Pass a few more tests

README.md

@@ -188,6 +188,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 ultilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
 1. **[SqueezeBert](https://huggingface.co/transformers/model_doc/squeezebert.html)** released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
 1. **[T5](https://huggingface.co/transformers/model_doc/t5.html)** (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.
+1. **[TAPAS](https://huggingface.co/transformers/master/model_doc/tapas.html)** released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[Transformer-XL](https://huggingface.co/transformers/model_doc/transformerxl.html)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[XLM](https://huggingface.co/transformers/model_doc/xlm.html)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/transformers/model_doc/xlmprophetnet.html)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
@@ -222,4 +223,4 @@ We now have a [paper](https://arxiv.org/abs/1910.03771) you can cite for the
   year={2019},
   volume={abs/1910.03771}
 }
-```
+```

docs/source/index.rst

@@ -145,22 +145,25 @@ conversion utilities for the following models:
 27. :doc:`T5 <model_doc/t5>` (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.
-28. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+28. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
+    Pre-training <https://arxiv.org/abs/2004.02349>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller,
+    Francesco Piccinno and Julian Martin Eisenschlos.
+29. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
     Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*,
     Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-29. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+30. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
     Pretraining <https://arxiv.org/abs/1901.07291>`__ by Guillaume Lample and Alexis Conneau.
-30. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+31. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
     Predicting Future N-gram for Sequence-to-Sequence Pre-training <https://arxiv.org/abs/2001.04063>`__ by Yu Yan,
     Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-31. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+32. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
     Cross-lingual Representation Learning at Scale <https://arxiv.org/abs/1911.02116>`__ by Alexis Conneau*, Kartikay
     Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke
     Zettlemoyer and Veselin Stoyanov.
-32. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
+33. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
     Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`__ by Zhilin Yang*, Zihang Dai*, Yiming
     Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
-33. `Other community models <https://huggingface.co/models>`__, contributed by the `community
+34. `Other community models <https://huggingface.co/models>`__, contributed by the `community
     <https://huggingface.co/users>`__.
 
 .. toctree::
@@ -258,6 +261,7 @@ conversion utilities for the following models:
     model_doc/roberta
     model_doc/squeezebert
     model_doc/t5
+    model_doc/tapas
     model_doc/transformerxl
     model_doc/xlm
     model_doc/xlmprophetnet

docs/source/model_doc/tapas.rst

@@ -0,0 +1,162 @@
+TAPAS
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The TAPAS model was proposed in `TAPAS: Weakly Supervised Table Parsing via Pre-training
+<https://arxiv.org/abs/2004.02349>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and
+Julian Martin Eisenschlos. It's a BERT-based model specifically designed (and pre-trained) for answering questions
+about tabular data. Compared to BERT, TAPAS uses relative position embeddings and has 7 token types that encode tabular
+structure. TAPAS is pre-trained on the masked language modeling (MLM) objective on a large dataset comprising millions
+of tables from English Wikipedia and corresponding texts. For question answering, TAPAS has 2 heads on top: a cell
+selection head and an aggregation head, for (optionally) performing aggregations (such as counting or summing) among
+selected cells. TAPAS has been fine-tuned on several datasets: SQA (Sequential Question Answering by Microsoft), WTQ
+(Wiki Table Questions by Stanford University) and WikiSQL (by Salesforce). It achieves state-of-the-art on both SQA and
+WTQ, while having comparable performance to SOTA on WikiSQL, with a much simpler architecture.
+
+The abstract from the paper is the following:
+
+*Answering natural language questions over tables is usually seen as a semantic parsing task. To alleviate the
+collection cost of full logical forms, one popular approach focuses on weak supervision consisting of denotations
+instead of logical forms. However, training semantic parsers from weak supervision poses difficulties, and in addition,
+the generated logical forms are only used as an intermediate step prior to retrieving the denotation. In this paper, we
+present TAPAS, an approach to question answering over tables without generating logical forms. TAPAS trains from weak
+supervision, and predicts the denotation by selecting table cells and optionally applying a corresponding aggregation
+operator to such selection. TAPAS extends BERT's architecture to encode tables as input, initializes from an effective
+joint pre-training of text segments and tables crawled from Wikipedia, and is trained end-to-end. We experiment with
+three different semantic parsing datasets, and find that TAPAS outperforms or rivals semantic parsing models by
+improving state-of-the-art accuracy on SQA from 55.1 to 67.2 and performing on par with the state-of-the-art on WIKISQL
+and WIKITQ, but with a simpler model architecture. We additionally find that transfer learning, which is trivial in our
+setting, from WIKISQL to WIKITQ, yields 48.7 accuracy, 4.2 points above the state-of-the-art.*
+
+In addition, the authors have further pre-trained TAPAS to recognize table entailment, by creating a balanced dataset
+of millions of automatically created training examples which are learned in an intermediate step prior to fine-tuning.
+The authors of TAPAS call this further pre-training intermediate pre-training (since TAPAS is first pre-trained on MLM,
+and then on another dataset). They found that intermediate pre-training further improves performance on SQA, achieving
+a new state-of-the-art as well as state-of-the-art on TabFact, a large-scale dataset with 16k Wikipedia tables for
+table entailment (a binary classification task). For more details, see their new paper: `Understanding tables with
+intermediate pre-training <https://arxiv.org/abs/2010.00571>`__ by Julian Martin Eisenschlos, Syrine Krichene and
+Thomas Müller.
+
+The original code can be found `here <https://github.com/google-research/tapas>`__.
+
+Tips:
+
+- TAPAS is a model that uses relative position embeddings by default (restarting the position embeddings at every cell
+  of the table). According to the authors, this usually results in a slightly better performance, and allows you to
+  encode longer sequences without running out of embeddings. If you don't want this, you can set the
+  `reset_position_index_per_cell` parameter of :class:`~transformers.TapasConfig` to False.
+- TAPAS has checkpoints fine-tuned on SQA, which are capable of answering questions related to a table in a
+  conversational set-up. This means that you can ask follow-up questions such as "what is his age?" related to the
+  previous question. Note that the forward pass of TAPAS is a bit different in case of a conversational set-up: in that
+  case, you have to feed every training example one by one to the model, such that the `prev_label_ids` token type ids
+  can be overwritten by the predicted `label_ids` of the model to the previous question.
+- TAPAS is similar to BERT and therefore relies on the masked language modeling (MLM) objective. It is therefore
+  efficient at predicting masked tokens and at NLU in general, but is not optimal for text generation. Models trained
+  with a causal language modeling (CLM) objective are better in that regard.
+
+
+Usage
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+If you just want to perform inference (i.e. making predictions) in a non-conversational setup, you can do the
+following:
+
+.. code-block::
+
+        >>> from transformers import TapasTokenizer, TapasForQuestionAnswering
+        >>> import pandas as pd 
+
+        >>> model_name = 'tapas-base-finetuned-wtq'
+        >>> model = TapasForQuestionAnswering.from_pretrained(model_name)
+        >>> tokenizer = TapasTokenizer.from_pretrained(model_name)
+
+        >>> data = {'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"], 'Number of movies': ["87", "53", "69"]}
+        >>> queries = ["What is the name of the first actor?", "How many movies has George Clooney played in?", "What is the total number of movies?"]
+        >>> table = pd.Dataframe(data)
+        >>> inputs = tokenizer(table, queries, return_tensors='pt')
+        >>> logits, logits_agg = model(**inputs)
+        >>> answer_coordinates_batch, aggregation_predictions = tokenizer.convert_logits_to_predictions(inputs, logits, logits_agg)
+
+        >>> # let's print out the results:
+        >>> id2aggregation = {0: "NONE", 1: "SUM", 2: "AVERAGE", 3:"COUNT"}
+        >>> aggregation_predictions_string = [id2aggregation[x] for x in aggregation_predictions]
+
+        >>> answers = []
+        >>> for coordinates in answer_coordinates_batch:
+        ...   if len(coordinates) == 1:
+        ...     # only a single cell:
+        ...     answers.append(df.iat[coordinates[0]])
+        ...   else:
+        ...     # multiple cells
+        ...     cell_values = []
+        ...     for coordinate in coordinates:
+        ...        cell_values.append(df.iat[coordinate])
+        ...     answers.append(", ".join(cell_values))
+
+        >>> display(df)
+        >>> print("")
+        >>> for query, answer, predicted_agg in zip(queries, answers, aggregation_predictions_string):
+        ...   print(query)
+        ...   if predicted_agg == "NONE":
+        ...     print("Predicted answer: " + answer)
+        ...   else:
+        ...     print("Predicted answer: " + predicted_agg + " > " + answer)    
+        When was Brad Pitt born?
+        Predicted answer: 18 december 1963
+        Which actor appeared in the least number of movies?
+        Predicted answer: Leonardo Di Caprio
+        What is the average number of movies?
+        Predicted answer: AVERAGE > 87, 53, 69
+
+
+Tapas specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_tapas.TableQuestionAnsweringOutput
+    :members:
+
+
+TapasConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasConfig
+    :members:
+
+
+TapasTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasTokenizer
+    :members: convert_logits_to_predictions, save_vocabulary
+
+
+TapasModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasModel
+    :members:
+
+
+TapasForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForMaskedLM
+    :members:
+
+
+TapasForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForSequenceClassification
+    :members: forward
+
+
+TapasForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TapasForQuestionAnswering
+    :members:
+
+

