TableQuestionAnsweringPipeline (#9145)

* AutoModelForTableQuestionAnswering

* TableQuestionAnsweringPipeline

* Apply suggestions from Patrick's code review

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Sylvain and Patrick comments

* Better PyTorch/TF error message

* Add integration tests

* Argument Handler naming

Co-authored-by: patrickvonplaten <patrick.v.platen@gmail.com>

* Fix docs to appease the documentation gods

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

docs/source/main_classes/pipelines.rst

@@ -34,6 +34,7 @@ There are two categories of pipeline abstractions to be aware about:
     - :class:`~transformers.TranslationPipeline`
     - :class:`~transformers.ZeroShotClassificationPipeline`
     - :class:`~transformers.Text2TextGenerationPipeline`
+    - :class:`~transformers.TableQuestionAnsweringPipeline`
 
 The pipeline abstraction
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -91,6 +92,13 @@ SummarizationPipeline
     :special-members: __call__
     :members:
 
+TableQuestionAnsweringPipeline
+=======================================================================================================================
+
+.. autoclass:: transformers.TableQuestionAnsweringPipeline
+    :special-members: __call__
+
+
 TextClassificationPipeline
 =======================================================================================================================
 

src/transformers/__init__.py

@@ -190,6 +190,7 @@
     PipelineDataFormat,
     QuestionAnsweringPipeline,
     SummarizationPipeline,
+    TableQuestionAnsweringPipeline,
     Text2TextGenerationPipeline,
     TextClassificationPipeline,
     TextGenerationPipeline,

src/transformers/file_utils.py

@@ -468,6 +468,13 @@ def wrapper(*args, **kwargs):
 """
 
 
+# docstyle-ignore
+PANDAS_IMPORT_ERROR = """
+{0} requires the pandas library but it was not found in your environment. You can install it with pip as
+explained here: https://pandas.pydata.org/pandas-docs/stable/getting_started/install.html.
+"""
+
+
 def requires_datasets(obj):
     name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
     if not is_datasets_available():
@@ -522,6 +529,12 @@ def requires_protobuf(obj):
         raise ImportError(PROTOBUF_IMPORT_ERROR.format(name))
 
 
+def requires_pandas(obj):
+    name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+    if not is_pandas_available():
+        raise ImportError(PANDAS_IMPORT_ERROR.format(name))
+
+
 def requires_scatter(obj):
     name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
     if not is_scatter_available():

src/transformers/models/tapas/configuration_tapas.py

@@ -87,41 +87,49 @@ class TapasConfig(PretrainedConfig):
             Importance weight for the regression loss.
         use_normalized_answer_loss (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether to normalize the answer loss by the maximum of the predicted and expected value.
-        huber_loss_delta: (:obj:`float`, `optional`):
+        huber_loss_delta (:obj:`float`, `optional`):
             Delta parameter used to calculate the regression loss.
-        temperature: (:obj:`float`, `optional`, defaults to 1.0):
+        temperature (:obj:`float`, `optional`, defaults to 1.0):
             Value used to control (OR change) the skewness of cell logits probabilities.
-        aggregation_temperature: (:obj:`float`, `optional`, defaults to 1.0):
+        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`):
+        use_gumbel_for_cells (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether to apply Gumbel-Softmax to cell selection.
-        use_gumbel_for_aggregation: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+        use_gumbel_for_aggregation (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether to apply Gumbel-Softmax to aggregation selection.
-        average_approximation_function: (:obj:`string`, `optional`, defaults to :obj:`"ratio"`):
+        average_approximation_function (:obj:`string`, `optional`, defaults to :obj:`"ratio"`):
             Method to calculate the expected average of cells in the weak supervision case. One of :obj:`"ratio"`,
             :obj:`"first_order"` or :obj:`"second_order"`.
-        cell_selection_preference: (:obj:`float`, `optional`):
+        cell_selection_preference (:obj:`float`, `optional`):
             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`):
+        answer_loss_cutoff (:obj:`float`, `optional`):
             Ignore examples with answer loss larger than cutoff.
-        max_num_rows: (:obj:`int`, `optional`, defaults to 64):
+        max_num_rows (:obj:`int`, `optional`, defaults to 64):
             Maximum number of rows.
-        max_num_columns: (:obj:`int`, `optional`, defaults to 32):
+        max_num_columns (:obj:`int`, `optional`, defaults to 32):
             Maximum number of columns.
-        average_logits_per_cell: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+        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`):
+        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_empty_column_selection (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether to allow not to select any column.
-        init_cell_selection_weights_to_zero: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+        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`):
+        reset_position_index_per_cell (:obj:`bool`, `optional`, defaults to :obj:`True`):
             Whether to restart position indexes at every cell (i.e. use relative position embeddings).
-        disable_per_token_loss: (:obj:`bool`, `optional`, defaults to :obj:`False`):
+        disable_per_token_loss (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether to disable any (strong or weak) supervision on cells.
+        aggregation_labels (:obj:`Dict[int, label]`, `optional`):
+            The aggregation labels used to aggregate the results. For example, the WTQ models have the following
+            aggregation labels: :obj:`{0: "NONE", 1: "SUM", 2: "AVERAGE", 3: "COUNT"}`
+        no_aggregation_label_index (:obj:`int`, `optional`):
+            If the aggregation labels are defined and one of these labels represents "No aggregation", this should be
+            set to its index. For example, the WTQ models have the "NONE" aggregation label at index 0, so that value
+            should be set to 0 for these models.
+
 
