[Core Tokenization] Support a fix for spm fast models (huggingface#…

…26678) * fix * last attempt * current work * fix forward compatibility * save all special tokens * current state * revert additional changes * updates * remove tokenizer.model * add a test and the fix * nit * revert one more break * fix typefield issue * quality * more tests * fix fields for FC * more nits? * new additional changes * how * some updates * the fix * where do we stand * nits * nits * revert unrelated changes * nits nits nits * styling * don't break llama just yet * revert llama changes * safe arg check * fixup * Add a test for T5 * Necessary changes * Tests passing, added tokens need to not be normalized. If the added tokens are normalized, it will the stripping which seems to be unwanted for a normal functioning * Add even more tests, when normalization is set to True (which does not work 😓 ) * Add even more tests, when normalization is set to True (which does not work 😓 ) * Update to main * nits * fmt * more and more test * comments * revert change as tests are failing * make the test more readble * nits * refactor the test * nit * updates * simplify * style * style * style convert slow * Update src/transformers/convert_slow_tokenizer.py
AjayP13 · Jan 22, 2024 · 31b3323 · 31b3323
1 parent 7127d67
commit 31b3323
Show file tree

Hide file tree

Showing 2 changed files with 47 additions and 5 deletions.
diff --git a/src/transformers/convert_slow_tokenizer.py b/src/transformers/convert_slow_tokenizer.py
@@ -552,15 +552,22 @@ def tokenizer(self, proto):
 
     def normalizer(self, proto):
         precompiled_charsmap = proto.normalizer_spec.precompiled_charsmap
+        _normalizers = [
+            normalizers.Strip(left=False, right=True),  # stripping is important
+            normalizers.Replace(Regex(" {2,}"), "▁"),
+        ]
         if not precompiled_charsmap:
-            return normalizers.Sequence([normalizers.Replace(Regex(" {2,}"), " ")])
+            return normalizers.Sequence(_normalizers)
         else:
-            return normalizers.Sequence(
-                [normalizers.Precompiled(precompiled_charsmap), normalizers.Replace(Regex(" {2,}"), " ")]
-            )
+            return normalizers.Sequence([normalizers.Precompiled(precompiled_charsmap)] + _normalizers)
 
     def pre_tokenizer(self, replacement, add_prefix_space):
-        return pre_tokenizers.Metaspace(replacement=replacement, add_prefix_space=add_prefix_space)
+        prepend_scheme = "always"
+        if hasattr(self.original_tokenizer, "legacy") and not self.original_tokenizer.legacy:
+            prepend_scheme = "first"
+        return pre_tokenizers.Metaspace(
+            replacement=replacement, add_prefix_space=add_prefix_space, prepend_scheme=prepend_scheme
+        )
 
     def post_processor(self):
         return None

diff --git a/tests/models/t5/test_tokenization_t5.py b/tests/models/t5/test_tokenization_t5.py
@@ -424,6 +424,41 @@ def test_some_edge_cases(self):
         self.assertEqual(tokens, [])
         self.assertEqual(tokens, tokenizer.sp_model.encode("▁", out_type=str))
 
+    def test_fast_slow_edge_cases(self):
+        # We are testing spaces before and spaces after special tokens + space transformations
+        slow_tokenizer = T5Tokenizer.from_pretrained("t5-base", legacy=False)
+        fast_tokenizer = T5TokenizerFast.from_pretrained("t5-base", legacy=False, from_slow=True)
+        slow_tokenizer.add_tokens(AddedToken("<new_token_test_>", rstrip=False, lstrip=False, normalized=False))
+        fast_tokenizer.add_tokens(AddedToken("<new_token_test_>", rstrip=False, lstrip=False, normalized=False))
+
+        edge_case = "Hey!<new_token_test_>. How</s>Hey <new_token_test_>!"
+        EXPECTED_SLOW = ["▁Hey", "!", "<new_token_test_>", ".", "▁How", "</s>", "He", "y", "<new_token_test_>", "!"]  # fmt: skip
+        with self.subTest(f"slow {edge_case} normalized = False"):
+            self.assertEqual(slow_tokenizer.tokenize(edge_case), EXPECTED_SLOW)
+        with self.subTest(f"Fast {edge_case} normalized = False"):
+            self.assertEqual(fast_tokenizer.tokenize(edge_case), EXPECTED_SLOW)
+
+        hard_case = "Hey! <new_token_test_>. How</s>   Hey   <new_token_test_>  !     .     "
+        EXPECTED_SLOW = ["▁Hey", "!", "<new_token_test_>", ".", "▁How", "</s>", "▁Hey", "<new_token_test_>", "▁", "!", "▁", "."]  # fmt: skip
+        with self.subTest(f"slow {edge_case} normalized = False"):
+            self.assertEqual(slow_tokenizer.tokenize(hard_case), EXPECTED_SLOW)
+        with self.subTest(f"fast {edge_case} normalized = False"):
+            self.assertEqual(fast_tokenizer.tokenize(hard_case), EXPECTED_SLOW)
+
+        fast_tokenizer = T5TokenizerFast.from_pretrained("t5-base", legacy=False, from_slow=True)
+        fast_tokenizer.add_tokens(AddedToken("<new_token_test_>", rstrip=False, lstrip=False, normalized=True))
+
+        # `normalized=True` is the default normalization scheme when adding a token. Normalize -> don't strip the space.
+        # the issue now is that our slow tokenizer should NOT strip the space if we want to simulate sentencepiece token addition.
+
+        EXPECTED_FAST = ["▁Hey", "!", "<new_token_test_>", ".", "▁How", "</s>", "He", "y", "▁", "<new_token_test_>", "!"]  # fmt: skip
+        with self.subTest(f"fast {edge_case} normalized = True"):
+            self.assertEqual(fast_tokenizer.tokenize(edge_case), EXPECTED_FAST)
+
+        EXPECTED_FAST = ['▁Hey', '!', '▁', '<new_token_test_>', '.', '▁How', '</s>', '▁Hey','▁', '<new_token_test_>', '▁', '!', '▁', '.']  # fmt: skip
+        with self.subTest(f"fast {edge_case} normalized = False"):
+            self.assertEqual(fast_tokenizer.tokenize(hard_case), EXPECTED_FAST)
+
 
 @require_sentencepiece
 @require_tokenizers