# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. # Copyright 2020 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import json import os import unittest from paddlenlp.transformers.bert.tokenizer import BasicTokenizer, WordpieceTokenizer from paddlenlp.transformers.roberta.tokenizer import ( RobertaBPETokenizer, RobertaChineseTokenizer, ) from paddlenlp.transformers.tokenizer_utils import AddedToken from ...testing_utils import slow from ...transformers.test_tokenizer_common import TokenizerTesterMixin VOCAB_FILES_NAMES = RobertaBPETokenizer.resource_files_names class RobertaBPETokenizationTest(TokenizerTesterMixin, unittest.TestCase): tokenizer_class = RobertaBPETokenizer test_offsets = False from_pretrained_kwargs = {"cls_token": ""} def setUp(self): super().setUp() # Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt vocab = [ "l", "o", "w", "e", "r", "s", "t", "i", "d", "n", "\u0120", "\u0120l", "\u0120n", "\u0120lo", "\u0120low", "er", "\u0120lowest", "\u0120newer", "\u0120wider", "", ] vocab_tokens = dict(zip(vocab, range(len(vocab)))) merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""] self.special_tokens_map = {"unk_token": ""} self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"]) self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"]) with open(self.vocab_file, "w", encoding="utf-8") as fp: fp.write(json.dumps(vocab_tokens) + "\n") with open(self.merges_file, "w", encoding="utf-8") as fp: fp.write("\n".join(merges)) def get_tokenizer(self, **kwargs): kwargs.update(self.special_tokens_map) return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs) def get_input_output_texts(self, tokenizer): input_text = "lower newer" output_text = "lower newer" return input_text, output_text def test_full_tokenizer(self): tokenizer = self.tokenizer_class(self.vocab_file, self.merges_file, **self.special_tokens_map) text = "lower newer" bpe_tokens = ["l", "o", "w", "er", "\u0120", "n", "e", "w", "er"] tokens = tokenizer.tokenize(text) # , add_prefix_space=True) self.assertListEqual(tokens, bpe_tokens) input_tokens = tokens + [tokenizer.unk_token] input_bpe_tokens = [0, 1, 2, 15, 10, 9, 3, 2, 15, 19] self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens) def roberta_dict_integration_testing(self): tokenizer = self.get_tokenizer() self.assertListEqual(tokenizer.encode("Hello world!", add_special_tokens=False), [0, 31414, 232, 328, 2]) self.assertListEqual( tokenizer.encode("Hello world! cécé herlolip 418", add_special_tokens=False), [0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2], ) @slow def test_sequence_builders(self): tokenizer = self.tokenizer_class.from_pretrained("roberta-base") text = tokenizer.encode("sequence builders", add_special_tokens=False)["input_ids"] text_2 = tokenizer.encode("multi-sequence build", add_special_tokens=False)["input_ids"] encoded_text_from_decode = tokenizer.encode( "sequence builders", add_special_tokens=True, add_prefix_space=False )["input_ids"] encoded_pair_from_decode = tokenizer.encode( "sequence builders", "multi-sequence build", add_special_tokens=True, add_prefix_space=False )["input_ids"] encoded_sentence = tokenizer.build_inputs_with_special_tokens(text) encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2) assert encoded_sentence == encoded_text_from_decode assert encoded_pair == encoded_pair_from_decode def test_space_encoding(self): tokenizer = self.get_tokenizer() sequence = "Encode this sequence." space_encoding = tokenizer.byte_encoder[" ".encode("utf-8")[0]] # Testing encoder arguments encoded = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=False)["input_ids"] first_char = tokenizer.convert_ids_to_tokens(encoded[0])[0] self.assertNotEqual(first_char, space_encoding) encoded = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=True)["input_ids"] first_char = tokenizer.convert_ids_to_tokens(encoded[0])[0] self.assertEqual(first_char, space_encoding) tokenizer.add_special_tokens({"bos_token": ""}) encoded = tokenizer.encode(sequence, add_special_tokens=True)["input_ids"] first_char = tokenizer.convert_ids_to_tokens(encoded[1])[0] self.assertNotEqual(first_char, space_encoding) # Testing