fix t5 special tokens (#8435)

src/transformers/tokenization_t5.py

@@ -249,8 +249,17 @@ class T5Tokenizer(PreTrainedTokenizer):
 
     def convert_tokens_to_string(self, tokens):
         """ Converts a sequence of tokens (string) in a single string. """
-        out_string = self.sp_model.decode_pieces(tokens)
-        return out_string
+        current_sub_tokens = []
+        out_string = ""
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                out_string += self.sp_model.decode_pieces(current_sub_tokens) + token + " "
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+        out_string += self.sp_model.decode_pieces(current_sub_tokens)
+        return out_string.strip()
 
     def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
         if not os.path.isdir(save_directory):

tests/test_tokenization_t5.py

@@ -222,3 +222,18 @@ class T5TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
 
         self.assertEqual(expected_src_tokens, src_ids)
         self.assertEqual(expected_tgt_tokens, tgt_ids)
+
+    def test_fast_and_slow_same_result(self):
+        src_text = "<pad> Today is <unk> nice day </s>"
+        tgt_ids = [0, 1960, 19, 2, 1245, 239, 1]
+        tgt_text = "<pad> Today is<unk> nice day</s>"
+
+        fast_ids = self.t5_base_tokenizer_fast(src_text, add_special_tokens=False).input_ids
+        slow_ids = self.t5_base_tokenizer(src_text, add_special_tokens=False).input_ids
+        self.assertEqual(tgt_ids, fast_ids)
+        self.assertEqual(tgt_ids, slow_ids)
+
+        fast_text = self.t5_base_tokenizer_fast.decode(fast_ids)
+        slow_text = self.t5_base_tokenizer.decode(fast_ids)
+        self.assertEqual(tgt_text, fast_text)
+        self.assertEqual(tgt_text, slow_text)