2019-07-05 17:20:27 +08:00
|
|
|
# coding=utf-8
|
|
|
|
# Copyright 2019 HuggingFace Inc.
|
|
|
|
#
|
|
|
|
# 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.
|
2019-12-22 23:20:32 +08:00
|
|
|
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2019-07-09 21:58:58 +08:00
|
|
|
import os
|
2019-12-23 01:12:11 +08:00
|
|
|
import pickle
|
2020-06-16 05:12:51 +08:00
|
|
|
import re
|
2019-12-21 22:57:32 +08:00
|
|
|
import shutil
|
2019-07-09 16:25:18 +08:00
|
|
|
import tempfile
|
2020-04-09 04:22:44 +08:00
|
|
|
from collections import OrderedDict
|
2020-05-15 01:14:26 +08:00
|
|
|
from typing import TYPE_CHECKING, Dict, Tuple, Union
|
2019-12-21 22:57:32 +08:00
|
|
|
|
2020-02-25 01:09:46 +08:00
|
|
|
from tests.utils import require_tf, require_torch
|
2020-06-16 05:12:51 +08:00
|
|
|
from transformers import PreTrainedTokenizer, PreTrainedTokenizerFast
|
2020-02-25 01:09:46 +08:00
|
|
|
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-05-15 01:14:26 +08:00
|
|
|
if TYPE_CHECKING:
|
|
|
|
from transformers import (
|
|
|
|
PretrainedConfig,
|
|
|
|
PreTrainedModel,
|
|
|
|
TFPreTrainedModel,
|
|
|
|
)
|
|
|
|
|
|
|
|
|
2020-04-09 04:22:44 +08:00
|
|
|
def merge_model_tokenizer_mappings(
|
2020-04-09 21:09:00 +08:00
|
|
|
model_mapping: Dict["PretrainedConfig", Union["PreTrainedModel", "TFPreTrainedModel"]],
|
|
|
|
tokenizer_mapping: Dict["PretrainedConfig", Tuple["PreTrainedTokenizer", "PreTrainedTokenizerFast"]],
|
|
|
|
) -> Dict[
|
|
|
|
Union["PreTrainedTokenizer", "PreTrainedTokenizerFast"],
|
|
|
|
Tuple["PretrainedConfig", Union["PreTrainedModel", "TFPreTrainedModel"]],
|
|
|
|
]:
|
2020-04-09 04:22:44 +08:00
|
|
|
configurations = list(model_mapping.keys())
|
|
|
|
model_tokenizer_mapping = OrderedDict([])
|
|
|
|
|
|
|
|
for configuration in configurations:
|
|
|
|
model = model_mapping[configuration]
|
|
|
|
tokenizer = tokenizer_mapping[configuration][0]
|
|
|
|
tokenizer_fast = tokenizer_mapping[configuration][1]
|
|
|
|
|
|
|
|
model_tokenizer_mapping.update({tokenizer: (configuration, model)})
|
|
|
|
if tokenizer_fast is not None:
|
|
|
|
model_tokenizer_mapping.update({tokenizer_fast: (configuration, model)})
|
|
|
|
|
|
|
|
return model_tokenizer_mapping
|
|
|
|
|
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
class TokenizerTesterMixin:
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
tokenizer_class = None
|
2019-12-25 02:29:01 +08:00
|
|
|
test_rust_tokenizer = False
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def setUp(self):
|
|
|
|
self.tmpdirname = tempfile.mkdtemp()
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def tearDown(self):
|
|
|
|
shutil.rmtree(self.tmpdirname)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
def get_input_output_texts(self, tokenizer):
|
|
|
|
input_txt = self.get_clean_sequence(tokenizer)[0]
|
|
|
|
return input_txt, input_txt
|
|
|
|
|
|
|
|
def get_clean_sequence(self, tokenizer, with_prefix_space=False, max_length=None) -> Tuple[str, list]:
|
|
|
|
toks = [(i, tokenizer.decode([i], clean_up_tokenization_spaces=False)) for i in range(len(tokenizer))]
|
|
|
|
toks = list(filter(lambda t: re.match(r"^[ a-zA-Z]+$", t[1]), toks))
|
|
|
|
toks = list(filter(lambda t: [t[0]] == tokenizer.encode(t[1], add_special_tokens=False), toks))
|
|
|
|
if max_length is not None and len(toks) > max_length:
|
|
|
|
toks = toks[:max_length]
|
|
|
|
# toks_str = [t[1] for t in toks]
|
|
|
|
toks_ids = [t[0] for t in toks]
|
|
|
|
|
|
|
|
# Ensure consistency
|
|
|
|
output_txt = tokenizer.decode(toks_ids, clean_up_tokenization_spaces=False)
|
|
|
|
if " " not in output_txt and len(toks_ids) > 1:
|
|
|
|
output_txt = (
|
|
|
|
tokenizer.decode([toks_ids[0]], clean_up_tokenization_spaces=False)
|
|
|
|
+ " "
|
|
|
|
+ tokenizer.decode(toks_ids[1:], clean_up_tokenization_spaces=False)
|
|
|
|
)
|
|
|
|
if with_prefix_space:
|
|
|
|
output_txt = " " + output_txt
|
|
|
|
output_ids = tokenizer.encode(output_txt, add_special_tokens=False)
|
|
|
|
return output_txt, output_ids
|
|
|
|
|
|
|
|
def get_tokenizers(self, fast=True, **kwargs) -> PreTrainedTokenizer:
|
|
|
|
if fast and self.test_rust_tokenizer:
|
|
|
|
return [self.get_tokenizer(**kwargs), self.get_rust_tokenizer(**kwargs)]
|
|
|
|
return [self.get_tokenizer(**kwargs)]
|
|
|
|
|
2020-05-19 22:46:55 +08:00
|
|
|
def get_tokenizer(self, **kwargs) -> PreTrainedTokenizer:
|
|
|
|
return self.tokenizer_class.from_pretrained(self.tmpdirname, **kwargs)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2019-12-25 02:29:01 +08:00
|
|
|
def get_rust_tokenizer(self, **kwargs):
|
|
|
|
raise NotImplementedError
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# def get_input_output_texts(self) -> Tuple[str, str]:
|
|
|
|
# """Feel free to overwrite"""
|
|
|
|
# # TODO: @property
|
|
|
|
# return (
|
|
|
|
# "This is a test",
|
|
|
|
# "This is a test",
|
|
|
|
# )
|
2019-07-09 21:58:58 +08:00
|
|
|
|
2020-02-25 01:09:46 +08:00
|
|
|
@staticmethod
|
|
|
|
def convert_batch_encode_plus_format_to_encode_plus(batch_encode_plus_sequences):
|
|
|
|
# Switch from batch_encode_plus format: {'input_ids': [[...], [...]], ...}
|
2020-05-19 22:46:55 +08:00
|
|
|
# to the list of examples/ encode_plus format: [{'input_ids': [...], ...}, {'input_ids': [...], ...}]
|
2020-02-25 01:09:46 +08:00
|
|
|
return [
|
|
|
|
{value: batch_encode_plus_sequences[value][i] for value in batch_encode_plus_sequences.keys()}
|
2020-03-10 01:48:58 +08:00
|
|
|
for i in range(len(batch_encode_plus_sequences["input_ids"]))
|
2020-02-25 01:09:46 +08:00
|
|
|
]
|
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_tokenizers_common_properties(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers()
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
attributes_list = [
|
|
|
|
"bos_token",
|
|
|
|
"eos_token",
|
|
|
|
"unk_token",
|
|
|
|
"sep_token",
|
|
|
|
"pad_token",
|
|
|
|
"cls_token",
|
|
|
|
"mask_token",
|
|
|
|
]
|
|
|
|
for attr in attributes_list:
|
|
|
|
self.assertTrue(hasattr(tokenizer, attr))
|
|
|
|
self.assertTrue(hasattr(tokenizer, attr + "_id"))
|
|
|
|
|
|
|
|
self.assertTrue(hasattr(tokenizer, "additional_special_tokens"))
|
|
|
|
self.assertTrue(hasattr(tokenizer, "additional_special_tokens_ids"))
|
|
|
|
|
|
|
|
attributes_list = [
|
|
|
|
"model_max_length",
|
|
|
|
"init_inputs",
|
|
|
|
"init_kwargs",
|
|
|
|
]
|
|
|
|
if not isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
attributes_list += [
|
|
|
|
"added_tokens_encoder",
|
|
|
|
"added_tokens_decoder",
|
|
|
|
]
|
|
|
|
for attr in attributes_list:
|
|
|
|
self.assertTrue(hasattr(tokenizer, attr))
|
2019-09-06 03:31:29 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_save_and_load_tokenizer(self):
|
|
|
|
# safety check on max_len default value so we are sure the test works
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(fast=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
self.assertNotEqual(tokenizer.max_len, 42)
|
2019-08-30 23:09:36 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
# Now let's start the test
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(fast=False, model_max_length=42)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sample_text = "He is very happy, UNwant\u00E9d,running"
|
|
|
|
before_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizer.save_pretrained(self.tmpdirname)
|
|
|
|
tokenizer = self.tokenizer_class.from_pretrained(self.tmpdirname)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
after_tokens = tokenizer.encode(sample_text, add_special_tokens=False)
|
|
|
|
self.assertListEqual(before_tokens, after_tokens)
|
2019-08-30 23:09:36 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(tokenizer.model_max_length, 42)
|
|
|
|
tokenizer = self.tokenizer_class.from_pretrained(self.tmpdirname, model_max_length=43)
|
|
|
|
self.assertEqual(tokenizer.model_max_length, 43)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_pickle_tokenizer(self):
|
2020-05-19 22:46:55 +08:00
|
|
|
"""Google pickle __getstate__ __setstate__ if you are struggling with this."""