spaces after special tokens mask = "" tokenizer.add_special_tokens( {"mask_token": AddedToken(mask, lstrip=True, rstrip=False)} ) # mask token has a left space mask_ind = tokenizer.convert_tokens_to_ids(mask) sequence = "Encode sequence" sequence_nospace = "Encode sequence" encoded = tokenizer.encode(sequence)["input_ids"] mask_loc = encoded.index(mask_ind) first_char = tokenizer.convert_ids_to_tokens(encoded[mask_loc + 1])[0] self.assertEqual(first_char, space_encoding) encoded = tokenizer.encode(sequence_nospace)["input_ids"] mask_loc = encoded.index(mask_ind) first_char = tokenizer.convert_ids_to_tokens(encoded[mask_loc + 1])[0] self.assertNotEqual(first_char, space_encoding) def test_pretokenized_inputs(self): pass def test_embeded_special_tokens(self): for tokenizer, pretrained_name, kwargs in self.tokenizers_list: with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"): tokenizer_p = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs) sentence = "A, AllenNLP sentence." tokens_p = tokenizer_p.encode( sentence, add_special_tokens=True, return_token_type_ids=True, return_attention_mask=True ) # token_type_ids should put 0 everywhere self.assertEqual(sum(tokens_p["token_type_ids"]), 0) # attention_mask should put 1 everywhere, so sum over length should be 1 self.assertEqual(sum(tokens_p["attention_mask"]) / len(tokens_p["attention_mask"]), 1) tokens_p_str = tokenizer_p.convert_ids_to_tokens(tokens_p["input_ids"]) # Rust correctly handles the space before the mask while python doesn't self.assertSequenceEqual(tokens_p["input_ids"], [0, 250, 6, 50264, 3823, 487, 21992, 3645, 4, 2]) self.assertSequenceEqual( tokens_p_str, ["", "A", ",", "", "ĠAllen", "N", "LP", "Ġsentence", ".", ""] ) def test_offsets_mapping_with_different_add_prefix_space_and_trim_space_arguments(self): # Test which aims to verify that the offsets are well adapted to the argument `add_prefix_space` and # `trim_offsets` for tokenizer, pretrained_name, kwargs in self.tokenizers_list: with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"): text_of_1_token = "hello" # `hello` is a token in the vocabulary of `pretrained_name` text = f"{text_of_1_token} {text_of_1_token}" tokenizer = self.tokenizer_class.from_pretrained(pretrained_name, add_prefix_space=True) encoding = tokenizer( text, return_offsets_mapping=True, add_special_tokens=False, return_token_type_ids=None ) self.assertEqual(encoding.offset_mapping[0], (1, len(text_of_1_token) + 1)) self.assertEqual( encoding.offset_mapping[1], (len(text_of_1_token) + 2, len(text_of_1_token) + 2 + len(text_of_1_token)), ) tokenizer = self.tokenizer_class.from_pretrained(pretrained_name, add_prefix_space=False) encoding = tokenizer( text, return_offsets_mapping=True, add_special_tokens=False, return_token_type_ids=None ) self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token))) self.assertEqual( encoding.offset_mapping[1], (len(text_of_1_token) + 1, len(text_of_1_token) + 1 + len(text_of_1_token)), ) class RobertaChineseTokenizationTest(TokenizerTesterMixin, unittest.TestCase): tokenizer_class = RobertaChineseTokenizer space_between_special_tokens = True test_seq2seq = True def setUp(self): super().setUp() vocab_tokens = [ "[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]", "want", "##want", "##ed", "wa", "un", "runn", "##ing", ",", "low", "lowest", "你", "我", "人", "天", ] self.vocab_file = os.path.join(self.tmpdirname, RobertaChineseTokenizer.resource_files_names["vocab_file"]) with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer: vocab_writer.write("".join([x + "\n" for x in vocab_tokens])) def get_input_output_texts(self, tokenizer): input_text = "UNwant\u00E9d,running" output_text = "unwanted, running" return input_text, output_text def test_full_tokenizer(self): tokenizer = self.tokenizer_class(self.vocab_file) tokens = tokenizer.tokenize("UNwant\u00E9d,running") self.assertListEqual(tokens, ["un", "##want", "##ed", ",", "runn", "##ing"]) self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [9, 6, 7, 12, 10, 11]) def test_chinese(self): tokenizer = BasicTokenizer() self.assertListEqual(tokenizer.tokenize("ah\u535A\u63A8zz"), ["ah", "\u535A", "\u63A8", "zz"]) def