multimodal / transformers /tests /utils /test_convert_slow_tokenizer.py
Li
add transformers
455a40f
import unittest
import warnings
from dataclasses import dataclass
from transformers.convert_slow_tokenizer import SpmConverter
from transformers.testing_utils import get_tests_dir
@dataclass
class FakeOriginalTokenizer:
vocab_file: str
class ConvertSlowTokenizerTest(unittest.TestCase):
def test_spm_converter_bytefallback_warning(self):
spm_model_file_without_bytefallback = get_tests_dir("fixtures/test_sentencepiece.model")
spm_model_file_with_bytefallback = get_tests_dir("fixtures/test_sentencepiece_with_bytefallback.model")
original_tokenizer_without_bytefallback = FakeOriginalTokenizer(vocab_file=spm_model_file_without_bytefallback)
with warnings.catch_warnings(record=True) as w:
_ = SpmConverter(original_tokenizer_without_bytefallback)
self.assertEqual(len(w), 0)
original_tokenizer_with_bytefallback = FakeOriginalTokenizer(vocab_file=spm_model_file_with_bytefallback)
with warnings.catch_warnings(record=True) as w:
_ = SpmConverter(original_tokenizer_with_bytefallback)
self.assertEqual(len(w), 1)
self.assertIn(
"The sentencepiece tokenizer that you are converting to a fast tokenizer uses the byte fallback option"
" which is not implemented in the fast tokenizers.",
str(w[0].message),
)