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers()
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
self.assertIsNotNone(tokenizer)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
text = "Munich and Berlin are nice cities"
|
|
|
|
subwords = tokenizer.tokenize(text)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
filename = os.path.join(self.tmpdirname, "tokenizer.bin")
|
|
|
|
with open(filename, "wb") as handle:
|
|
|
|
pickle.dump(tokenizer, handle)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
with open(filename, "rb") as handle:
|
|
|
|
tokenizer_new = pickle.load(handle)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
subwords_loaded = tokenizer_new.tokenize(text)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertListEqual(subwords, subwords_loaded)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_added_tokens_do_lower_case(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
# TODO(thom) activate fast tokenizer tests once Rust tokenizers accepts white spaces in added tokens
|
|
|
|
tokenizers = self.get_tokenizers(fast=False, do_lower_case=True)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
special_token = tokenizer.all_special_tokens[0]
|
2019-12-04 23:53:25 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
text = special_token + " aaaaa bbbbbb low cccccccccdddddddd l " + special_token
|
|
|
|
text2 = special_token + " AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l " + special_token
|
Consider do_lower_case in PreTrainedTokenizer
As pointed out in #1545, when using an uncased model, and adding
a new uncased token, the tokenizer does not correctly identify this
in the case that the input text contains the token in a cased format.
For instance, if we load bert-base-uncased into BertTokenizer, and
then use .add_tokens() to add "cool-token", we get the expected
result for .tokenize('this is a cool-token'). However, we get a
possibly unexpected result for .tokenize('this is a cOOl-Token'),
which in fact mirrors the result for the former from before the new
token was added.
This commit adds
- functionality to PreTrainedTokenizer to handle this
situation in case a tokenizer (currently Bert, DistilBert,
and XLNet) has the do_lower_case=True kwarg by:
1) lowercasing tokens added with .add_tokens()
2) lowercasing text at the beginning of .tokenize()
- new common test case for tokenizers
https://github.com/huggingface/transformers/issues/1545
2019-11-06 19:18:16 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
toks0 = tokenizer.tokenize(text) # toks before adding new_toks
|
Consider do_lower_case in PreTrainedTokenizer
As pointed out in #1545, when using an uncased model, and adding
a new uncased token, the tokenizer does not correctly identify this
in the case that the input text contains the token in a cased format.
For instance, if we load bert-base-uncased into BertTokenizer, and
then use .add_tokens() to add "cool-token", we get the expected
result for .tokenize('this is a cool-token'). However, we get a
possibly unexpected result for .tokenize('this is a cOOl-Token'),
which in fact mirrors the result for the former from before the new
token was added.
This commit adds
- functionality to PreTrainedTokenizer to handle this
situation in case a tokenizer (currently Bert, DistilBert,
and XLNet) has the do_lower_case=True kwarg by:
1) lowercasing tokens added with .add_tokens()
2) lowercasing text at the beginning of .tokenize()
- new common test case for tokenizers
https://github.com/huggingface/transformers/issues/1545
2019-11-06 19:18:16 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd", "AAAAA BBBBBB", "CCCCCCCCCDDDDDDDD"]
|
|
|
|
added = tokenizer.add_tokens(new_toks)
|
|
|
|
self.assertEqual(added, 2)
|
Consider do_lower_case in PreTrainedTokenizer
As pointed out in #1545, when using an uncased model, and adding
a new uncased token, the tokenizer does not correctly identify this
in the case that the input text contains the token in a cased format.
For instance, if we load bert-base-uncased into BertTokenizer, and
then use .add_tokens() to add "cool-token", we get the expected
result for .tokenize('this is a cool-token'). However, we get a
possibly unexpected result for .tokenize('this is a cOOl-Token'),
which in fact mirrors the result for the former from before the new
token was added.
This commit adds
- functionality to PreTrainedTokenizer to handle this
situation in case a tokenizer (currently Bert, DistilBert,
and XLNet) has the do_lower_case=True kwarg by:
1) lowercasing tokens added with .add_tokens()
2) lowercasing text at the beginning of .tokenize()
- new common test case for tokenizers
https://github.com/huggingface/transformers/issues/1545
2019-11-06 19:18:16 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
toks = tokenizer.tokenize(text)
|
|
|
|
toks2 = tokenizer.tokenize(text2)
|
Consider do_lower_case in PreTrainedTokenizer
As pointed out in #1545, when using an uncased model, and adding
a new uncased token, the tokenizer does not correctly identify this
in the case that the input text contains the token in a cased format.
For instance, if we load bert-base-uncased into BertTokenizer, and
then use .add_tokens() to add "cool-token", we get the expected
result for .tokenize('this is a cool-token'). However, we get a
possibly unexpected result for .tokenize('this is a cOOl-Token'),
which in fact mirrors the result for the former from before the new
token was added.
This commit adds
- functionality to PreTrainedTokenizer to handle this
situation in case a tokenizer (currently Bert, DistilBert,
and XLNet) has the do_lower_case=True kwarg by:
1) lowercasing tokens added with .add_tokens()
2) lowercasing text at the beginning of .tokenize()
- new common test case for tokenizers
https://github.com/huggingface/transformers/issues/1545
2019-11-06 19:18:16 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(len(toks), len(toks2))
|
|
|
|
self.assertListEqual(toks, toks2)
|
|
|
|
if not isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
# Python tokenizers can have added tokens with spaces inside them
|
|
|
|
# cf https://github.com/huggingface/tokenizers/issues/302
|
|
|
|
self.assertNotEqual(len(toks), len(toks0)) # toks0 should be longer
|
Consider do_lower_case in PreTrainedTokenizer
As pointed out in #1545, when using an uncased model, and adding
a new uncased token, the tokenizer does not correctly identify this
in the case that the input text contains the token in a cased format.
For instance, if we load bert-base-uncased into BertTokenizer, and
then use .add_tokens() to add "cool-token", we get the expected
result for .tokenize('this is a cool-token'). However, we get a
possibly unexpected result for .tokenize('this is a cOOl-Token'),
which in fact mirrors the result for the former from before the new
token was added.