test_basic_tokenizer_lower(self): tokenizer = BasicTokenizer(do_lower_case=True) self.assertListEqual( tokenizer.tokenize(" \tHeLLo!how \n Are yoU? "), ["hello", "!", "how", "are", "you", "?"] ) self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"]) def test_basic_tokenizer_lower_strip_accents_false(self): tokenizer = BasicTokenizer(do_lower_case=True, strip_accents=False) self.assertListEqual( tokenizer.tokenize(" \tHäLLo!how \n Are yoU? "), ["hällo", "!", "how", "are", "you", "?"] ) self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["h\u00E9llo"]) def test_basic_tokenizer_lower_strip_accents_true(self): tokenizer = BasicTokenizer(do_lower_case=True) self.assertListEqual( tokenizer.tokenize(" \tHäLLo!how \n Are yoU? "), ["hallo", "!", "how", "are", "you", "?"] ) self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"]) def test_basic_tokenizer_lower_strip_accents_default(self): tokenizer = BasicTokenizer(do_lower_case=True) self.assertListEqual( tokenizer.tokenize(" \tHäLLo!how \n Are yoU? "), ["hallo", "!", "how", "are", "you", "?"] ) self.assertListEqual(tokenizer.tokenize("H\u00E9llo"), ["hello"]) def test_basic_tokenizer_no_lower(self): tokenizer = BasicTokenizer(do_lower_case=False) self.assertListEqual( tokenizer.tokenize(" \tHeLLo!how \n Are yoU? "), ["HeLLo", "!", "how", "Are", "yoU", "?"] ) def test_basic_tokenizer_no_lower_strip_accents_false(self): tokenizer = BasicTokenizer(do_lower_case=False, strip_accents=False) self.assertListEqual( tokenizer.tokenize(" \tHäLLo!how \n Are yoU? "), ["HäLLo", "!", "how", "Are", "yoU", "?"] ) def test_basic_tokenizer_no_lower_strip_accents_true(self): tokenizer = BasicTokenizer(do_lower_case=False, strip_accents=True) self.assertListEqual( tokenizer.tokenize(" \tHäLLo!how \n Are yoU? "), ["HaLLo", "!", "how", "Are", "yoU", "?"] ) def test_basic_tokenizer_respects_never_split_tokens(self): tokenizer = BasicTokenizer(do_lower_case=False, never_split=["[UNK]"]) self.assertListEqual( tokenizer.tokenize(" \tHeLLo!how \n Are yoU? [UNK]"), ["HeLLo", "!", "how", "Are", "yoU", "?", "[UNK]"] ) def test_wordpiece_tokenizer(self): vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", "##ing"] vocab = {} for (i, token) in enumerate(vocab_tokens): vocab[token] = i tokenizer = WordpieceTokenizer(vocab=vocab, unk_token="[UNK]") self.assertListEqual(tokenizer.tokenize(""), []) self.assertListEqual(tokenizer.tokenize("unwanted running"), ["un", "##want", "##ed", "runn", "##ing"]) self.assertListEqual(tokenizer.tokenize("unwantedX running"), ["[UNK]", "runn", "##ing"]) def test_clean_text(self): tokenizer = self.get_tokenizer() # Example taken from the issue https://github.com/huggingface/tokenizers/issues/340 self.assertListEqual([tokenizer.tokenize(t) for t in ["Test", "\xad", "test"]], [["[UNK]"], [], ["[UNK]"]]) @slow def test_sequence_builders(self): tokenizer = self.tokenizer_class.from_pretrained("hfl/roberta-wwm-ext") text = tokenizer.encode("sequence builders", return_token_type_ids=None, add_special_tokens=False)["input_ids"] text_2 = tokenizer.encode("multi-sequence build", return_token_type_ids=None, add_special_tokens=False)[ "input_ids" ] encoded_sentence = tokenizer.build_inputs_with_special_tokens(text) encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2) assert encoded_sentence == [101] + text + [102] assert encoded_pair == [101] + text + [102] + text_2 + [102] def test_change_tokenize_chinese_chars(self): list_of_commun_chinese_char = ["的", "人", "有"] text_with_chinese_char = "".join(list_of_commun_chinese_char) for tokenizer, pretrained_name, kwargs in self.tokenizers_list: with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"): kwargs["tokenize_chinese_chars"] = True tokenizer = self.tokenizer_class.from_pretrained(pretrained_name, **kwargs) ids_without_spe_char_p = tokenizer.encode( text_with_chinese_char, return_token_type_ids=None, add_special_tokens=False )["input_ids"] tokens_without_spe_char_p = tokenizer.convert_ids_to_tokens(ids_without_spe_char_p) # it is expected that each Chinese character is not preceded by "##" self.assertListEqual(tokens_without_spe_char_p, list_of_commun_chinese_char)