     Example::
 
@@ -174,6 +182,8 @@ def __init__(
         init_cell_selection_weights_to_zero=False,
         reset_position_index_per_cell=True,
         disable_per_token_loss=False,
+        aggregation_labels=None,
+        no_aggregation_label_index=None,
         **kwargs
     ):
 
@@ -217,3 +227,10 @@ def __init__(
         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
+
+        # Aggregation hyperparameters
+        self.aggregation_labels = aggregation_labels
+        self.no_aggregation_label_index = no_aggregation_label_index
+
+        if isinstance(self.aggregation_labels, dict):
+            self.aggregation_labels = {int(k): v for k, v in aggregation_labels.items()}

src/transformers/models/tapas/tokenization_tapas.py

@@ -1905,12 +1905,14 @@ def convert_logits_to_predictions(self, data, logits, logits_agg=None, cell_clas
                 this threshold will be selected.
 
         Returns:
-            :obj:`tuple` comprising various elements depending on the inputs: predicted_answer_coordinates
-            (``List[List[[tuple]]`` of length ``batch_size``): Predicted answer coordinates as a list of lists of
-            tuples. Each element in the list contains the predicted answer coordinates of a single example in the
-            batch, as a list of tuples. Each tuple is a cell, i.e. (row index, column index).
-            predicted_aggregation_indices (`optional`, returned when ``logits_aggregation`` is provided) ``List[int]``
-            of length ``batch_size``: Predicted aggregation operator indices of the aggregation head.
+            :obj:`tuple` comprising various elements depending on the inputs:
+
+            - predicted_answer_coordinates (``List[List[[tuple]]`` of length ``batch_size``): Predicted answer
+              coordinates as a list of lists of tuples. Each element in the list contains the predicted answer
+              coordinates of a single example in the batch, as a list of tuples. Each tuple is a cell, i.e. (row index,
+              column index).
+            - predicted_aggregation_indices (``List[int]``of length ``batch_size``, `optional`, returned when
+              ``logits_aggregation`` is provided): Predicted aggregation operator indices of the aggregation head.
         """
         # input data is of type float32
         # np.log(np.finfo(np.float32).max) = 88.72284
@@ -1969,11 +1971,11 @@ def convert_logits_to_predictions(self, data, logits, logits_agg=None, cell_clas
             answer_coordinates = sorted(answer_coordinates)
             predicted_answer_coordinates.append(answer_coordinates)
 
-        output = predicted_answer_coordinates
+        output = (predicted_answer_coordinates,)
 
         if logits_agg is not None:
             predicted_aggregation_indices = logits_agg.argmax(dim=-1)
-            output = (output, predicted_aggregation_indices.tolist())
+            output = (predicted_answer_coordinates, predicted_aggregation_indices.tolist())
 
         return output