This commit adds
- functionality to PreTrainedTokenizer to handle this
situation in case a tokenizer (currently Bert, DistilBert,
and XLNet) has the do_lower_case=True kwarg by:
1) lowercasing tokens added with .add_tokens()
2) lowercasing text at the beginning of .tokenize()
- new common test case for tokenizers
https://github.com/huggingface/transformers/issues/1545
2019-11-06 19:18:16 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# Check that none of the special tokens are lowercased
|
|
|
|
sequence_with_special_tokens = "A " + " yEs ".join(tokenizer.all_special_tokens) + " B"
|
|
|
|
tokenized_sequence = tokenizer.tokenize(sequence_with_special_tokens)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
for special_token in tokenizer.all_special_tokens:
|
|
|
|
self.assertTrue(special_token in tokenized_sequence)
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(fast=False, do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
special_token = tokenizer.all_special_tokens[0]
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
text = special_token + " aaaaa bbbbbb low cccccccccdddddddd l " + special_token
|
|
|
|
text2 = special_token + " AAAAA BBBBBB low CCCCCCCCCDDDDDDDD l " + special_token
|
2019-07-09 16:25:18 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd", "AAAAA BBBBBB", "CCCCCCCCCDDDDDDDD"]
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
toks0 = tokenizer.tokenize(text) # toks before adding new_toks
|
2019-09-02 08:27:39 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
added = tokenizer.add_tokens(new_toks)
|
|
|
|
self.assertEqual(added, 4)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
toks = tokenizer.tokenize(text)
|
|
|
|
toks2 = tokenizer.tokenize(text2)
|
2019-07-05 17:20:27 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(len(toks), len(toks2)) # Length should still be the same
|
|
|
|
self.assertNotEqual(toks[1], toks2[1]) # But at least the first non-special tokens should differ
|
|
|
|
if not isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
# Python tokenizers can have added tokens with spaces inside them
|
|
|
|
# cf https://github.com/huggingface/tokenizers/issues/302
|
|
|
|
self.assertNotEqual(len(toks), len(toks0)) # toks0 should be longer
|
2019-07-15 23:30:42 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
def test_add_tokens_tokenizer(self):
|
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
vocab_size = tokenizer.vocab_size
|
|
|
|
all_size = len(tokenizer)
|
|
|
|
|
|
|
|
self.assertNotEqual(vocab_size, 0)
|
|
|
|
self.assertEqual(vocab_size, all_size)
|
|
|
|
|
|
|
|
new_toks = ["aaaaa bbbbbb", "cccccccccdddddddd"]
|
|
|
|
added_toks = tokenizer.add_tokens(new_toks)
|
|
|
|
vocab_size_2 = tokenizer.vocab_size
|
|
|
|
all_size_2 = len(tokenizer)
|
|
|
|
|
|
|
|
self.assertNotEqual(vocab_size_2, 0)
|
|
|
|
self.assertEqual(vocab_size, vocab_size_2)
|
|
|
|
self.assertEqual(added_toks, len(new_toks))
|
|
|
|
self.assertEqual(all_size_2, all_size + len(new_toks))
|
|
|
|
|
|
|
|
tokens = tokenizer.encode("aaaaa bbbbbb low cccccccccdddddddd l", add_special_tokens=False)
|
|
|
|
|
|
|
|
self.assertGreaterEqual(len(tokens), 4)
|
|
|
|
self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
|
|
|
|
self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
|
|
|
|
|
|
|
|
new_toks_2 = {"eos_token": ">>>>|||<||<<|<<", "pad_token": "<<<<<|||>|>>>>|>"}
|
|
|
|
added_toks_2 = tokenizer.add_special_tokens(new_toks_2)
|
|
|
|
vocab_size_3 = tokenizer.vocab_size
|
|
|
|
all_size_3 = len(tokenizer)
|
|
|
|
|
|
|
|
self.assertNotEqual(vocab_size_3, 0)
|
|
|
|
self.assertEqual(vocab_size, vocab_size_3)
|
|
|
|
self.assertEqual(added_toks_2, len(new_toks_2))
|
|
|
|
self.assertEqual(all_size_3, all_size_2 + len(new_toks_2))
|
|
|
|
|
|
|
|
tokens = tokenizer.encode(
|
|
|
|
">>>>|||<||<<|<< aaaaabbbbbb low cccccccccdddddddd <<<<<|||>|>>>>|> l", add_special_tokens=False
|
|
|
|
)
|
|
|
|
|
|
|
|
self.assertGreaterEqual(len(tokens), 6)
|
|
|
|
self.assertGreater(tokens[0], tokenizer.vocab_size - 1)
|
|
|
|
self.assertGreater(tokens[0], tokens[1])
|
|
|
|
self.assertGreater(tokens[-2], tokenizer.vocab_size - 1)
|
|
|
|
self.assertGreater(tokens[-2], tokens[-3])
|
|
|
|
self.assertEqual(tokens[0], tokenizer.eos_token_id)
|
|
|
|
self.assertEqual(tokens[-2], tokenizer.pad_token_id)
|
2019-07-16 21:11:29 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_add_special_tokens(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
input_text, ids = self.get_clean_sequence(tokenizer)
|
2019-11-13 04:27:57 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
special_token = "[SPECIAL_TOKEN]"
|
2019-11-13 04:27:57 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizer.add_special_tokens({"cls_token": special_token})
|
|
|
|
encoded_special_token = tokenizer.encode(special_token, add_special_tokens=False)
|
|
|
|
self.assertEqual(len(encoded_special_token), 1)
|
2019-11-13 04:27:57 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
text = tokenizer.decode(ids + encoded_special_token, clean_up_tokenization_spaces=False)
|
|
|
|
encoded = tokenizer.encode(text, add_special_tokens=False)
|
2019-11-13 04:27:57 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
input_encoded = tokenizer.encode(input_text, add_special_tokens=False)
|
|
|
|
special_token_id = tokenizer.encode(special_token, add_special_tokens=False)
|
|
|
|
self.assertEqual(encoded, input_encoded + special_token_id)
|
2019-11-13 04:27:57 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
decoded = tokenizer.decode(encoded, skip_special_tokens=True)
|
|
|
|
self.assertTrue(special_token not in decoded)
|
2019-07-16 21:11:29 +08:00
|
|
|
|
2020-05-19 22:46:55 +08:00
|
|
|
def test_internal_consistency(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers()
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
input_text, output_text = self.get_input_output_texts(tokenizer)
|
2019-07-16 21:11:29 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokens = tokenizer.tokenize(input_text)
|
|
|
|
ids = tokenizer.convert_tokens_to_ids(tokens)
|
|
|
|
ids_2 = tokenizer.encode(input_text, add_special_tokens=False)
|
|
|
|
self.assertListEqual(ids, ids_2)
|
2019-07-16 21:11:29 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokens_2 = tokenizer.convert_ids_to_tokens(ids)
|
|
|
|
self.assertNotEqual(len(tokens_2), 0)
|
|
|
|
text_2 = tokenizer.decode(ids)
|
|
|
|
self.assertIsInstance(text_2, str)
|
2019-08-05 20:08:56 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(text_2, output_text)
|
2019-08-05 20:08:56 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_encode_decode_with_spaces(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
2019-12-12 01:36:37 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
new_toks = ["[ABC]", "[DEF]"] # TODO(thom) add this one back when Rust toks are ready: , "GHI IHG"]
|
|
|
|
tokenizer.add_tokens(new_toks)
|
|
|
|
input = "[ABC] [DEF] [ABC] [DEF]" # TODO(thom) add back cf above: "[ABC] [DEF] [ABC] GHI IHG [DEF]"
|
|
|
|
encoded = tokenizer.encode(input, add_special_tokens=False)
|
|
|
|
decoded = tokenizer.decode(encoded)
|
|
|
|
self.assertEqual(decoded, input)
|
2019-08-05 20:08:56 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_pretrained_model_lists(self):
|
|
|
|
weights_list = list(self.tokenizer_class.max_model_input_sizes.keys())
|
|
|
|
weights_lists_2 = []
|
|
|
|
for file_id, map_list in self.tokenizer_class.pretrained_vocab_files_map.items():
|
|
|
|
weights_lists_2.append(list(map_list.keys()))
|
2019-08-05 20:08:56 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
for weights_list_2 in weights_lists_2:
|
|
|
|
self.assertListEqual(weights_list, weights_list_2)
|
2019-09-03 05:47:16 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_mask_output(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(fast=False, do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
|
|
|
|
if (
|
|
|
|
tokenizer.build_inputs_with_special_tokens.__qualname__.split(".")[0] != "PreTrainedTokenizer"
|
|
|
|
and "token_type_ids" in tokenizer.model_input_names
|
|
|
|
):
|
|
|
|
seq_0 = "Test this method."
|
|
|
|
seq_1 = "With these inputs."
|
|
|
|
information = tokenizer.encode_plus(seq_0, seq_1, add_special_tokens=True)
|
|
|
|
sequences, mask = information["input_ids"], information["token_type_ids"]
|
|
|
|
self.assertEqual(len(sequences), len(mask))
|
2019-12-22 22:34:15 +08:00
|
|
|
|
|
|
|
def test_number_of_added_tokens(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
seq_0 = "Test this method."
|
|
|
|
seq_1 = "With these inputs."