src/transformers/__init__.py

@@ -61,6 +61,7 @@ from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, Re
 from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
 from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
 from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
+from .configuration_tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
 from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
@@ -190,6 +191,7 @@ from .tokenization_rag import RagTokenizer
 from .tokenization_retribert import RetriBertTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_squeezebert import SqueezeBertTokenizer
+from .tokenization_tapas import TapasTokenizer
 from .tokenization_transfo_xl import TransfoXLCorpus, TransfoXLTokenizer
 from .tokenization_utils import PreTrainedTokenizer
 from .tokenization_utils_base import (
@@ -558,6 +560,14 @@ if is_torch_available():
         T5PreTrainedModel,
         load_tf_weights_in_t5,
     )
+    from .modeling_tapas import (
+        TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TapasForMaskedLM,
+        TapasForQuestionAnswering,
+        TapasForSequenceClassification,
+        TapasModel,
+        load_tf_weights_in_tapas,
+    )
     from .modeling_transfo_xl import (
         TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_LIST,
         AdaptiveEmbedding,

src/transformers/commands/convert.py

@@ -130,6 +130,13 @@ class ConvertCommand(BaseTransformersCLICommand):
                 raise ImportError(IMPORT_ERROR_MESSAGE)
 
             convert_gpt2_checkpoint_to_pytorch(self._tf_checkpoint, self._config, self._pytorch_dump_output)
+        elif self._model_type == "tapas":
+            try:
+                from transformers.convert_tapas_original_tf_checkpoint_to_pytorch import (
+                    convert_tf_checkpoint_to_pytorch,
+                )
+            except ImportError:
+                raise ImportError(IMPORT_ERROR_MESSAGE)
         elif self._model_type == "xlnet":
             try:
                 from transformers.convert_xlnet_original_tf_checkpoint_to_pytorch import (

src/transformers/configuration_auto.py

@@ -48,6 +48,7 @@ from .configuration_retribert import RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, Re
 from .configuration_roberta import ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, RobertaConfig
 from .configuration_squeezebert import SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, SqueezeBertConfig
 from .configuration_t5 import T5_PRETRAINED_CONFIG_ARCHIVE_MAP, T5Config
+from .configuration_tapas import TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP, TapasConfig
 from .configuration_transfo_xl import TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP, TransfoXLConfig
 from .configuration_utils import PretrainedConfig
 from .configuration_xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig
@@ -88,6 +89,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
         PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -131,6 +133,7 @@ CONFIG_MAPPING = OrderedDict(
         ("dpr", DPRConfig),
         ("layoutlm", LayoutLMConfig),
         ("rag", RagConfig),
+        ("tapas", TapasConfig),
     ]
 )
 
@@ -172,6 +175,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("rag", "RAG"),
         ("xlm-prophetnet", "XLMProphetNet"),
         ("prophetnet", "ProphetNet"),
+        ("tapas", "TAPAS"),
     ]
 )
 

src/transformers/configuration_tapas.py

@@ -0,0 +1,209 @@
+# coding=utf-8
+# Copyright 2020 Google Research and 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.
+""" TAPAS configuration. Adds additional hyperparameters to the configuration of BERT."""
+
+
+from .configuration_utils import PretrainedConfig
+
+
+TAPAS_PRETRAINED_CONFIG_ARCHIVE_MAP = {"tapas-base": "", "tapas-large": ""}  # to be added  # to be added
+
+
+class TapasConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.TapasModel`. It is used to
+    instantiate a TAPAS model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the TAPAS `tapas-base-finetuned-sqa`
+    architecture. Configuration objects inherit from :class:`~transformers.PreTrainedConfig` and can be used to control
+    the model outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+    Hyperparameters additional to BERT are taken from run_task_main.py and hparam_utils.py of the original
+    implementation. Original implementation available at https://github.com/google-research/tapas/tree/master.
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 30522):
+            Vocabulary size of the TAPAS model. Defines the number of different tokens that can be represented by the
+            :obj:`inputs_ids` passed when calling :class:`~transformers.TapasModel`.
+        hidden_size (:obj:`int`, `optional`, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, `optional`, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"swish"` and :obj:`"gelu_new"` are supported.
+        hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`int`, `optional`, defaults to 1024):
+            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_sizes (:obj:`List[int]`, `optional`, defaults to [3, 256, 256, 2, 256, 256, 10]):
+            The vocabulary sizes of the :obj:`token_type_ids` passed when calling :class:`~transformers.TapasModel`.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        positive_label_weight (:obj:`float`, `optional`, defaults to 10.0):
+            Weight for positive labels.
+        num_aggregation_labels (:obj:`int`, `optional`, defaults to 0):
+            The number of aggregation operators to predict.
+        aggregation_loss_weight (:obj:`float`, `optional`, defaults to 1.0):
+            Importance weight for the aggregation loss.
+        use_answer_as_supervision (:obj:`bool`, `optional`, defaults to :obj:`None`):
+            Whether to use the answer as the only supervision for aggregation examples.
+        answer_loss_importance (:obj:`float`, `optional`, defaults to 1.0):
+            Importance weight for the regression loss.
+        use_normalized_answer_loss (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Normalize loss by max of predicted and expected value.
+        huber_loss_delta: (:obj:`float`, `optional`, defaults to None):
+            Delta parameter used to calculate the regression loss.
+        temperature: (:obj:`float`, `optional`, defaults to 1.0):
+            Scales cell logits to control the skewness of probabilities.
+        aggregation_temperature: (:obj:`float`, `optional`, defaults to 1.0):
+            Scales aggregation logits to control the skewness of probabilities.
+        use_gumbel_for_cells: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Applies Gumbel-Softmax to cell selection.
+        use_gumbel_for_aggregation: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Applies Gumbel-Softmax to aggregation selection.
+        average_approximation_function: (:obj:`string`, `optional`, defaults to :obj:`"ratio"`):
+            Method to calculate expected average of cells in the relaxed case.
+        cell_selection_preference: (:obj:`float`, `optional`, defaults to None):
+            Preference for cell selection in ambiguous cases. Only applicable in case of weak supervision for
+            aggregation (WTQ, WikiSQL). If the total mass of the aggregation probabilities (excluding the "NONE"
+            operator) is higher than this hyperparameter, then aggregation is predicted for an example.
+        answer_loss_cutoff: (:obj:`float`, `optional`, defaults to None):
+            Ignore examples with answer loss larger than cutoff.
+        max_num_rows: (:obj:`int`, `optional`, defaults to 64):
+            Maximum number of rows.
+        max_num_columns: (:obj:`int`, `optional`, defaults to 32):
+            Maximum number of columns.
+        average_logits_per_cell: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to average logits per cell.
+        select_one_column: (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether to constrain the model to only select cells from a single column.
+        allow_empty_column_selection: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Allow not to select any column.
+        init_cell_selection_weights_to_zero: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether to initialize cell selection weights to 0 so that the initial probabilities are 50%.
+        reset_position_index_per_cell: (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Restart position indexes at every cell.
+        disable_per_token_loss: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Disable any (strong or weak) supervision on cells.
+        span_prediction: (:obj:`string`, `optional`, defaults to :obj:`"none"`):
+            Span selection mode to use. Currently only "none" is supported.
+
+    Example::
+
+        >>> from transformers import TapasModel, TapasConfig
+        >>> # Initializing a Tapas configuration
+        >>> configuration = TapasConfig()
+        >>> # Initializing a model from the configuration
+        >>> model = TapasModel(configuration)
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+
+    model_type = "tapas"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=1024,
+        type_vocab_sizes=[3, 256, 256, 2, 256, 256, 10],
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        gradient_checkpointing=False,
+        positive_label_weight=10.0,
+        num_aggregation_labels=0,
+        aggregation_loss_weight=1.0,
+        use_answer_as_supervision=None,
+        answer_loss_importance=1.0,
+        use_normalized_answer_loss=False,
+        huber_loss_delta=None,
+        temperature=1.0,
+        aggregation_temperature=1.0,
+        use_gumbel_for_cells=False,
+        use_gumbel_for_aggregation=False,
+        average_approximation_function="ratio",
+        cell_selection_preference=None,
+        answer_loss_cutoff=None,
+        max_num_rows=64,
+        max_num_columns=32,
+        average_logits_per_cell=False,
+        select_one_column=True,
+        allow_empty_column_selection=False,
+        init_cell_selection_weights_to_zero=False,
+        reset_position_index_per_cell=True,
+        disable_per_token_loss=False,
+        span_prediction="none",
+        **kwargs
+    ):
+
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        # BERT hyperparameters (with updated max_position_embeddings and type_vocab_sizes)
+        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_sizes = type_vocab_sizes
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.gradient_checkpointing = gradient_checkpointing
+
+        # Fine-tuning task hyperparameters
+        self.positive_label_weight = positive_label_weight
+        self.num_aggregation_labels = num_aggregation_labels
+        self.aggregation_loss_weight = aggregation_loss_weight
+        self.use_answer_as_supervision = use_answer_as_supervision
+        self.answer_loss_importance = answer_loss_importance
+        self.use_normalized_answer_loss = use_normalized_answer_loss
+        self.huber_loss_delta = huber_loss_delta
+        self.temperature = temperature
+        self.aggregation_temperature = aggregation_temperature
+        self.use_gumbel_for_cells = use_gumbel_for_cells
+        self.use_gumbel_for_aggregation = use_gumbel_for_aggregation
+        self.average_approximation_function = average_approximation_function
+        self.cell_selection_preference = cell_selection_preference
+        self.answer_loss_cutoff = answer_loss_cutoff
+        self.max_num_rows = max_num_rows
+        self.max_num_columns = max_num_columns
+        self.average_logits_per_cell = average_logits_per_cell
+        self.select_one_column = select_one_column
+        self.allow_empty_column_selection = allow_empty_column_selection
+        self.init_cell_selection_weights_to_zero = init_cell_selection_weights_to_zero
+        self.reset_position_index_per_cell = reset_position_index_per_cell
+        self.disable_per_token_loss = disable_per_token_loss
+        self.span_prediction = span_prediction

