁is', u'▁a', u'▁t', u'est'])
+
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(tokens), [285, 46, 10, 170, 382])
+
+        tokens = tokenizer.tokenize(u"I was born in 92000, and this is falsé.")
+        self.assertListEqual(tokens, [SPIECE_UNDERLINE + u'I', SPIECE_UNDERLINE + u'was', SPIECE_UNDERLINE + u'b',
+                                    u'or', u'n', SPIECE_UNDERLINE + u'in', SPIECE_UNDERLINE + u'',
+                                    u'9', u'2', u'0', u'0', u'0', u',', SPIECE_UNDERLINE + u'and', SPIECE_UNDERLINE + u'this',
+                                    SPIECE_UNDERLINE + u'is', SPIECE_UNDERLINE + u'f', u'al', u's', u'é', u'.'])
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(
+            ids, [8, 21, 84, 55, 24, 19, 7, 0,
+                602, 347, 347, 347, 3, 12, 66,
+                46, 72, 80, 6, 0, 4])
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(back_tokens, [SPIECE_UNDERLINE + u'I', SPIECE_UNDERLINE + u'was', SPIECE_UNDERLINE + u'b',
+                                        u'or', u'n', SPIECE_UNDERLINE + u'in',
+                                        SPIECE_UNDERLINE + u'', u'<unk>', u'2', u'0', u'0', u'0', u',',
+                                        SPIECE_UNDERLINE + u'and', SPIECE_UNDERLINE + u'this',
+                                        SPIECE_UNDERLINE + u'is', SPIECE_UNDERLINE + u'f', u'al', u's',
+                                        u'<unk>', u'.'])
+
+
+if __name__ == '__main__':
+    unittest.main()

transformers/tokenization_auto.py

@@ -30,6 +30,7 @@ from .tokenization_roberta import RobertaTokenizer
 from .tokenization_distilbert import DistilBertTokenizer
 from .tokenization_camembert import CamembertTokenizer
 from .tokenization_albert import AlbertTokenizer
+from .tokenization_t5 import T5Tokenizer
 
 logger = logging.getLogger(__name__)
 
@@ -44,6 +45,7 @@ class AutoTokenizer(object):
 
         The tokenizer class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Tokenizer (T5 model)
             - contains `distilbert`: DistilBertTokenizer (DistilBert model)
             - contains `albert`: AlbertTokenizer (ALBERT model)
             - contains `camembert`: CamembertTokenizer (CamemBERT model)
@@ -69,6 +71,7 @@ class AutoTokenizer(object):
 
         The tokenizer class to instantiate is selected as the first pattern matching
         in the `pretrained_model_name_or_path` string (in the following order):
+            - contains `t5`: T5Tokenizer (T5 model)
             - contains `distilbert`: DistilBertTokenizer (DistilBert model)
             - contains `albert`: AlbertTokenizer (ALBERT model)
             - contains `camembert`: CamembertTokenizer (CamemBERT model)
@@ -119,7 +122,9 @@ class AutoTokenizer(object):
             tokenizer = AutoTokenizer.from_pretrained('./test/bert_saved_model/')
 
         """
-        if 'distilbert' in pretrained_model_name_or_path:
+        if 't5' in pretrained_model_name_or_path:
+            return T5Tokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+        elif 'distilbert' in pretrained_model_name_or_path:
             return DistilBertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
         elif 'albert' in pretrained_model_name_or_path:
             return AlbertTokenizer.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)

transformers/tokenization_bert.py

@@ -46,6 +46,8 @@ PRETRAINED_VOCAB_FILES_MAP = {
         'bert-base-cased-finetuned-mrpc': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-finetuned-mrpc-vocab.txt",
         'bert-base-german-dbmdz-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-cased-vocab.txt",
         'bert-base-german-dbmdz-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-german-dbmdz-uncased-vocab.txt",
+        'bert-base-finnish-cased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-cased-v1/vocab.txt",
+        'bert-base-finnish-uncased-v1': "https://s3.amazonaws.com/models.huggingface.co/bert/TurkuNLP/bert-base-finnish-uncased-v1/vocab.txt",
     }
 }
 
@@ -65,6 +67,8 @@ PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     'bert-base-cased-finetuned-mrpc': 512,
     'bert-base-german-dbmdz-cased': 512,
     'bert-base-german-dbmdz-uncased': 512,
+    'bert-base-finnish-cased-v1': 512,
+    'bert-base-finnish-uncased-v1': 512,
 }
 
 PRETRAINED_INIT_CONFIGURATION = {
@@ -83,6 +87,8 @@ PRETRAINED_INIT_CONFIGURATION = {
     'bert-base-cased-finetuned-mrpc': {'do_lower_case': False},
     'bert-base-german-dbmdz-cased': {'do_lower_case': False},
     'bert-base-german-dbmdz-uncased': {'do_lower_case': True},
+    'bert-base-finnish-cased-v1': {'do_lower_case': False},
+    'bert-base-finnish-uncased-v1': {'do_lower_case': True},
 }
 
 
@@ -113,12 +119,12 @@ class BertTokenizer(PreTrainedTokenizer):
 
     Args:
         vocab_file: Path to a one-wordpiece-per-line vocabulary file
-        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
         do_basic_tokenize: Whether to do basic tokenization before wordpiece.
         max_len: An artificial maximum length to truncate tokenized sequences to; Effective maximum length is always the
             minimum of this value (if specified) and the underlying BERT model's sequence length.
         never_split: List of tokens which will never be split during tokenization. Only has an effect when
-            do_wordpiece_only=False
+            do_basic_tokenize=True
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

transformers/tokenization_distilbert.py

@@ -53,12 +53,12 @@ class DistilBertTokenizer(BertTokenizer):
 