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
sequences = tokenizer.encode(seq_0, seq_1, add_special_tokens=False)
|
2020-06-23 19:36:57 +08:00
|
|
|
attached_sequences = tokenizer.encode(seq_0, seq_1, add_special_tokens=True)
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# Method is implemented (e.g. not GPT-2)
|
|
|
|
if len(attached_sequences) != 2:
|
|
|
|
self.assertEqual(
|
|
|
|
tokenizer.num_special_tokens_to_add(pair=True), len(attached_sequences) - len(sequences)
|
|
|
|
)
|
2019-12-22 22:34:15 +08:00
|
|
|
|
|
|
|
def test_maximum_encoding_length_single_input(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
seq_0, ids = self.get_clean_sequence(tokenizer)
|
|
|
|
stride = 2
|
|
|
|
|
|
|
|
sequence = tokenizer.encode(seq_0, add_special_tokens=False)
|
|
|
|
# self.assertEqual(sequence, ids)
|
|
|
|
|
|
|
|
total_length = len(sequence)
|
|
|
|
information = tokenizer.encode_plus(
|
|
|
|
seq_0,
|
|
|
|
max_length=total_length - 2,
|
|
|
|
add_special_tokens=False,
|
|
|
|
stride=stride,
|
|
|
|
truncation="longest_first",
|
|
|
|
return_overflowing_tokens=True,
|
2020-06-23 19:36:57 +08:00
|
|
|
# add_prefix_space=False,
|
2020-06-16 05:12:51 +08:00
|
|
|
)
|
|
|
|
|
|
|
|
# Overflowing tokens are handled quite differently in slow and fast tokenizers
|
|
|
|
if isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
truncated_sequence = information["input_ids"][0]
|
|
|
|
overflowing_tokens = information["input_ids"][1]
|
|
|
|
self.assertEqual(len(information["input_ids"]), 2)
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), total_length - 2)
|
|
|
|
self.assertEqual(truncated_sequence, sequence[:-2])
|
|
|
|
|
|
|
|
self.assertEqual(len(overflowing_tokens), 2 + stride)
|
|
|
|
self.assertEqual(overflowing_tokens, sequence[-(2 + stride) :])
|
|
|
|
else:
|
|
|
|
truncated_sequence = information["input_ids"]
|
|
|
|
overflowing_tokens = information["overflowing_tokens"]
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(len(truncated_sequence), total_length - 2)
|
|
|
|
self.assertEqual(truncated_sequence, sequence[:-2])
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(
|
|
|
|
len(overflowing_tokens), 0
|
|
|
|
) # No overflowing tokens when using 'longest' in python tokenizers
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
def test_maximum_encoding_length_pair_input(self):
|
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
# Build a sequence from our model's vocabulary
|
|
|
|
stride = 2
|
|
|
|
seq_0, ids = self.get_clean_sequence(tokenizer)
|
|
|
|
if len(ids) <= 2 + stride:
|
|
|
|
seq_0 = [s for s in seq_0 for _ in range(2 + stride)]
|
|
|
|
ids = [i for i in ids for _ in range(2 + stride)]
|
|
|
|
|
|
|
|
seq0_tokens = tokenizer.encode(seq_0, add_special_tokens=False)
|
|
|
|
assert len(seq0_tokens) > 2 + stride
|
|
|
|
|
|
|
|
seq_1 = "This is another sentence to be encoded."
|
|
|
|
seq1_tokens = tokenizer.encode(seq_1, add_special_tokens=False)
|
|
|
|
if len(seq0_tokens) == len(seq1_tokens):
|
|
|
|
seq1_tokens = seq1_tokens + seq1_tokens
|
|
|
|
seq_1 = tokenizer.decode(seq1_tokens, clean_up_tokenization_spaces=False)
|
|
|
|
seq1_tokens = tokenizer.encode(seq_1, add_special_tokens=False)
|
|
|
|
|
|
|
|
assert len(seq1_tokens) > 2 + stride
|
|
|
|
|
|
|
|
smallest = seq1_tokens if len(seq0_tokens) > len(seq1_tokens) else seq0_tokens
|
|
|
|
|
|
|
|
# We are not using the special tokens - a bit too hard to test all the tokenizers with this
|
|
|
|
# TODO try this again later
|
2020-06-23 19:36:57 +08:00
|
|
|
sequence = tokenizer.encode(seq_0, seq_1, add_special_tokens=False) # , add_prefix_space=False)
|
2020-06-16 05:12:51 +08:00
|
|
|
truncated_first_sequence = tokenizer.encode(seq_0, add_special_tokens=False)[:-2] + tokenizer.encode(
|
|
|
|
seq_1, add_special_tokens=False
|
|
|
|
)
|
|
|
|
truncated_second_sequence = (
|
|
|
|
tokenizer.encode(seq_0, add_special_tokens=False)
|
|
|
|
+ tokenizer.encode(seq_1, add_special_tokens=False)[:-2]
|
|
|
|
)
|
|
|
|
truncated_longest_sequence = (
|
|
|
|
truncated_first_sequence if len(seq0_tokens) > len(seq1_tokens) else truncated_second_sequence
|
|
|
|
)
|
|
|
|
|
|
|
|
overflow_first_sequence = tokenizer.encode(seq_0, add_special_tokens=False)[
|
|
|
|
-(2 + stride) :
|
|
|
|
] + tokenizer.encode(seq_1, add_special_tokens=False)
|
|
|
|
overflow_second_sequence = (
|
|
|
|
tokenizer.encode(seq_0, add_special_tokens=False)
|
|
|
|
+ tokenizer.encode(seq_1, add_special_tokens=False)[-(2 + stride) :]
|
|
|
|
)
|
|
|
|
overflow_longest_sequence = (
|
|
|
|
overflow_first_sequence if len(seq0_tokens) > len(seq1_tokens) else overflow_second_sequence
|
|
|
|
)
|
|
|
|
|
|
|
|
information = tokenizer.encode_plus(
|
|
|
|
seq_0,
|
|
|
|
seq_1,
|
|
|
|
max_length=len(sequence) - 2,
|
|
|
|
add_special_tokens=False,
|
|
|
|
stride=stride,
|
|
|
|
truncation="longest_first",
|
|
|
|
return_overflowing_tokens=True,
|
2020-06-23 19:36:57 +08:00
|
|
|
# add_prefix_space=False,
|
2020-06-16 05:12:51 +08:00
|
|
|
)
|
|
|
|
# Overflowing tokens are handled quite differently in slow and fast tokenizers
|
|
|
|
if isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
truncated_sequence = information["input_ids"][0]
|
|
|
|
overflowing_tokens = information["input_ids"][1]
|
|
|
|
self.assertEqual(len(information["input_ids"]), 2)
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
|
|
|
|
self.assertEqual(truncated_sequence, truncated_longest_sequence)
|
|
|
|
|
|
|
|
self.assertEqual(len(overflowing_tokens), 2 + stride + len(smallest))
|
|
|
|
self.assertEqual(overflowing_tokens, overflow_longest_sequence)
|
|
|
|
else:
|
|
|
|
truncated_sequence = information["input_ids"]
|
|
|
|
overflowing_tokens = information["overflowing_tokens"]
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
|
|
|
|
self.assertEqual(truncated_sequence, truncated_longest_sequence)
|
|
|
|
|
|
|
|
self.assertEqual(
|
|
|
|
len(overflowing_tokens), 0
|
|
|
|
) # No overflowing tokens when using 'longest' in python tokenizers
|
|
|
|
|
|
|
|
information_first_truncated = tokenizer.encode_plus(
|
|
|
|
seq_0,
|
|
|
|
seq_1,
|
|
|
|
max_length=len(sequence) - 2,
|
|
|
|
add_special_tokens=False,
|
|
|
|
stride=stride,
|
|
|
|
truncation=True,
|
|
|
|
return_overflowing_tokens=True,
|
2020-06-23 19:36:57 +08:00
|
|
|
# add_prefix_space=False,
|
2020-06-16 05:12:51 +08:00
|
|
|
)
|
|
|
|
# Overflowing tokens are handled quite differently in slow and fast tokenizers
|
|
|
|
if isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
truncated_sequence = information_first_truncated["input_ids"][0]
|
|
|
|
overflowing_tokens = information_first_truncated["input_ids"][1]
|
|
|
|
self.assertEqual(len(information_first_truncated["input_ids"]), 2)
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
|
|
|
|
self.assertEqual(truncated_sequence, truncated_first_sequence)
|
|
|
|
|
|
|
|
self.assertEqual(len(overflowing_tokens), 2 + stride + len(seq1_tokens))
|
|
|
|
self.assertEqual(overflowing_tokens, overflow_first_sequence)
|
|
|
|
else:
|
|
|
|
truncated_sequence = information_first_truncated["input_ids"]
|
|
|
|
overflowing_tokens = information_first_truncated["overflowing_tokens"]
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
|
|
|
|
self.assertEqual(truncated_sequence, truncated_first_sequence)
|
|
|
|
|
|
|
|
self.assertEqual(len(overflowing_tokens), 2 + stride)
|
|
|