src/transformers/convert_tapas_original_tf_checkpoint_to_pytorch.py

@@ -0,0 +1,84 @@
+# 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.
+"""Convert TAPAS checkpoint."""
+
+
+import argparse
+
+import torch
+
+from transformers import (
+    TapasConfig,
+    TapasForQuestionAnswering,
+    TapasForSequenceClassification,
+    load_tf_weights_in_tapas,
+)
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, tapas_config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+
+    # WTQ config
+    # config = TapasConfig(# run_task_main.py hparams
+    #         num_aggregation_labels = 4,
+    #         use_answer_as_supervision = True,
+    #         # hparam_utils.py hparams
+    #         answer_loss_cutoff = 0.664694,
+    #         cell_select_pref = 0.207951,
+    #         huber_loss_delta = 0.121194,
+    #         init_cell_selection_weights_to_zero = True,
+    #         select_one_column = True,
+    #         allow_empty_column_selection = False,
+    #         temperature = 0.0352513)
+
+    # SQA config
+    config = TapasConfig()
+
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    # model = TapasForMaskedLM(config)
+    model = TapasForQuestionAnswering(config)
+    # model = TapasForSequenceClassification(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_tapas(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(
+        "--tapas_config_file",
+        default=None,
+        type=str,
+        required=True,
+        help="The config json file corresponding to the pre-trained TAPAS 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.tapas_config_file, args.pytorch_dump_path)

src/transformers/file_utils.py

@@ -193,6 +193,20 @@ except ImportError:
     _tokenizers_available = False
 
 
+try:
+    import torch_scatter
+
+    # Check we're not importing a "torch_scatter" directory somewhere
+    _scatter_available = hasattr(torch_scatter, "__version__") and hasattr(torch_scatter, "scatter")
+    if _scatter_available:
+        logger.debug(f"Succesfully imported torch-scatter version {torch_scatter.__version__}")
+    else:
+        logger.debug("Imported a torch_scatter object but this doesn't seem to be the torch-scatter library.")
+
+except ImportError:
+    _scatter_available = False
+
+
 default_cache_path = os.path.join(torch_cache_home, "transformers")
 
 
@@ -289,6 +303,14 @@ def torch_only_method(fn):
 
