Modify Return Signature of TorchScript BERT

Summary:
* Return actual answer text instead of spans. (Blank means no answer, no need for exception handling in caller.)
* Return answer confidence score.
* Return has_answer score.

Differential Revision: D17983996

fbshipit-source-id: fc3084681e9d1b0453b8e4a28eeb1293dcb4d541

Author: Debojeet Chatterjee

pytext/models/output_layers/squad_output_layer.py

@@ -64,7 +64,7 @@ def get_position_preds(
         max_span_length: int,
     ):
         # the following is to enforce end_pos > start_pos.  We create a matrix
-        # of start_positions X end_positions, fill it with the sum logits,
+        # of start_position X end_position, fill it with the sum logits,
         # then mask it to be upper-triangular
         # e.g. start_pos_logits = [1, 3, 0, 5, 2]
         #      end_pos_logits = [2, 4, 6, 3, 5]
@@ -94,14 +94,10 @@ def get_position_preds(
         for i in range(logit_sum_matrix.size()[1]):
             logit_sum_matrix[:, i, i + max_span_length :] = 0
         vals, ids = logit_sum_matrix.max(-1)
-        _, start_positions = vals.max(-1)
-        end_positions = ids.gather(-1, start_positions.unsqueeze(-1)).squeeze(-1)
+        _, start_position = vals.max(-1)
+        end_position = ids.gather(-1, start_position.unsqueeze(-1)).squeeze(-1)
 
-        return (
-            start_positions,
-            end_positions,
-            logit_sum_matrix[0, start_positions, end_positions],
-        )
+        return start_position, end_position
 
     def get_pred(
         self,
@@ -110,7 +106,7 @@ def get_pred(
         contexts: Dict[str, List[Any]],
     ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]:
         start_pos_logits, end_pos_logits, has_answer_logits = logits
-        start_pos_preds, end_pos_preds, _ = self.get_position_preds(
+        start_pos_preds, end_pos_preds = self.get_position_preds(
             start_pos_logits, end_pos_logits, self.max_answer_len
         )
         has_answer_preds = has_answer_logits.argmax(-1)