|
self.assertEqual(overflowing_tokens, seq0_tokens[-(2 + stride) :])
|
|
|
|
|
|
|
|
information_second_truncated = tokenizer.encode_plus(
|
|
|
|
seq_0,
|
|
|
|
seq_1,
|
|
|
|
max_length=len(sequence) - 2,
|
|
|
|
add_special_tokens=False,
|
|
|
|
stride=stride,
|
|
|
|
truncation="only_second",
|
|
|
|
return_overflowing_tokens=True,
|
2020-06-23 19:36:57 +08:00
|
|
|
# add_prefix_space=False,
|
2020-06-16 05:12:51 +08:00
|
|
|
)
|
|
|
|
# Overflowing tokens are handled quite differently in slow and fast tokenizers
|
|
|
|
if isinstance(tokenizer, PreTrainedTokenizerFast):
|
|
|
|
truncated_sequence = information_second_truncated["input_ids"][0]
|
|
|
|
overflowing_tokens = information_second_truncated["input_ids"][1]
|
|
|
|
self.assertEqual(len(information_second_truncated["input_ids"]), 2)
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
|
|
|
|
self.assertEqual(truncated_sequence, truncated_second_sequence)
|
|
|
|
|
|
|
|
self.assertEqual(len(overflowing_tokens), 2 + stride + len(seq0_tokens))
|
|
|
|
self.assertEqual(overflowing_tokens, overflow_second_sequence)
|
|
|
|
else:
|
|
|
|
truncated_sequence = information_second_truncated["input_ids"]
|
|
|
|
overflowing_tokens = information_second_truncated["overflowing_tokens"]
|
|
|
|
|
|
|
|
self.assertEqual(len(truncated_sequence), len(sequence) - 2)
|
|
|
|
self.assertEqual(truncated_sequence, truncated_second_sequence)
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(len(overflowing_tokens), 2 + stride)
|
|
|
|
self.assertEqual(overflowing_tokens, seq1_tokens[-(2 + stride) :])
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# def test_encode_input_type(self):
|
|
|
|
# tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
# for tokenizer in tokenizers:
|
|
|
|
# with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
# sequence = "Let's encode this sequence"
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# tokens = sequence.split() # tokenizer.tokenize(sequence)
|
|
|
|
# # input_ids = tokenizer.convert_tokens_to_ids(tokens)
|
|
|
|
# formatted_input = tokenizer.encode(sequence, add_special_tokens=True, add_prefix_space=False)
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# self.assertEqual(
|
|
|
|
# tokenizer.encode(tokens, is_pretokenized=True, add_special_tokens=True), formatted_input
|
|
|
|
# )
|
|
|
|
# # This is not supported with the Rust tokenizers
|
|
|
|
# # self.assertEqual(tokenizer.encode(input_ids, add_special_tokens=True), formatted_input)
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
def test_swap_special_token(self):
|
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
mask = "<mask>"
|
|
|
|
sequence = "Encode this sequence"
|
|
|
|
sequence_masked_0 = "Encode <mask> sequence"
|
|
|
|
sequence_masked_1 = "<mask> this sequence"
|
|
|
|
|
|
|
|
# Add tokens so that masked token isn't split
|
|
|
|
tokenizer.add_tokens(sequence.split())
|
|
|
|
tokenizer.add_special_tokens({"mask_token": mask})
|
|
|
|
mask_ind = tokenizer.convert_tokens_to_ids(mask)
|
|
|
|
encoded = tokenizer.encode(sequence, add_special_tokens=False)
|
|
|
|
|
|
|
|
# Test first masked sequence
|
|
|
|
encoded_masked = tokenizer.encode(sequence_masked_0, add_special_tokens=False)
|
|
|
|
mask_loc = encoded_masked.index(mask_ind)
|
|
|
|
encoded_masked[mask_loc] = encoded[mask_loc]
|
|
|
|
|
|
|
|
self.assertEqual(encoded_masked, encoded)
|
|
|
|
|
|
|
|
# Test second masked sequence
|
|
|
|
encoded_masked = tokenizer.encode(sequence_masked_1, add_special_tokens=False)
|
|
|
|
mask_loc = encoded_masked.index(mask_ind)
|
|
|
|
encoded_masked[mask_loc] = encoded[mask_loc]
|
|
|
|
|
|
|
|
self.assertEqual(encoded_masked, encoded)
|
2020-02-14 02:29:43 +08:00
|
|
|
|
2019-12-22 22:34:15 +08:00
|
|
|
def test_special_tokens_mask(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequence_0 = "Encode this."
|
|
|
|
# Testing single inputs
|
|
|
|
encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
|
|
|
|
encoded_sequence_dict = tokenizer.encode_plus(
|
2020-06-23 19:36:57 +08:00
|
|
|
sequence_0, add_special_tokens=True, return_special_tokens_mask=True # , add_prefix_space=False
|
2020-06-16 05:12:51 +08:00
|
|
|
)
|
|
|
|
encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
|
|
|
|
special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
|
|
|
|
self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
|
|
|
|
|
|
|
|
filtered_sequence = [x for i, x in enumerate(encoded_sequence_w_special) if not special_tokens_mask[i]]
|
|
|
|
self.assertEqual(encoded_sequence, filtered_sequence)
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-05-19 22:46:55 +08:00
|
|
|
def test_special_tokens_mask_input_pairs(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequence_0 = "Encode this."
|
|
|
|
sequence_1 = "This one too please."
|
|
|
|
encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
|
|
|
|
encoded_sequence += tokenizer.encode(sequence_1, add_special_tokens=False)
|
|
|
|
encoded_sequence_dict = tokenizer.encode_plus(
|
|
|
|
sequence_0,
|
|
|
|
sequence_1,
|
|
|
|
add_special_tokens=True,
|
|
|
|
return_special_tokens_mask=True,
|
2020-06-23 19:36:57 +08:00
|
|
|
# add_prefix_space=False,
|
2020-06-16 05:12:51 +08:00
|
|
|
)
|
|
|
|
encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
|
|
|
|
special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
|
|
|
|
self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
|
|
|
|
|
|
|
|
filtered_sequence = [
|
|
|
|
(x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)
|
|
|
|
]
|
|
|
|
filtered_sequence = [x for x in filtered_sequence if x is not None]
|
|
|
|
self.assertEqual(encoded_sequence, filtered_sequence)
|
2019-12-22 22:34:15 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
def test_right_and_left_padding(self):
|
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequence = "Sequence"
|
|
|
|
padding_size = 10
|
|
|
|
|
|
|
|
# check correct behaviour if no pad_token_id exists and add it eventually
|
|
|
|
self._check_no_pad_token_padding(tokenizer, sequence)
|
|
|
|
|
|
|
|
padding_idx = tokenizer.pad_token_id
|
|
|
|
|
|
|
|
# RIGHT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
encoded_sequence = tokenizer.encode(sequence)
|
|
|
|
sequence_length = len(encoded_sequence)
|
|
|
|
padded_sequence = tokenizer.encode(
|
|
|
|
sequence, max_length=sequence_length + padding_size, padding="max_length"
|
|
|
|
)
|
|
|
|
padded_sequence_length = len(padded_sequence)
|
|
|
|
assert sequence_length + padding_size == padded_sequence_length
|
|
|
|
assert encoded_sequence + [padding_idx] * padding_size == padded_sequence
|
|
|
|
|
|
|
|
# LEFT PADDING - Check that it correctly pads when a maximum length is specified along with the padding flag set to True
|
|
|
|
tokenizer.padding_side = "left"
|
|
|
|
encoded_sequence = tokenizer.encode(sequence)
|
|
|
|
sequence_length = len(encoded_sequence)
|
|
|
|
padded_sequence = tokenizer.encode(
|
|
|
|
sequence, max_length=sequence_length + padding_size, padding="max_length"
|
|
|
|
)
|
|
|
|
padded_sequence_length = len(padded_sequence)
|
|
|
|
assert sequence_length + padding_size == padded_sequence_length
|
|
|
|
assert [padding_idx] * padding_size + encoded_sequence == padded_sequence
|
|
|
|
|
|
|
|
# RIGHT & LEFT PADDING - Check that nothing is done for 'longest' and 'no_padding'
|
|
|
|
encoded_sequence = tokenizer.encode(sequence)