 
 # docstyle-ignore
+def is_sklearn_available():
+    return _has_sklearn
+
+
+def is_scatter_available():
+    return _scatter_available
+
+
 DATASETS_IMPORT_ERROR = """
 {0} requires the 🤗 Datasets library but it was not found in your environment. You can install it with:
 ```
@@ -368,6 +390,12 @@ FLAX_IMPORT_ERROR = """
 installation page: https://github.com/google/flax and follow the ones that match your environment.
 """
 
+SCATTER_IMPORT_ERROR = """
+{0} requires the torch-scatter library but it was not found in your environment. You can install it with pip as
+explained here: https://github.com/rusty1s/pytorch_scatter.
+
+"""
+
 
 def requires_datasets(obj):
     name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
@@ -417,6 +445,12 @@ def requires_sentencepiece(obj):
         raise ImportError(SENTENCEPIECE_IMPORT_ERROR.format(name))
 
 
+def requires_scatter(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_scatter_available():
+        raise ImportError(SCATTER_IMPORT_ERROR.format(name))
+
+
 def add_start_docstrings(*docstr):
     def docstring_decorator(fn):
         fn.__doc__ = "".join(docstr) + (fn.__doc__ if fn.__doc__ is not None else "")

src/transformers/modeling_auto.py

@@ -49,6 +49,7 @@ from .configuration_auto import (
     RobertaConfig,
     SqueezeBertConfig,
     T5Config,
+    TapasConfig,
     TransfoXLConfig,
     XLMConfig,
     XLMProphetNetConfig,
@@ -188,6 +189,7 @@ from .modeling_squeezebert import (
     SqueezeBertModel,
 )
 from .modeling_t5 import T5ForConditionalGeneration, T5Model
+from .modeling_tapas import TapasForMaskedLM, TapasForQuestionAnswering, TapasForSequenceClassification, TapasModel
 from .modeling_transfo_xl import TransfoXLLMHeadModel, TransfoXLModel
 from .modeling_xlm import (
     XLMForMultipleChoice,
@@ -229,6 +231,7 @@ MODEL_MAPPING = OrderedDict(
     [
         (RetriBertConfig, RetriBertModel),
         (T5Config, T5Model),
+        (TapasConfig, TapasModel),
         (DistilBertConfig, DistilBertModel),
         (AlbertConfig, AlbertModel),
         (CamembertConfig, CamembertModel),
@@ -265,6 +268,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (LayoutLMConfig, LayoutLMForMaskedLM),
         (RetriBertConfig, RetriBertModel),
         (T5Config, T5ForConditionalGeneration),
+        (TapasConfig, TapasForMaskedLM),
         (DistilBertConfig, DistilBertForMaskedLM),
         (AlbertConfig, AlbertForPreTraining),
         (CamembertConfig, CamembertForMaskedLM),
@@ -292,6 +296,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
     [
         (LayoutLMConfig, LayoutLMForMaskedLM),
         (T5Config, T5ForConditionalGeneration),
+        (TapasConfig, TapasForMaskedLM),
         (DistilBertConfig, DistilBertForMaskedLM),
         (AlbertConfig, AlbertForMaskedLM),
         (CamembertConfig, CamembertForMaskedLM),
@@ -351,6 +356,7 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
         (LongformerConfig, LongformerForMaskedLM),
         (RobertaConfig, RobertaForMaskedLM),
         (SqueezeBertConfig, SqueezeBertForMaskedLM),
+        (TapasConfig, TapasForMaskedLM),
         (BertConfig, BertForMaskedLM),
         (MobileBertConfig, MobileBertForMaskedLM),
         (FlaubertConfig, FlaubertWithLMHeadModel),
@@ -396,6 +402,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (DebertaConfig, DebertaForSequenceClassification),
         (GPT2Config, GPT2ForSequenceClassification),
         (OpenAIGPTConfig, OpenAIGPTForSequenceClassification),
+        (TapasConfig, TapasForSequenceClassification),
     ]
 )
 
@@ -410,6 +417,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
         (RobertaConfig, RobertaForQuestionAnswering),
         (SqueezeBertConfig, SqueezeBertForQuestionAnswering),
         (BertConfig, BertForQuestionAnswering),
+        (TapasConfig, TapasForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),
         (FlaubertConfig, FlaubertForQuestionAnsweringSimple),
         (MobileBertConfig, MobileBertForQuestionAnswering),

src/transformers/tokenization_auto.py

@@ -50,6 +50,7 @@ from .configuration_auto import (
     RobertaConfig,
     SqueezeBertConfig,
     T5Config,
+    TapasConfig,
     TransfoXLConfig,
     XLMConfig,
     XLMProphetNetConfig,
@@ -85,6 +86,7 @@ from .tokenization_rag import RagTokenizer
 from .tokenization_retribert import RetriBertTokenizer
 from .tokenization_roberta import RobertaTokenizer
 from .tokenization_squeezebert import SqueezeBertTokenizer
+from .tokenization_tapas import TapasTokenizer
 from .tokenization_transfo_xl import TransfoXLTokenizer
 from .tokenization_xlm import XLMTokenizer
 from .utils import logging
@@ -210,6 +212,7 @@ TOKENIZER_MAPPING = OrderedDict(
         (RagConfig, (RagTokenizer, None)),
         (XLMProphetNetConfig, (XLMProphetNetTokenizer, None)),
         (ProphetNetConfig, (ProphetNetTokenizer, None)),
+        (TapasConfig, (TapasTokenizer, None)),
     ]
 )
 

tests/test_modeling_tapas.py

@@ -0,0 +1,612 @@
+# coding=utf-8
+# Copyright 2020 Google Research and 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.
+
+
+import unittest
+
+import numpy as np
+
+from transformers import is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from .test_configuration_common import ConfigTester
+from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        TAPAS_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TapasConfig,
+        TapasForMaskedLM,
+        TapasForQuestionAnswering,
+        TapasForSequenceClassification,
+        TapasModel,
+    )
+
+
+class TapasModelTester:
+    """You can also import this e.g from .test_modeling_tapas import TapasModelTester """
+
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        initializer_range=0.02,
+        max_position_embeddings=512,
+        type_vocab_sizes=[3, 256, 256, 2, 256, 256, 10],
+        type_sequence_label_size=2,
+        positive_weight=10.0,
+        num_aggregation_labels=4,
+        num_labels=2,
+        aggregation_loss_importance=0.8,
+        use_answer_as_supervision=True,
+        answer_loss_importance=0.001,
+        use_normalized_answer_loss=False,
+        huber_loss_delta=25.0,
+        temperature=1.0,
+        agg_temperature=1.0,
+        use_gumbel_for_cells=False,
+        use_gumbel_for_agg=False,
+        average_approximation_function="ratio",
+        cell_select_pref=0.5,
+        answer_loss_cutoff=100,
+        max_num_rows=64,
+        max_num_columns=32,
+        average_logits_per_cell=True,
+        select_one_column=True,
+        allow_empty_column_selection=False,
+        init_cell_selection_weights_to_zero=False,
+        reset_position_index_per_cell=False,
+        disable_per_token_loss=False,
+        span_prediction="none",
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_sizes = type_vocab_sizes
+        self.type_sequence_label_size = type_sequence_label_size
+        self.positive_weight = positive_weight
+        self.num_aggregation_labels = num_aggregation_labels
+        self.num_labels = num_labels
+        self.aggregation_loss_importance = aggregation_loss_importance
+        self.use_answer_as_supervision = use_answer_as_supervision
+        self.answer_loss_importance = answer_loss_importance
+        self.use_normalized_answer_loss = use_normalized_answer_loss
+        self.huber_loss_delta = huber_loss_delta
+        self.temperature = temperature
+        self.agg_temperature = agg_temperature
+        self.use_gumbel_for_cells = use_gumbel_for_cells
+        self.use_gumbel_for_agg = use_gumbel_for_agg
+        self.average_approximation_function = average_approximation_function
+        self.cell_select_pref = cell_select_pref
+        self.answer_loss_cutoff = answer_loss_cutoff
+        self.max_num_rows = max_num_rows
+        self.max_num_columns = max_num_columns
+        self.average_logits_per_cell = average_logits_per_cell
+        self.select_one_column = select_one_column
+        self.allow_empty_column_selection = allow_empty_column_selection
+        self.init_cell_selection_weights_to_zero = init_cell_selection_weights_to_zero
+        self.reset_position_index_per_cell = reset_position_index_per_cell
+        self.disable_per_token_loss = disable_per_token_loss
+        self.span_prediction = span_prediction
+        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 = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = []
+        for type_vocab_size in self.type_vocab_sizes:
+            token_type_ids.append(ids_tensor(shape=[self.batch_size, self.seq_length], vocab_size=type_vocab_size))
+        token_type_ids = torch.stack(token_type_ids, dim=2)
+
+        sequence_labels = None
+        token_labels = None
+        label_ids = None
+        answer = None
+        numeric_values = None
+        numeric_values_scale = None
+        aggregation_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            label_ids = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+            answer = floats_tensor([self.batch_size])
+            numeric_values = floats_tensor([self.batch_size, self.seq_length])
+            numeric_values_scale = floats_tensor([self.batch_size, self.seq_length])
+            aggregation_labels = ids_tensor([self.batch_size], self.num_aggregation_labels)
+
+        config = TapasConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_sizes=self.type_vocab_sizes,
