Merge pull request #2303 from patrickvonplaten/fix_error_with_repetition_penalty

fix repetition penalty error in modeling_utils.py

src/transformers/modeling_utils.py

@@ -728,7 +728,11 @@ class PreTrainedModel(nn.Module):
             if repetition_penalty != 1.0:
                 for i in range(batch_size):
                     for previous_tokens in set(input_ids[i].tolist()):
-                        next_token_logits[i, previous_tokens] /= repetition_penalty
+                        # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                        if next_token_logits[i, previous_tokens] < 0:
+                            next_token_logits[i, previous_tokens] *= repetition_penalty
+                        else:
+                            next_token_logits[i, previous_tokens] /= repetition_penalty
 
             if do_sample:
                 # Temperature (higher temperature => more likely to sample low probability tokens)
@@ -807,7 +811,11 @@ class PreTrainedModel(nn.Module):
             if repetition_penalty != 1.0:
                 for i in range(batch_size * num_beams):
                     for previous_tokens in set(input_ids[i].tolist()):
-                        scores[i, previous_tokens] /= repetition_penalty
+                        # if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                        if scores[i, previous_tokens] < 0:
+                            scores[i, previous_tokens] *= repetition_penalty
+                        else:
+                            scores[i, previous_tokens] /= repetition_penalty
 
             if do_sample:
                 # Temperature (higher temperature => more likely to sample low probability tokens)