|
|
|
|
sequence_length = len(encoded_sequence)
|
|
|
|
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
padded_sequence_right = tokenizer.encode(sequence, padding=True)
|
|
|
|
padded_sequence_right_length = len(padded_sequence_right)
|
|
|
|
assert sequence_length == padded_sequence_right_length
|
|
|
|
assert encoded_sequence == padded_sequence_right
|
|
|
|
|
|
|
|
tokenizer.padding_side = "left"
|
|
|
|
padded_sequence_left = tokenizer.encode(sequence, padding="longest")
|
|
|
|
padded_sequence_left_length = len(padded_sequence_left)
|
|
|
|
assert sequence_length == padded_sequence_left_length
|
|
|
|
assert encoded_sequence == padded_sequence_left
|
|
|
|
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
padded_sequence_right = tokenizer.encode(sequence)
|
|
|
|
padded_sequence_right_length = len(padded_sequence_right)
|
|
|
|
assert sequence_length == padded_sequence_right_length
|
|
|
|
assert encoded_sequence == padded_sequence_right
|
|
|
|
|
|
|
|
tokenizer.padding_side = "left"
|
|
|
|
padded_sequence_left = tokenizer.encode(sequence, padding=False)
|
|
|
|
padded_sequence_left_length = len(padded_sequence_left)
|
|
|
|
assert sequence_length == padded_sequence_left_length
|
|
|
|
assert encoded_sequence == padded_sequence_left
|
2019-12-22 22:34:15 +08:00
|
|
|
|
|
|
|
def test_padding_to_max_length(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
""" We keep this test for backward compatibility but it should be remove when `pad_to_max_length` will e deprecated
|
|
|
|
"""
|
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequence = "Sequence"
|
|
|
|
padding_size = 10
|
|
|
|
|
|
|
|
# check correct behaviour if no pad_token_id exists and add it eventually
|
|
|
|
self._check_no_pad_token_padding(tokenizer, sequence)
|
|
|
|
|
|
|
|
padding_idx = tokenizer.pad_token_id
|
|
|
|
|
|
|
|
# Check that it correctly pads when a maximum length is specified along with the padding flag set to True
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
encoded_sequence = tokenizer.encode(sequence)
|
|
|
|
sequence_length = len(encoded_sequence)
|
|
|
|
padded_sequence = tokenizer.encode(
|
|
|
|
sequence, max_length=sequence_length + padding_size, pad_to_max_length=True
|
|
|
|
)
|
|
|
|
padded_sequence_length = len(padded_sequence)
|
|
|
|
assert sequence_length + padding_size == padded_sequence_length
|
|
|
|
assert encoded_sequence + [padding_idx] * padding_size == padded_sequence
|
|
|
|
|
|
|
|
# Check that nothing is done when a maximum length is not specified
|
|
|
|
encoded_sequence = tokenizer.encode(sequence)
|
|
|
|
sequence_length = len(encoded_sequence)
|
|
|
|
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
padded_sequence_right = tokenizer.encode(sequence, pad_to_max_length=True)
|
|
|
|
padded_sequence_right_length = len(padded_sequence_right)
|
|
|
|
assert sequence_length == padded_sequence_right_length
|
|
|
|
assert encoded_sequence == padded_sequence_right
|
2019-12-22 22:34:15 +08:00
|
|
|
|
|
|
|
def test_encode_plus_with_padding(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequence = "Sequence"
|
|
|
|
|
|
|
|
# check correct behaviour if no pad_token_id exists and add it eventually
|
|
|
|
self._check_no_pad_token_padding(tokenizer, sequence)
|
|
|
|
|
|
|
|
padding_size = 10
|
|
|
|
padding_idx = tokenizer.pad_token_id
|
|
|
|
token_type_padding_idx = tokenizer.pad_token_type_id
|
|
|
|
|
|
|
|
encoded_sequence = tokenizer.encode_plus(sequence, return_special_tokens_mask=True)
|
|
|
|
input_ids = encoded_sequence["input_ids"]
|
|
|
|
special_tokens_mask = encoded_sequence["special_tokens_mask"]
|
|
|
|
sequence_length = len(input_ids)
|
|
|
|
|
|
|
|
# Test 'longest' and 'no_padding' don't do anything
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
|
|
|
|
not_padded_sequence = tokenizer.encode_plus(sequence, padding=True, return_special_tokens_mask=True,)
|
|
|
|
not_padded_input_ids = not_padded_sequence["input_ids"]
|
|
|
|
|
|
|
|
not_padded_special_tokens_mask = not_padded_sequence["special_tokens_mask"]
|
|
|
|
not_padded_sequence_length = len(not_padded_input_ids)
|
|
|
|
|
|
|
|
assert sequence_length == not_padded_sequence_length
|
|
|
|
assert input_ids == not_padded_input_ids
|
|
|
|
assert special_tokens_mask == not_padded_special_tokens_mask
|
|
|
|
|
|
|
|
not_padded_sequence = tokenizer.encode_plus(sequence, padding=False, return_special_tokens_mask=True,)
|
|
|
|
not_padded_input_ids = not_padded_sequence["input_ids"]
|
|
|
|
|
|
|
|
not_padded_special_tokens_mask = not_padded_sequence["special_tokens_mask"]
|
|
|
|
not_padded_sequence_length = len(not_padded_input_ids)
|
|
|
|
|
|
|
|
assert sequence_length == not_padded_sequence_length
|
|
|
|
assert input_ids == not_padded_input_ids
|
|
|
|
assert special_tokens_mask == not_padded_special_tokens_mask
|
|
|
|
|
|
|
|
# Test right padding
|
|
|
|
tokenizer.padding_side = "right"
|
|
|
|
|
|
|
|
right_padded_sequence = tokenizer.encode_plus(
|
|
|
|
sequence,
|
|
|
|
max_length=sequence_length + padding_size,
|
|
|
|
padding="max_length",
|
|
|
|
return_special_tokens_mask=True,
|
|
|
|
)
|
|
|
|
right_padded_input_ids = right_padded_sequence["input_ids"]
|
|
|
|
|
|
|
|
right_padded_special_tokens_mask = right_padded_sequence["special_tokens_mask"]
|
|
|
|
right_padded_sequence_length = len(right_padded_input_ids)
|
|
|
|
|
|
|
|
assert sequence_length + padding_size == right_padded_sequence_length
|
|
|
|
assert input_ids + [padding_idx] * padding_size == right_padded_input_ids
|
|
|
|
assert special_tokens_mask + [1] * padding_size == right_padded_special_tokens_mask
|
|
|
|
|
|
|
|
# Test left padding
|
|
|
|
tokenizer.padding_side = "left"
|
|
|
|
left_padded_sequence = tokenizer.encode_plus(
|
|
|
|
sequence,
|
|
|
|
max_length=sequence_length + padding_size,
|
|
|
|
padding="max_length",
|
|
|
|
return_special_tokens_mask=True,
|
|
|
|
)
|
|
|
|
left_padded_input_ids = left_padded_sequence["input_ids"]
|
|
|
|
left_padded_special_tokens_mask = left_padded_sequence["special_tokens_mask"]
|
|
|
|
left_padded_sequence_length = len(left_padded_input_ids)
|
|
|
|
|
|
|
|
assert sequence_length + padding_size == left_padded_sequence_length
|
|
|
|
assert [padding_idx] * padding_size + input_ids == left_padded_input_ids
|
|
|
|
assert [1] * padding_size + special_tokens_mask == left_padded_special_tokens_mask
|
|
|
|
|
|
|
|
if "token_type_ids" in tokenizer.model_input_names:
|
|
|
|
token_type_ids = encoded_sequence["token_type_ids"]
|
|
|
|
left_padded_token_type_ids = left_padded_sequence["token_type_ids"]
|
|
|
|
right_padded_token_type_ids = right_padded_sequence["token_type_ids"]
|
|
|
|
|
|
|
|
assert token_type_ids + [token_type_padding_idx] * padding_size == right_padded_token_type_ids
|
|
|
|
assert [token_type_padding_idx] * padding_size + token_type_ids == left_padded_token_type_ids
|
|
|
|
|
|
|
|
if "attention_mask" in tokenizer.model_input_names:
|
|
|
|
attention_mask = encoded_sequence["attention_mask"]
|
|
|
|
right_padded_attention_mask = right_padded_sequence["attention_mask"]
|
|
|
|
left_padded_attention_mask = left_padded_sequence["attention_mask"]
|
|
|
|
|
|
|
|
assert attention_mask + [0] * padding_size == right_padded_attention_mask
|
|
|
|
assert [0] * padding_size + attention_mask == left_padded_attention_mask
|
2020-01-30 05:15:39 +08:00
|
|
|
|
|
|
|
def test_separate_tokenizers(self):