+            initializer_range=self.initializer_range,
+            positive_weight=self.positive_weight,
+            num_aggregation_labels=self.num_aggregation_labels,
+            num_labels=self.num_labels,
+            aggregation_loss_importance=self.aggregation_loss_importance,
+            use_answer_as_supervision=self.use_answer_as_supervision,
+            answer_loss_importance=self.answer_loss_importance,
+            use_normalized_answer_loss=self.use_normalized_answer_loss,
+            huber_loss_delta=self.huber_loss_delta,
+            temperature=self.temperature,
+            agg_temperature=self.agg_temperature,
+            use_gumbel_for_cells=self.use_gumbel_for_cells,
+            use_gumbel_for_agg=self.use_gumbel_for_agg,
+            average_approximation_function=self.average_approximation_function,
+            cell_select_pref=self.cell_select_pref,
+            answer_loss_cutoff=self.answer_loss_cutoff,
+            max_num_rows=self.max_num_rows,
+            max_num_columns=self.max_num_columns,
+            average_logits_per_cell=self.average_logits_per_cell,
+            select_one_column=self.select_one_column,
+            allow_empty_column_selection=self.allow_empty_column_selection,
+            init_cell_selection_weights_to_zero=self.init_cell_selection_weights_to_zero,
+            reset_position_index_per_cell=self.reset_position_index_per_cell,
+            disable_per_token_loss=self.disable_per_token_loss,
+            span_prediction=self.span_prediction,
+            return_dict=True,
+        )
+
+        return (
+            config,
+            input_ids,
+            input_mask,
+            token_type_ids,
+            sequence_labels,
+            token_labels,
+            label_ids,
+            answer,
+            numeric_values,
+            numeric_values_scale,
+            aggregation_labels,
+        )
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        model = TapasModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_for_masked_lm(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        model = TapasForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_question_answering(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        model = TapasForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            label_ids=label_ids,
+            answer=answer,
+            numeric_values=numeric_values,
+            numeric_values_scale=numeric_values_scale,
+            aggregation_labels=aggregation_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.logits_aggregation.shape, (self.batch_size, self.num_aggregation_labels))
+
+    def create_and_check_for_sequence_classification(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_type_ids,
+        sequence_labels,
+        token_labels,
+        label_ids,
+        answer,
+        numeric_values,
+        numeric_values_scale,
+        aggregation_labels,
+    ):
+        config.num_labels = self.num_labels
+        model = TapasForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            input_mask,
+            token_type_ids,
+            sequence_labels,
+            token_labels,
+            label_ids,
+            answer,
+            numeric_values,
+            numeric_values_scale,
+            aggregation_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class TapasModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            TapasModel,
+            TapasForMaskedLM,
+            TapasForQuestionAnswering,
+            TapasForSequenceClassification,
+        )
+        if is_torch_available()
+        else None
+    )
+    test_pruning = False
+    test_torchscript = True
+    test_resize_embeddings = True
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = TapasModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=TapasConfig, dim=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
+
+    # @slow
+    # def test_lm_outputs_same_as_reference_model(self):
+    #     """Write something that could help someone fixing this here."""
+    #     checkpoint_path = "XXX/bart-large"
+    #     model = self.big_model
+    #     tokenizer = AutoTokenizer.from_pretrained(
+    #         checkpoint_path
+    #     )  # same with AutoTokenizer (see tokenization_auto.py). This is not mandatory
+    #     # MODIFY THIS DEPENDING ON YOUR MODELS RELEVANT TASK.
+    #     batch = tokenizer(["I went to the <mask> yesterday"]).to(torch_device)
+    #     desired_mask_result = tokenizer.decode("store")  # update this
+    #     logits = model(**batch).logits
+    #     masked_index = (batch.input_ids == self.tokenizer.mask_token_id).nonzero()
+    #     assert model.num_parameters() == 175e9  # a joke
+    #     mask_entry_logits = logits[0, masked_index.item(), :]
+    #     probs = mask_entry_logits.softmax(dim=0)
+    #     _, predictions = probs.topk(1)
+    #     self.assertEqual(tokenizer.decode(predictions), desired_mask_result)
+
+    # @cached_property
+    # def big_model(self):
+    #     """Cached property means this code will only be executed once."""
+    #     checkpoint_path = "XXX/bart-large"
+    #     model = AutoModelForMaskedLM.from_pretrained(checkpoint_path).to(
+    #         torch_device
+    #     )  # test whether AutoModel can determine your model_class from checkpoint name
+    #     if torch_device == "cuda":
+    #         model.half()
+
+    # optional: do more testing! This will save you time later!
+    # @slow
+    # def test_that_XXX_can_be_used_in_a_pipeline(self):
+    #     """We can use self.big_model here without calling __init__ again."""
+    #     pass
+
+    # def test_XXX_loss_doesnt_change_if_you_add_padding(self):
+    #     pass
+
+    # def test_XXX_bad_args(self):
+    #     pass
+
+    # def test_XXX_backward_pass_reduces_loss(self):
+    #     """Test loss/gradients same as reference implementation, for example."""
+    #     pass
+
+    # @require_torch_and_cuda
+    # def test_large_inputs_in_fp16_dont_cause_overflow(self):
+    #     pass
+
+
+# Below: tests for Tapas utilities, based on segmented_tensor_test.py of the original implementation.
+# These test the operations on segmented tensors.
+class TapasUtilitiesTest(unittest.TestCase):
+    def _prepare_tables(self):
+        """Prepares two tables, both with three distinct rows.
+        The first table has two columns:
+        1.0, 2.0 | 3.0
+        2.0, 0.0 | 1.0
+        1.0, 3.0 | 4.0
+        The second table has three columns:
+        1.0 | 2.0 | 3.0
+        2.0 | 0.0 | 1.0
+        1.0 | 3.0 | 4.0
+        Returns:
+        SegmentedTensors with the tables.
+        """
+        values = torch.tensor(
+            [
+                [[1.0, 2.0, 3.0], [2.0, 0.0, 1.0], [1.0, 3.0, 4.0]],
+                [[1.0, 2.0, 3.0], [2.0, 0.0, 1.0], [1.0, 3.0, 4.0]],
+            ]
+        )
+        row_index = utils.IndexMap(
+            indices=torch.tensor(
+                [
+                    [[0, 0, 0], [1, 1, 1], [2, 2, 2]],
+                    [[0, 0, 0], [1, 1, 1], [2, 2, 2]],
+                ]
+            ),
+            num_segments=3,
+            batch_dims=1,
+        )
+        col_index = utils.IndexMap(
+            indices=torch.tensor(
+                [
+                    [[0, 0, 1], [0, 0, 1], [0, 0, 1]],
+                    [[0, 1, 2], [0, 1, 2], [0, 1, 2]],
+                ]
+            ),
+            num_segments=3,
+            batch_dims=1,
+        )
+        return values, row_index, col_index
+
+    def test_product_index(self):
+        _, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+        row_index_proj = cell_index.project_outer(cell_index)
+        col_index_proj = cell_index.project_inner(cell_index)
+
+        ind = cell_index.indices
+        self.assertEqual(cell_index.num_segments, 9)
+
+        # Projections should give back the original indices.
+        # we use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(row_index.indices.numpy(), row_index_proj.indices.numpy())
+        self.assertEqual(row_index.num_segments, row_index_proj.num_segments)
+        self.assertEqual(row_index.batch_dims, row_index_proj.batch_dims)
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(col_index.indices.numpy(), col_index_proj.indices.numpy())
+        self.assertEqual(col_index.batch_dims, col_index_proj.batch_dims)
+
+        # The first and second "column" are identified in the first table.
+        for i in range(3):
+            self.assertEqual(ind[0, i, 0], ind[0, i, 1])
+            self.assertNotEqual(ind[0, i, 0], ind[0, i, 2])
+
+        # All rows are distinct in the first table.
+        for i, i_2 in zip(range(3), range(3)):
+            for j, j_2 in zip(range(3), range(3)):
+                if i != i_2 and j != j_2:
+                    self.assertNotEqual(ind[0, i, j], ind[0, i_2, j_2])
+
+        # All cells are distinct in the second table.
+        for i, i_2 in zip(range(3), range(3)):
+            for j, j_2 in zip(range(3), range(3)):
+                if i != i_2 or j != j_2:
+                    self.assertNotEqual(ind[1, i, j], ind[1, i_2, j_2])
+
+    def test_flatten(self):
+        _, row_index, col_index = self._prepare_tables()
+        row_index_flat = utils.flatten(row_index)
+        col_index_flat = utils.flatten(col_index)
+
+        shape = [3, 4, 5]
+        batched_index = utils.IndexMap(indices=torch.zeros(shape).type(torch.LongTensor), num_segments=1, batch_dims=3)
