Avoid erasing the attention mask when double padding (#8915)

2020-12-03 10:45:07 -05:00
parent 0deece9c53
commit 8453201cfe
1 changed files with 2 additions and 3 deletions
--- a/src/transformers/tokenization_utils_base.py
+++ b/src/transformers/tokenization_utils_base.py
@@ -3047,8 +3047,7 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
                encoded_inputs["input_ids"] = [self.pad_token_id] * difference + encoded_inputs["input_ids"]
            else:
                raise ValueError("Invalid padding strategy:" + str(self.padding_side))
-        else:
-            if return_attention_mask:
+        elif return_attention_mask and "attention_mask" not in encoded_inputs:
            encoded_inputs["attention_mask"] = [1] * len(encoded_inputs["input_ids"])

        return encoded_inputs