|
|
|
|
# This tests that tokenizers don't impact others. Unfortunately the case where it fails is when
|
|
|
|
# we're loading an S3 configuration from a pre-trained identifier, and we have no way of testing those today.
|
|
|
|
|
|
|
|
tokenizer = self.get_tokenizer(random_argument=True)
|
2020-01-30 05:29:20 +08:00
|
|
|
assert tokenizer.init_kwargs["random_argument"] is True
|
2020-01-30 05:15:39 +08:00
|
|
|
new_tokenizer = self.get_tokenizer(random_argument=False)
|
2020-01-30 05:29:20 +08:00
|
|
|
assert tokenizer.init_kwargs["random_argument"] is True
|
|
|
|
assert new_tokenizer.init_kwargs["random_argument"] is False
|
2020-02-21 04:25:46 +08:00
|
|
|
|
|
|
|
def test_get_vocab(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
vocab = tokenizer.get_vocab()
|
2020-02-21 04:25:46 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertIsInstance(vocab, dict)
|
|
|
|
self.assertEqual(len(vocab), len(tokenizer))
|
2020-02-21 04:25:46 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
for word, ind in vocab.items():
|
|
|
|
self.assertEqual(tokenizer.convert_tokens_to_ids(word), ind)
|
|
|
|
self.assertEqual(tokenizer.convert_ids_to_tokens(ind), word)
|
2020-02-21 04:25:46 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizer.add_tokens(["asdfasdfasdfasdf"])
|
|
|
|
vocab = tokenizer.get_vocab()
|
|
|
|
self.assertIsInstance(vocab, dict)
|
|
|
|
self.assertEqual(len(vocab), len(tokenizer))
|
2020-02-21 04:25:46 +08:00
|
|
|
|
2020-05-19 22:46:55 +08:00
|
|
|
def test_conversion_reversible(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
vocab = tokenizer.get_vocab()
|
|
|
|
for word, ind in vocab.items():
|
|
|
|
self.assertEqual(tokenizer.convert_tokens_to_ids(word), ind)
|
|
|
|
self.assertEqual(tokenizer.convert_ids_to_tokens(ind), word)
|
|
|
|
|
|
|
|
def test_call(self):
|
|
|
|
# Tests that all call wrap to encode_plus and batch_encode_plus
|
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequences = [
|
|
|
|
"Testing batch encode plus",
|
|
|
|
"Testing batch encode plus with different sequence lengths",
|
|
|
|
"Testing batch encode plus with different sequence lengths correctly pads",
|
|
|
|
]
|
|
|
|
|
|
|
|
# Test not batched
|
|
|
|
encoded_sequences_1 = tokenizer.encode_plus(sequences[0])
|
|
|
|
encoded_sequences_2 = tokenizer(sequences[0])
|
|
|
|
self.assertEqual(encoded_sequences_1, encoded_sequences_2)
|
|
|
|
|
|
|
|
# Test not batched pairs
|
|
|
|
encoded_sequences_1 = tokenizer.encode_plus(sequences[0], sequences[1])
|
|
|
|
encoded_sequences_2 = tokenizer(sequences[0], sequences[1])
|
|
|
|
self.assertEqual(encoded_sequences_1, encoded_sequences_2)
|
|
|
|
|
|
|
|
# Test batched
|
|
|
|
encoded_sequences_1 = tokenizer.batch_encode_plus(sequences)
|
|
|
|
encoded_sequences_2 = tokenizer(sequences)
|
|
|
|
self.assertEqual(encoded_sequences_1, encoded_sequences_2)
|
|
|
|
|
|
|
|
# Test batched pairs
|
|
|
|
encoded_sequences_1 = tokenizer.batch_encode_plus(list(zip(sequences, sequences)))
|
|
|
|
encoded_sequences_2 = tokenizer(sequences, sequences)
|
|
|
|
self.assertEqual(encoded_sequences_1, encoded_sequences_2)
|
2020-02-25 01:09:46 +08:00
|
|
|
|
|
|
|
def test_batch_encode_plus_batch_sequence_length(self):
|
|
|
|
# Tests that all encoded values have the correct size
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequences = [
|
|
|
|
"Testing batch encode plus",
|
|
|
|
"Testing batch encode plus with different sequence lengths",
|
|
|
|
"Testing batch encode plus with different sequence lengths correctly pads",
|
|
|
|
]
|
|
|
|
|
|
|
|
encoded_sequences = [tokenizer.encode_plus(sequence) for sequence in sequences]
|
|
|
|
encoded_sequences_batch = tokenizer.batch_encode_plus(sequences, padding=False)
|
|
|
|
self.assertListEqual(
|
|
|
|
encoded_sequences, self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch)
|
|
|
|
)
|
|
|
|
|
|
|
|
maximum_length = len(
|
|
|
|
max([encoded_sequence["input_ids"] for encoded_sequence in encoded_sequences], key=len)
|
|
|
|
)
|
|
|
|
|
|
|
|
# check correct behaviour if no pad_token_id exists and add it eventually
|
|
|
|
self._check_no_pad_token_padding(tokenizer, sequences)
|
|
|
|
|
|
|
|
encoded_sequences_padded = [
|
|
|
|
tokenizer.encode_plus(sequence, max_length=maximum_length, padding="max_length")
|
|
|
|
for sequence in sequences
|
|
|
|
]
|
|
|
|
|
|
|
|
encoded_sequences_batch_padded = tokenizer.batch_encode_plus(sequences, padding=True)
|
|
|
|
self.assertListEqual(
|
|
|
|
encoded_sequences_padded,
|
|
|
|
self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch_padded),
|
|
|
|
)
|
|
|
|
|
|
|
|
# check 'longest' is unsensitive to a max length
|
|
|
|
encoded_sequences_batch_padded_1 = tokenizer.batch_encode_plus(sequences, padding=True)
|
|
|
|
encoded_sequences_batch_padded_2 = tokenizer.batch_encode_plus(
|
|
|
|
sequences, max_length=maximum_length + 10, padding="longest"
|
|
|
|
)
|
|
|
|
for key in encoded_sequences_batch_padded_1.keys():
|
|
|
|
self.assertListEqual(
|
|
|
|
encoded_sequences_batch_padded_1[key], encoded_sequences_batch_padded_2[key],
|
|
|
|
)
|
|
|
|
|
|
|
|
# check 'no_padding' is unsensitive to a max length
|
|
|
|
encoded_sequences_batch_padded_1 = tokenizer.batch_encode_plus(sequences, padding=False)
|
|
|
|
encoded_sequences_batch_padded_2 = tokenizer.batch_encode_plus(
|
|
|
|
sequences, max_length=maximum_length + 10, padding=False
|
|
|
|
)
|
|
|
|
for key in encoded_sequences_batch_padded_1.keys():
|
|
|
|
self.assertListEqual(
|
|
|
|
encoded_sequences_batch_padded_1[key], encoded_sequences_batch_padded_2[key],
|
|
|
|
)
|
2020-02-25 01:09:46 +08:00
|
|
|
|
|
|
|
def test_batch_encode_plus_padding(self):
|
|
|
|
# Test that padded sequences are equivalent between batch_encode_plus and encode_plus
|
|
|
|
|
|
|
|
# Right padding tests
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequences = [
|
|
|
|
"Testing batch encode plus",
|
|
|
|
"Testing batch encode plus with different sequence lengths",
|
|
|
|
"Testing batch encode plus with different sequence lengths correctly pads",
|
|
|
|
]
|
|
|
|
|
|
|
|
max_length = 100
|
|
|
|
|
|
|
|
# check correct behaviour if no pad_token_id exists and add it eventually
|
|
|
|
self._check_no_pad_token_padding(tokenizer, sequences)
|
|
|
|
|
|
|
|
encoded_sequences = [
|
|
|
|
tokenizer.encode_plus(sequence, max_length=max_length, padding="max_length")
|
|
|
|
for sequence in sequences
|
|
|
|
]
|
|
|
|
encoded_sequences_batch = tokenizer.batch_encode_plus(
|
|
|
|
sequences, max_length=max_length, padding="max_length"
|
|
|
|
)
|
|
|
|
self.assertListEqual(
|
|
|
|
encoded_sequences, self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch)
|
|
|
|
)
|
2020-02-25 01:09:46 +08:00
|
|
|
|
|
|
|
# Left padding tests
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
tokenizer.padding_side = "left"
|
|
|
|
sequences = [
|
|
|
|
"Testing batch encode plus",
|
|
|
|
"Testing batch encode plus with different sequence lengths",
|
|
|
|
"Testing batch encode plus with different sequence lengths correctly pads",