+        batched_index_flat = utils.flatten(batched_index)
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(
+            row_index_flat.indices.numpy(), [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 5]
+        )
+        np.testing.assert_array_equal(
+            col_index_flat.indices.numpy(), [0, 0, 1, 0, 0, 1, 0, 0, 1, 3, 4, 5, 3, 4, 5, 3, 4, 5]
+        )
+        self.assertEqual(batched_index_flat.num_segments.numpy(), np.prod(shape))
+        np.testing.assert_array_equal(batched_index_flat.indices.numpy(), range(np.prod(shape)))
+
+    def test_range_index_map(self):
+        batch_shape = [3, 4]
+        num_segments = 5
+        index = utils.range_index_map(batch_shape, num_segments)
+
+        self.assertEqual(num_segments, index.num_segments)
+        self.assertEqual(2, index.batch_dims)
+        indices = index.indices
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(list(indices.size()), [3, 4, 5])
+        for i in range(batch_shape[0]):
+            for j in range(batch_shape[1]):
+                # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+                np.testing.assert_array_equal(indices[i, j, :].numpy(), range(num_segments))
+
+    def test_reduce_sum(self):
+        values, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+        row_sum, _ = utils.reduce_sum(values, row_index)
+        col_sum, _ = utils.reduce_sum(values, col_index)
+        cell_sum, _ = utils.reduce_sum(values, cell_index)
+
+        # We use np.testing.assert_allclose rather than Tensorflow's assertAllClose
+        np.testing.assert_allclose(row_sum.numpy(), [[6.0, 3.0, 8.0], [6.0, 3.0, 8.0]])
+        np.testing.assert_allclose(col_sum.numpy(), [[9.0, 8.0, 0.0], [4.0, 5.0, 8.0]])
+        np.testing.assert_allclose(
+            cell_sum.numpy(),
+            [[3.0, 3.0, 0.0, 2.0, 1.0, 0.0, 4.0, 4.0, 0.0], [1.0, 2.0, 3.0, 2.0, 0.0, 1.0, 1.0, 3.0, 4.0]],
+        )
+
+    def test_reduce_mean(self):
+        values, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+        row_mean, _ = utils.reduce_mean(values, row_index)
+        col_mean, _ = utils.reduce_mean(values, col_index)
+        cell_mean, _ = utils.reduce_mean(values, cell_index)
+
+        # We use np.testing.assert_allclose rather than Tensorflow's assertAllClose
+        np.testing.assert_allclose(
+            row_mean.numpy(), [[6.0 / 3.0, 3.0 / 3.0, 8.0 / 3.0], [6.0 / 3.0, 3.0 / 3.0, 8.0 / 3.0]]
+        )
+        np.testing.assert_allclose(col_mean.numpy(), [[9.0 / 6.0, 8.0 / 3.0, 0.0], [4.0 / 3.0, 5.0 / 3.0, 8.0 / 3.0]])
+        np.testing.assert_allclose(
+            cell_mean.numpy(),
+            [
+                [3.0 / 2.0, 3.0, 0.0, 2.0 / 2.0, 1.0, 0.0, 4.0 / 2.0, 4.0, 0.0],
+                [1.0, 2.0, 3.0, 2.0, 0.0, 1.0, 1.0, 3.0, 4.0],
+            ],
+        )
+
+    def test_reduce_max(self):
+        values = torch.as_tensor([2.0, 1.0, 0.0, 3.0])
+        index = utils.IndexMap(indices=torch.as_tensor([0, 1, 0, 1]), num_segments=2)
+        maximum, _ = utils.reduce_max(values, index)
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(maximum.numpy(), [2, 3])
+
+    def test_reduce_sum_vectorized(self):
+        values = torch.as_tensor([[1.0, 2.0, 3.0], [2.0, 3.0, 4.0], [3.0, 4.0, 5.0]])
+        index = utils.IndexMap(indices=torch.as_tensor([0, 0, 1]), num_segments=2, batch_dims=0)
+        sums, new_index = utils.reduce_sum(values, index)
+
+        # We use np.testing.assert_allclose rather than Tensorflow's assertAllClose
+        np.testing.assert_allclose(sums.numpy(), [[3.0, 5.0, 7.0], [3.0, 4.0, 5.0]])
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(new_index.indices.numpy(), [0, 1])
+        np.testing.assert_array_equal(new_index.num_segments.numpy(), 2)
+        np.testing.assert_array_equal(new_index.batch_dims, 0)
+
+    def test_gather(self):
+        values, row_index, col_index = self._prepare_tables()
+        cell_index = utils.ProductIndexMap(row_index, col_index)
+
+        # Compute sums and then gather. The result should have the same shape as
+        # the original table and each element should contain the sum the values in
+        # its cell.
+        sums, _ = utils.reduce_sum(values, cell_index)
+        cell_sum = utils.gather(sums, cell_index)
+        assert cell_sum.size() == values.size()
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_allclose(
+            cell_sum.numpy(),
+            [[[3.0, 3.0, 3.0], [2.0, 2.0, 1.0], [4.0, 4.0, 4.0]], [[1.0, 2.0, 3.0], [2.0, 0.0, 1.0], [1.0, 3.0, 4.0]]],
+        )
+
+    def test_gather_vectorized(self):
+        values = torch.as_tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])
+        index = utils.IndexMap(indices=torch.as_tensor([[0, 1], [1, 0]]), num_segments=2, batch_dims=1)
+        result = utils.gather(values, index)
+
+        # We use np.testing.assert_array_equal rather than Tensorflow's assertAllEqual
+        np.testing.assert_array_equal(result.numpy(), [[[1, 2], [3, 4]], [[7, 8], [5, 6]]])

tests/test_tokenization_tapas.py

@@ -0,0 +1,573 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import functools
+import os
+import unittest
+from typing import List, Tuple
+
+import pandas as pd
+
+import regex as re
+from transformers import AddedToken, PreTrainedTokenizerBase
+from transformers.testing_utils import require_tokenizers, slow
+from transformers.tokenization_tapas import (
+    VOCAB_FILES_NAMES,
+    BasicTokenizer,
+    TapasTokenizer,
+    WordpieceTokenizer,
+    _is_control,
+    _is_punctuation,
+    _is_whitespace,
+)
+
+from .test_tokenization_common import TokenizerTesterMixin, filter_non_english
+
+
+@require_tokenizers
+class TapasTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = TapasTokenizer
+    test_rust_tokenizer = False
+    space_between_special_tokens = True
+    from_pretrained_filter = filter_non_english
+
+    def get_clean_sequence(
+        self, tokenizer: TapasTokenizer, empty_table: bool = False, add_special_tokens: bool = True
+    ) -> Tuple[pd.DataFrame, str, List[int]]:
+        if empty_table:
+            table = pd.DataFrame.from_dict({})
+        else:
+            data = {
+                "Actors": ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"],
+                "Age": ["56", "45", "59"],
+                "Number of movies": ["87", "53", "69"],
+                "Date of birth": ["18 december 1963", "11 november 1974", "6 may 1961"],
+            }
+            table = pd.DataFrame.from_dict(data)
+        query = "Which actor appeared in the least number of movies?"
+
+        inputs = tokenizer.encode(table, query, add_special_tokens=add_special_tokens)
+
+        return table, query, inputs
+
+    # def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=20, min_length=5) -> Tuple[str, list]:
+    #     data = {
+    #         'Actors': ["Brad Pitt", "Leonardo Di Caprio", "George Clooney"],
+    #         'Age': ["56", "45", "59"],
+    #         'Number of movies': ["87", "53", "69"],
+    #         'Date of birth': ["18 december 1963", "11 november 1974", "6 may 1961"]
+    #     }
+    #     table = pd.DataFrame.from_dict(data)
+    #     output_ids = tokenizer.encode(table, add_special_tokens=False, max_length=max_length)
+    #     output_txt = tokenizer.decode(output_ids)
+    #
+    #     return output_txt, output_ids
+
+    def setUp(self):
+        super().setUp()
+
+        vocab_tokens = [
+            "[UNK]",
+            "[CLS]",
+            "[SEP]",
+            "[PAD]",
+            "[MASK]",
+            "want",
+            "##want",
+            "##ed",
+            "wa",
+            "un",
+            "runn",
+            "##ing",
+            ",",
+            "low",
+            "lowest",
+        ]
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "UNwant\u00E9d,running"
+        output_text = "unwanted, running"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file)
+
+        tokens = tokenizer.tokenize("UNwant\u00E9d,running")
+        self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [9, 6, 7, 12, 10, 11])
+
+    def test_rust_and_python_full_tokenizers(self):
+        if not self.test_rust_tokenizer:
+            return
+
+        tokenizer = self.get_tokenizer()
+        rust_tokenizer = self.get_rust_tokenizer()
+
+        sequence = "UNwant\u00E9d,running"
+
+        tokens = tokenizer.tokenize(sequence)
+        rust_tokens = rust_tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, rust_tokens)
+
+        ids = tokenizer.encode(sequence, add_special_tokens=False)
+        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
+        self.assertListEqual(ids, rust_ids)
+
+        rust_tokenizer = self.get_rust_tokenizer()
+        ids = tokenizer.encode(sequence)
+        rust_ids = rust_tokenizer.encode(sequence)
+        self.assertListEqual(ids, rust_ids)
+
+        # With lower casing
+        tokenizer = self.get_tokenizer(do_lower_case=True)
+        rust_tokenizer = self.get_rust_tokenizer(do_lower_case=True)
+
+        sequence = "UNwant\u00E9d,running"
+
+        tokens = tokenizer.tokenize(sequence)
+        rust_tokens = rust_tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, rust_tokens)
+
+        ids = tokenizer.encode(sequence, add_special_tokens=False)
+        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
