Update pipelines.py

Modified QA pipeline to consider all features for each example before generating topk answers. 
Current pipeline only takes one SquadExample, one SquadFeature, one start logit list, one end logit list to retrieve the answer, this is not correct as one SquadExample can produce multiple SquadFeatures.

src/transformers/pipelines.py

@@ -705,9 +705,16 @@ class QuestionAnsweringPipeline(Pipeline):
 
         # Convert inputs to features
         examples = self._args_parser(*texts, **kwargs)
-        features = squad_convert_examples_to_features(
-            examples, self.tokenizer, kwargs["max_seq_len"], kwargs["doc_stride"], kwargs["max_question_len"], False
-        )
+        features_list = [ squad_convert_examples_to_features(
+                            [example], 
+                            self.tokenizer, 
+                            kwargs["max_seq_len"], 
+                            kwargs["doc_stride"], 
+                            kwargs["max_question_len"], 
+                            False
+                            ) for example in examples ]
+        all_answers = []
+        for features, example in zip(features_list, examples):
             fw_args = self.inputs_for_model([f.__dict__ for f in features])
 
             # Manage tensor allocation on correct device
@@ -724,7 +731,7 @@ class QuestionAnsweringPipeline(Pipeline):
                         start, end = start.cpu().numpy(), end.cpu().numpy()
 
             answers = []
-        for (example, feature, start_, end_) in zip(examples, features, start, end):
+            for (feature, start_, end_) in zip(features, start, end):
                 # Normalize logits and spans to retrieve the answer
                 start_ = np.exp(start_) / np.sum(np.exp(start_))
                 end_ = np.exp(end_) / np.sum(np.exp(end_))
@@ -751,9 +758,12 @@ class QuestionAnsweringPipeline(Pipeline):
                     }
                     for s, e, score in zip(starts, ends, scores)
                 ]
-        if len(answers) == 1:
-            return answers[0]
-        return answers
+            answers = sorted(answers, key = lambda x:x['score'], reverse=True)[:kwargs["topk"]]    
+            all_answers+=answers
+            
+        if len(all_answers) == 1:
+           return all_answers[0]
+        return all_answers
 
     def decode(self, start: np.ndarray, end: np.ndarray, topk: int, max_answer_len: int) -> Tuple:
         """