Entailment needed [CLS] token. Now all tests pass!

textattack/attack_methods/greedy_word_swap_wir.py

@@ -91,6 +91,7 @@ class GreedyWordSwapWIR(Attack):
                     original_result,
                     best_result
                 )
+            else:
                 tokenized_text = results[0].tokenized_text
         
         return FailedAttackResult(original_result, results[0])

textattack/goal_functions/classification/untargeted_classification.py

@@ -7,7 +7,6 @@ class UntargetedClassification(ClassificationGoalFunction):
     """
     
     def _is_goal_complete(self, model_output, correct_output):
-        import pdb; pdb.set_trace()
         return model_output.argmax() != correct_output 
 
     def _get_score(self, model_output, correct_output):

textattack/goal_functions/goal_function.py

@@ -43,7 +43,6 @@ class GoalFunction:
         display purposes, and a score.
         """
         model_outputs = self._call_model(tokenized_text_list)
-        import pdb; pdb.set_trace()
         results = []
         for tokenized_text, raw_output in zip(tokenized_text_list, model_outputs):
             succeeded = self._is_goal_complete(raw_output, ground_truth_output)

textattack/tokenizers/bert_entailment_tokenizer.py

@@ -44,7 +44,7 @@ class BERTEntailmentTokenizer(BERTTokenizer):
         # Ensure they will fit in self.max_seq_length.
         self._truncate_seq_pair(tokens_a, tokens_b)
         # Concatenate and return.
-        return tokens_a + ['[SEP]'] + tokens_b + ['[SEP]']
+        return ['[CLS]'] + tokens_a + ['[SEP]'] + tokens_b + ['[SEP]']
     
     
     def convert_tokens_to_ids(self, tokens):