+        self.assertListEqual(ids, rust_ids)
+
+        rust_tokenizer = self.get_rust_tokenizer()
+        ids = tokenizer.encode(sequence)
+        rust_ids = rust_tokenizer.encode(sequence)
+        self.assertListEqual(ids, rust_ids)
+
+    def test_chinese(self):
+        tokenizer = BasicTokenizer()
+
+        self.assertListEqual(tokenizer.tokenize("ah\u535A\u63A8zz"), ["ah", "\u535A", "\u63A8", "zz"])
+
+    def test_basic_tokenizer_lower(self):
+        tokenizer = BasicTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU?  "), ["hello", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"])
+
+    def test_basic_tokenizer_lower_strip_accents_false(self):
+        tokenizer = BasicTokenizer(do_lower_case=True, strip_accents=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hällo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["h\u00E9llo"])
+
+    def test_basic_tokenizer_lower_strip_accents_true(self):
+        tokenizer = BasicTokenizer(do_lower_case=True, strip_accents=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hallo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"])
+
+    def test_basic_tokenizer_lower_strip_accents_default(self):
+        tokenizer = BasicTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hallo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"])
+
+    def test_basic_tokenizer_no_lower(self):
+        tokenizer = BasicTokenizer(do_lower_case=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU?  "), ["HeLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    def test_basic_tokenizer_no_lower_strip_accents_false(self):
+        tokenizer = BasicTokenizer(do_lower_case=False, strip_accents=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["HäLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    def test_basic_tokenizer_no_lower_strip_accents_true(self):
+        tokenizer = BasicTokenizer(do_lower_case=False, strip_accents=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["HaLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    def test_basic_tokenizer_respects_never_split_tokens(self):
+        tokenizer = BasicTokenizer(do_lower_case=False, never_split=["[UNK]"])
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU? [UNK]"), ["HeLLo", "!", "how", "Are", "yoU", "?", "[UNK]"]
+        )
+
+    def test_wordpiece_tokenizer(self):
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", "##ing"]
+
+        vocab = {}
+        for (i, token) in enumerate(vocab_tokens):
+            vocab[token] = i
+        tokenizer = WordpieceTokenizer(vocab=vocab, unk_token="[UNK]")
+
+        self.assertListEqual(tokenizer.tokenize(""), [])
+
+        self.assertListEqual(tokenizer.tokenize("unwanted running"), ["un", "##want", "##ed", "runn", "##ing"])
+
+        self.assertListEqual(tokenizer.tokenize("unwantedX running"), ["[UNK]", "runn", "##ing"])
+
+    def test_is_whitespace(self):
+        self.assertTrue(_is_whitespace(" "))
+        self.assertTrue(_is_whitespace("\t"))
+        self.assertTrue(_is_whitespace("\r"))
+        self.assertTrue(_is_whitespace("\n"))
+        self.assertTrue(_is_whitespace("\u00A0"))
+
+        self.assertFalse(_is_whitespace("A"))
+        self.assertFalse(_is_whitespace("-"))
+
+    def test_is_control(self):
+        self.assertTrue(_is_control("\u0005"))
+
+        self.assertFalse(_is_control("A"))
+        self.assertFalse(_is_control(" "))
+        self.assertFalse(_is_control("\t"))
+        self.assertFalse(_is_control("\r"))
+
+    def test_is_punctuation(self):
+        self.assertTrue(_is_punctuation("-"))
+        self.assertTrue(_is_punctuation("$"))
+        self.assertTrue(_is_punctuation("`"))
+        self.assertTrue(_is_punctuation("."))
+
+        self.assertFalse(_is_punctuation("A"))
+        self.assertFalse(_is_punctuation(" "))
+
+    def test_clean_text(self):
+        tokenizer = self.get_tokenizer()
+        # rust_tokenizer = self.get_rust_tokenizer()
+
+        # Example taken from the issue https://github.com/huggingface/tokenizers/issues/340
+        self.assertListEqual([tokenizer.tokenize(t) for t in ["Test", "\xad", "test"]], [["[UNK]"], [], ["[UNK]"]])
+
+        # self.assertListEqual(
+        #     [rust_tokenizer.tokenize(t) for t in ["Test", "\xad", "test"]], [["[UNK]"], [], ["[UNK]"]]
+        # )
+
+    @slow
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class.from_pretrained("tapas-base-uncased")
+
+        text = tokenizer.encode("sequence builders", add_special_tokens=False)
+        text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == [101] + text + [102]
+        assert encoded_pair == [101] + text + [102] + text_2 + [102]
+
+    def test_offsets_with_special_characters(self):
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest("{} ({})".format(tokenizer.__class__.__name__, pretrained_name)):
+                tokenizer_r = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+
+                sentence = f"A, naïve {tokenizer_r.mask_token} AllenNLP sentence."
+                tokens = tokenizer_r.encode_plus(
+                    sentence,
+                    return_attention_mask=False,
+                    return_token_type_ids=False,
+                    return_offsets_mapping=True,
+                    add_special_tokens=True,
+                )
+
+                do_lower_case = tokenizer_r.do_lower_case if hasattr(tokenizer_r, "do_lower_case") else False
+                expected_results = (
+                    [
+                        ((0, 0), tokenizer_r.cls_token),
+                        ((0, 1), "A"),
+                        ((1, 2), ","),
+                        ((3, 5), "na"),
+                        ((5, 6), "##ï"),
+                        ((6, 8), "##ve"),
+                        ((9, 15), tokenizer_r.mask_token),
+                        ((16, 21), "Allen"),
+                        ((21, 23), "##NL"),
+                        ((23, 24), "##P"),
+                        ((25, 33), "sentence"),
+                        ((33, 34), "."),
+                        ((0, 0), tokenizer_r.sep_token),
+                    ]
+                    if not do_lower_case
+                    else [
+                        ((0, 0), tokenizer_r.cls_token),
+                        ((0, 1), "a"),
+                        ((1, 2), ","),
+                        ((3, 8), "naive"),
+                        ((9, 15), tokenizer_r.mask_token),
+                        ((16, 21), "allen"),
+                        ((21, 23), "##nl"),
+                        ((23, 24), "##p"),
+                        ((25, 33), "sentence"),
+                        ((33, 34), "."),
+                        ((0, 0), tokenizer_r.sep_token),
+                    ]
+                )
+
+                self.assertEqual(
+                    [e[1] for e in expected_results], tokenizer_r.convert_ids_to_tokens(tokens["input_ids"])
+                )
+                self.assertEqual([e[0] for e in expected_results], tokens["offset_mapping"])
+
+    def test_add_special_tokens(self):
+        tokenizers: List[TapasTokenizer] = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                input_table, input_query, ids = self.get_clean_sequence(
+                    tokenizer, empty_table=True, add_special_tokens=False
+                )
+
+                special_token = "[SPECIAL_TOKEN]"
+
+                tokenizer.add_special_tokens({"cls_token": special_token})
+                encoded_special_token = tokenizer.encode(input_table, special_token, add_special_tokens=False)
+                self.assertEqual(len(encoded_special_token), 1)
+
+                text = tokenizer.decode(ids + encoded_special_token, clean_up_tokenization_spaces=False)
+                encoded = tokenizer.encode(input_table, text, add_special_tokens=False)
+
+                input_encoded = tokenizer.encode(input_table, input_query, add_special_tokens=False)
+                special_token_id = tokenizer.encode(input_table, special_token, add_special_tokens=False)
+                self.assertEqual(encoded, input_encoded + special_token_id)
+
+                decoded = tokenizer.decode(encoded, skip_special_tokens=True)
+                self.assertTrue(special_token not in decoded)
+
+    def test_add_tokens_tokenizer(self):
+        tokenizers: List[TapasTokenizer] = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                input_table, input_query, ids = self.get_clean_sequence(tokenizer, empty_table=True)
+                vocab_size = tokenizer.vocab_size
+                all_size = len(tokenizer)
+
+                self.assertNotEqual(vocab_size, 0)
+
+                # We usually have added tokens from the start in tests because our vocab fixtures are
+                # smaller than the original vocabs - let's not assert this
+                # self.assertEqual(vocab_size, all_size)
+
+                new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"]
+                added_toks = tokenizer.add_tokens(new_toks)
+                vocab_size_2 = tokenizer.vocab_size
+                all_size_2 = len(tokenizer)
+
+                self.assertNotEqual(vocab_size_2, 0)
+                self.assertEqual(vocab_size, vocab_size_2)
+                self.assertEqual(added_toks, len(new_toks))
+                self.assertEqual(all_size_2, all_size + len(new_toks))
+
+                tokens = tokenizer.encode(