     Args:
         vocab_file: Path to a one-wordpiece-per-line vocabulary file
-        do_lower_case: Whether to lower case the input. Only has an effect when do_wordpiece_only=False
+        do_lower_case: Whether to lower case the input. Only has an effect when do_basic_tokenize=True
         do_basic_tokenize: Whether to do basic tokenization before wordpiece.
         max_len: An artificial maximum length to truncate tokenized sequences to; Effective maximum length is always the
             minimum of this value (if specified) and the underlying BERT model's sequence length.
         never_split: List of tokens which will never be split during tokenization. Only has an effect when
-            do_wordpiece_only=False
+            do_basic_tokenize=True
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

transformers/tokenization_t5.py

@@ -0,0 +1,176 @@
+# coding=utf-8
+# Copyright 2018 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.
+""" Tokenization class for model T5."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import logging
+import os
+import re
+import six
+from shutil import copyfile
+
+from .tokenization_utils import PreTrainedTokenizer
+
+logger = logging.getLogger(__name__)
+
+SPIECE_UNDERLINE = u'▁'
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to file names for serializing Tokenizer instances
+####################################################
+VOCAB_FILES_NAMES = {'vocab_file': 'spiece.model'}
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to pretrained vocabulary URL for all the model shortcut names.
+####################################################
+PRETRAINED_VOCAB_FILES_MAP = {
+    'vocab_file':
+    {
+        't5-small': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-base': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-large': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-3b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+        't5-11b': "https://s3.amazonaws.com/models.huggingface.co/bert/t5-spiece.model",
+    }
+}
+
+####################################################
+# Mapping from model shortcut names to max length of inputs
+####################################################
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    't5-small': 512,
+    't5-base': 512,
+    't5-large': 512,
+    't5-3b': 512,
+    't5-11b': 512,
+}
+
+class T5Tokenizer(PreTrainedTokenizer):
+    """
+        SentencePiece based tokenizer. Peculiarities:
+
+            - requires `SentencePiece <https://github.com/google/sentencepiece>`_
+            - `extra_ids` add a number of extra ids added to the end of the vocabulary for use as sentinels.
+                These tokens are accessible as `<extra_id_{%d}>` where `{%d}` is a number between 0 and extra_ids-1.
+                Extra tokens are indexed from the end of the vocabulary up to beginnning (<extra_id_0> is the last token in the vocabulary)
+                (like in T5 preprocessing
+                see: https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117)
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(self, vocab_file, eos_token="</s>", unk_token="<unk>",
+                 pad_token="<pad>", extra_ids=100, additional_special_tokens=None, **kwargs):
+        # Add extra_ids to the special token list
+        if extra_ids > 0:
+            if additional_special_tokens is None:
+                additional_special_tokens = []
+            additional_special_tokens.extend([u"<extra_id_{}>".format(i) for i in range(extra_ids)])
+
+        super(T5Tokenizer, self).__init__(eos_token=eos_token, unk_token=unk_token,
+                                          pad_token=pad_token, additional_special_tokens=additional_special_tokens,
+                                          **kwargs)
+
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning("You need to install SentencePiece to use T5Tokenizer:"
+                           "https://github.com/google/sentencepiece"
+                           "pip install sentencepiece")
+
+        self.vocab_file = vocab_file
+        self._extra_ids = extra_ids
+
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(vocab_file)
+
+    @property
+    def vocab_size(self):
+        return self.sp_model.get_piece_size() + self._extra_ids
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        try:
+            import sentencepiece as spm
+        except ImportError:
+            logger.warning("You need to install SentencePiece to use XLNetTokenizer: https://github.com/google/sentencepiece"
+                           "pip install sentencepiece")
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(self.vocab_file)
+
+    def _tokenize(self, text, return_unicode=True, sample=False):
+        """ Take as input a string and return a list of strings (tokens) for words/sub-words
+        """
+        if not sample:
+            pieces = self.sp_model.EncodeAsPieces(text)
+        else:
+            pieces = self.sp_model.SampleEncodeAsPieces(text, 64, 0.1)
+
+        # convert back to unicode for py2
+        if six.PY2 and return_unicode:
+            ret_pieces = []
+            for piece in pieces:
+                if isinstance(piece, str):
+                    piece = piece.decode('utf-8')
+                ret_pieces.append(piece)
+            pieces = ret_pieces
+
+        return pieces
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str/unicode) in an id using the vocab. """
+        if token.startswith(u"<extra_id_"):
+            l = re.match(r'<extra_id_(\d+)>', token)
+            num = int(l.group(1))
+            return self.vocab_size - num - 1
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index, return_unicode=True):
+        """Converts an index (integer) in a token (string/unicode) using the vocab."""
+        if index < self.sp_model.get_piece_size():
+            token = self.sp_model.IdToPiece(index)
+        else:
+            token = u"<extra_id_{}>".format(self.vocab_size - 1 - index)
+        if six.PY2 and return_unicode and isinstance(token, str):
+            token = token.decode('utf-8')
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """ Converts a sequence of tokens (string) in a single string. """
+        out_string = self.sp_model.decode_pieces(tokens)
+        return out_string
+
+    def save_vocabulary(self, save_directory):
+        """ Save the sentencepiece vocabulary (copy original file) and special tokens file
+            to a directory.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error("Vocabulary path ({}) should be a directory".format(save_directory))
+            return
+        out_vocab_file = os.path.join(save_directory, VOCAB_FILES_NAMES['vocab_file'])
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)