+                    input_table, "aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False
+                )
+
+                self.assertGreaterEqual(len(tokens), 4)
+                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
+
+                new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"}
+                added_toks_2 = tokenizer.add_special_tokens(new_toks_2)
+                vocab_size_3 = tokenizer.vocab_size
+                all_size_3 = len(tokenizer)
+
+                self.assertNotEqual(vocab_size_3, 0)
+                self.assertEqual(vocab_size, vocab_size_3)
+                self.assertEqual(added_toks_2, len(new_toks_2))
+                self.assertEqual(all_size_3, all_size_2 + len(new_toks_2))
+
+                tokens = tokenizer.encode(
+                    input_table,
+                    ">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l",
+                    add_special_tokens=False,
+                )
+
+                self.assertGreaterEqual(len(tokens), 6)
+                self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[0], tokens[1])
+                self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
+                self.assertGreater(tokens[-2], tokens[-3])
+                self.assertEqual(tokens[0], tokenizer.eos_token_id)
+                self.assertEqual(tokens[-2], tokenizer.pad_token_id)
+
+    @require_tokenizers
+    def test_encode_decode_with_spaces(self):
+        tokenizers = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                input_table, input_query, ids = self.get_clean_sequence(tokenizer, empty_table=True)
+
+                # new_toks = ["[ABC]", "[DEF]"]  # TODO(thom) add this one back when Rust toks are ready: , "GHI IHG"]
+                new_toks = [AddedToken("[ABC]", normalized=False), AddedToken("[DEF]", normalized=False)]
+                tokenizer.add_tokens(new_toks)
+                input = "[ABC][DEF][ABC][DEF]"  # TODO(thom) add back cf above: "[ABC] [DEF] [ABC] GHI IHG [DEF]"
+                if self.space_between_special_tokens:
+                    output = "[ABC] [DEF] [ABC] [DEF]"
+                else:
+                    output = input
+                encoded = tokenizer.encode(input_table, input, add_special_tokens=False)
+                decoded = tokenizer.decode(encoded, spaces_between_special_tokens=self.space_between_special_tokens)
+                self.assertIn(decoded, [output, output.lower()])
+
+    def test_encode_plus_with_padding(self):
+        tokenizers = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                input_table, input_query, ids = self.get_clean_sequence(tokenizer, empty_table=True)
+                sequence = "Sequence"
+
+                # check correct behaviour if no pad_token_id exists and add it eventually
+                self._check_no_pad_token_padding(tokenizer, sequence)
+
+                padding_size = 10
+                padding_idx = tokenizer.pad_token_id
+                token_type_padding_idx = tokenizer.pad_token_type_id
+
+                encoded_sequence = tokenizer.encode_plus(input_table, sequence, return_special_tokens_mask=True)
+                input_ids = encoded_sequence["input_ids"]
+                special_tokens_mask = encoded_sequence["special_tokens_mask"]
+                sequence_length = len(input_ids)
+
+                # Test 'longest' and 'no_padding' don't do anything
+                tokenizer.padding_side = "right"
+
+                not_padded_sequence = tokenizer.encode_plus(
+                    input_table,
+                    sequence,
+                    padding=True,
+                    return_special_tokens_mask=True,
+                )
+                not_padded_input_ids = not_padded_sequence["input_ids"]
+
+                not_padded_special_tokens_mask = not_padded_sequence["special_tokens_mask"]
+                not_padded_sequence_length = len(not_padded_input_ids)
+
+                assert sequence_length == not_padded_sequence_length
+                assert input_ids == not_padded_input_ids
+                assert special_tokens_mask == not_padded_special_tokens_mask
+
+                not_padded_sequence = tokenizer.encode_plus(
+                    input_table,
+                    sequence,
+                    padding=False,
+                    return_special_tokens_mask=True,
+                )
+                not_padded_input_ids = not_padded_sequence["input_ids"]
+
+                not_padded_special_tokens_mask = not_padded_sequence["special_tokens_mask"]
+                not_padded_sequence_length = len(not_padded_input_ids)
+
+                assert sequence_length == not_padded_sequence_length
+                assert input_ids == not_padded_input_ids
+                assert special_tokens_mask == not_padded_special_tokens_mask
+
+                # Test right padding
+                tokenizer.padding_side = "right"
+
+                right_padded_sequence = tokenizer.encode_plus(
+                    input_table,
+                    sequence,
+                    max_length=sequence_length + padding_size,
+                    padding="max_length",
+                    return_special_tokens_mask=True,
+                )
+                right_padded_input_ids = right_padded_sequence["input_ids"]
+
+                right_padded_special_tokens_mask = right_padded_sequence["special_tokens_mask"]
+                right_padded_sequence_length = len(right_padded_input_ids)
+
+                assert sequence_length + padding_size == right_padded_sequence_length
+                assert input_ids + [padding_idx] * padding_size == right_padded_input_ids
+                assert special_tokens_mask + [1] * padding_size == right_padded_special_tokens_mask
+
+                # Test left padding
+                tokenizer.padding_side = "left"
+                left_padded_sequence = tokenizer.encode_plus(
+                    input_table,
+                    sequence,
+                    max_length=sequence_length + padding_size,
+                    padding="max_length",
+                    return_special_tokens_mask=True,
+                )
+                left_padded_input_ids = left_padded_sequence["input_ids"]
+                left_padded_special_tokens_mask = left_padded_sequence["special_tokens_mask"]
+                left_padded_sequence_length = len(left_padded_input_ids)
+
+                assert sequence_length + padding_size == left_padded_sequence_length
+                assert [padding_idx] * padding_size + input_ids == left_padded_input_ids
+                assert [1] * padding_size + special_tokens_mask == left_padded_special_tokens_mask
+
+                if "token_type_ids" in tokenizer.model_input_names:
+                    token_type_ids = encoded_sequence["token_type_ids"]
+                    left_padded_token_type_ids = left_padded_sequence["token_type_ids"]
+                    right_padded_token_type_ids = right_padded_sequence["token_type_ids"]
+
+                    assert (
+                        token_type_ids + [[token_type_padding_idx] * 7] * padding_size == right_padded_token_type_ids
+                    )
+                    assert [[token_type_padding_idx] * 7] * padding_size + token_type_ids == left_padded_token_type_ids
+
+                if "attention_mask" in tokenizer.model_input_names:
+                    attention_mask = encoded_sequence["attention_mask"]
+                    right_padded_attention_mask = right_padded_sequence["attention_mask"]
+                    left_padded_attention_mask = left_padded_sequence["attention_mask"]
+
+                    assert attention_mask + [0] * padding_size == right_padded_attention_mask
+                    assert [0] * padding_size + attention_mask == left_padded_attention_mask
+
+    def test_internal_consistency(self):
+        tokenizers = self.get_tokenizers()
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                input_table, input_query, ids = self.get_clean_sequence(tokenizer, empty_table=True)
+                input_text, output_text = self.get_input_output_texts(tokenizer)
+
+                tokens = tokenizer.tokenize(input_text)
+                ids = tokenizer.convert_tokens_to_ids(tokens)
+                ids_2 = tokenizer.encode(input_table, input_text, add_special_tokens=False)
+                self.assertListEqual(ids, ids_2)
+
+                tokens_2 = tokenizer.convert_ids_to_tokens(ids)
+                self.assertNotEqual(len(tokens_2), 0)
+                text_2 = tokenizer.decode(ids)
+                self.assertIsInstance(text_2, str)
+
+                self.assertEqual(text_2, output_text)
+
+    def test_mask_output(self):
+        tokenizers = self.get_tokenizers(fast=False, do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                input_table, input_query, ids = self.get_clean_sequence(tokenizer)
+
+                if (
+                    tokenizer.build_inputs_with_special_tokens.__qualname__.split(".")[0] != "PreTrainedTokenizer"
+                    and "token_type_ids" in tokenizer.model_input_names
+                ):
+                    information = tokenizer.encode_plus(input_table, input_query, add_special_tokens=True)
+                    sequences, mask = information["input_ids"], information["token_type_ids"]
+                    self.assertEqual(len(sequences), len(mask))