|
|
|
|
]
|
|
|
|
|
|
|
|
max_length = 100
|
|
|
|
|
|
|
|
# check correct behaviour if no pad_token_id exists and add it eventually
|
|
|
|
self._check_no_pad_token_padding(tokenizer, sequences)
|
|
|
|
|
|
|
|
encoded_sequences = [
|
|
|
|
tokenizer.encode_plus(sequence, max_length=max_length, padding="max_length")
|
|
|
|
for sequence in sequences
|
|
|
|
]
|
|
|
|
encoded_sequences_batch = tokenizer.batch_encode_plus(
|
|
|
|
sequences, max_length=max_length, padding="max_length"
|
|
|
|
)
|
|
|
|
self.assertListEqual(
|
|
|
|
encoded_sequences, self.convert_batch_encode_plus_format_to_encode_plus(encoded_sequences_batch)
|
|
|
|
)
|
|
|
|
|
|
|
|
def test_pretokenized_inputs(self):
|
|
|
|
# Test when inputs are pretokenized
|
|
|
|
|
2020-06-23 19:36:57 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False) # , add_prefix_space=True)
|
2020-06-16 05:12:51 +08:00
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
|
|
|
|
# Prepare a sequence from our tokenizer vocabulary
|
|
|
|
sequence, ids = self.get_clean_sequence(tokenizer, with_prefix_space=True, max_length=20)
|
|
|
|
# sequence = " " + sequence # To be sure the byte-level tokenizers are feeling good
|
|
|
|
token_sequence = sequence.split()
|
|
|
|
# sequence_no_prefix_space = sequence.strip()
|
|
|
|
|
|
|
|
# Test encode for pretokenized inputs
|
|
|
|
output = tokenizer.encode(token_sequence, is_pretokenized=True, add_special_tokens=False)
|
|
|
|
output_sequence = tokenizer.encode(sequence, add_special_tokens=False)
|
|
|
|
self.assertEqual(output, output_sequence)
|
|
|
|
|
|
|
|
output = tokenizer.encode(token_sequence, is_pretokenized=True, add_special_tokens=True)
|
|
|
|
output_sequence = tokenizer.encode(sequence, add_special_tokens=True)
|
|
|
|
self.assertEqual(output, output_sequence)
|
|
|
|
|
|
|
|
# Test encode_plus for pretokenized inputs
|
|
|
|
output = tokenizer.encode_plus(token_sequence, is_pretokenized=True, add_special_tokens=False)
|
|
|
|
output_sequence = tokenizer.encode_plus(sequence, add_special_tokens=False)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
output = tokenizer.encode_plus(token_sequence, is_pretokenized=True, add_special_tokens=True)
|
|
|
|
output_sequence = tokenizer.encode_plus(sequence, add_special_tokens=True)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
|
|
|
|
# Test batch_encode_plus for pretokenized inputs
|
|
|
|
sequence_batch = [sequence.strip()] * 2 + [sequence.strip() + " " + sequence.strip()]
|
|
|
|
token_sequence_batch = [s.split() for s in sequence_batch]
|
|
|
|
sequence_batch_cleaned_up_spaces = [" " + " ".join(s) for s in token_sequence_batch]
|
|
|
|
|
|
|
|
output = tokenizer.batch_encode_plus(
|
|
|
|
token_sequence_batch, is_pretokenized=True, add_special_tokens=False
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.batch_encode_plus(
|
|
|
|
sequence_batch_cleaned_up_spaces, add_special_tokens=False
|
|
|
|
)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
output = tokenizer.batch_encode_plus(
|
|
|
|
token_sequence_batch, is_pretokenized=True, add_special_tokens=True
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.batch_encode_plus(
|
|
|
|
sequence_batch_cleaned_up_spaces, add_special_tokens=True
|
|
|
|
)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
|
|
|
|
# Test encode for pretokenized inputs pairs
|
|
|
|
output = tokenizer.encode(
|
|
|
|
token_sequence, token_sequence, is_pretokenized=True, add_special_tokens=False
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.encode(sequence, sequence, add_special_tokens=False)
|
|
|
|
self.assertEqual(output, output_sequence)
|
|
|
|
output = tokenizer.encode(
|
|
|
|
token_sequence, token_sequence, is_pretokenized=True, add_special_tokens=True
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.encode(sequence, sequence, add_special_tokens=True)
|
|
|
|
self.assertEqual(output, output_sequence)
|
|
|
|
|
|
|
|
# Test encode_plus for pretokenized inputs pairs
|
|
|
|
output = tokenizer.encode_plus(
|
|
|
|
token_sequence, token_sequence, is_pretokenized=True, add_special_tokens=False
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.encode_plus(sequence, sequence, add_special_tokens=False)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
output = tokenizer.encode_plus(
|
|
|
|
token_sequence, token_sequence, is_pretokenized=True, add_special_tokens=True
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.encode_plus(sequence, sequence, add_special_tokens=True)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
|
|
|
|
# Test batch_encode_plus for pretokenized inputs pairs
|
|
|
|
sequence_pair_batch = [(sequence.strip(), sequence.strip())] * 2 + [
|
|
|
|
(sequence.strip() + " " + sequence.strip(), sequence.strip())
|
|
|
|
]
|
|
|
|
token_sequence_pair_batch = [tuple(s.split() for s in pair) for pair in sequence_pair_batch]
|
|
|
|
sequence_pair_batch_cleaned_up_spaces = [
|
|
|
|
tuple(" " + " ".join(s) for s in pair) for pair in token_sequence_pair_batch
|
|
|
|
]
|
|
|
|
|
|
|
|
output = tokenizer.batch_encode_plus(
|
|
|
|
token_sequence_pair_batch, is_pretokenized=True, add_special_tokens=False
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.batch_encode_plus(
|
|
|
|
sequence_pair_batch_cleaned_up_spaces, add_special_tokens=False
|
|
|
|
)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
|
|
|
output = tokenizer.batch_encode_plus(
|
|
|
|
token_sequence_pair_batch, is_pretokenized=True, add_special_tokens=True
|
|
|
|
)
|
|
|
|
output_sequence = tokenizer.batch_encode_plus(
|
|
|
|
sequence_pair_batch_cleaned_up_spaces, add_special_tokens=True
|
|
|
|
)
|
|
|
|
for key in output.keys():
|
|
|
|
self.assertEqual(output[key], output_sequence[key])
|
2020-02-25 01:09:46 +08:00
|
|
|
|
|
|
|
@require_torch
|
|
|
|
@require_tf
|
|
|
|
def test_batch_encode_plus_tensors(self):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
sequences = [
|
|
|
|
"Testing batch encode plus",
|
|
|
|
"Testing batch encode plus with different sequence lengths",
|
|
|
|
"Testing batch encode plus with different sequence lengths correctly pads",
|
|
|
|
]
|
|
|
|
|
|
|
|
# A Tensor cannot be build by sequences which are not the same size
|
|
|
|
self.assertRaises(ValueError, tokenizer.batch_encode_plus, sequences, return_tensors="pt")
|
|
|
|
self.assertRaises(ValueError, tokenizer.batch_encode_plus, sequences, return_tensors="tf")
|
|
|
|
|
|
|
|
if tokenizer.pad_token_id is None:
|
|
|
|
self.assertRaises(
|
|
|
|
ValueError, tokenizer.batch_encode_plus, sequences, padding=True, return_tensors="pt",
|
|
|
|
)
|
|
|
|
self.assertRaises(
|
|
|
|
ValueError, tokenizer.batch_encode_plus, sequences, padding="longest", return_tensors="tf",
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
pytorch_tensor = tokenizer.batch_encode_plus(sequences, padding=True, return_tensors="pt")
|
|
|
|
tensorflow_tensor = tokenizer.batch_encode_plus(sequences, padding="longest", return_tensors="tf")
|
|
|
|
encoded_sequences = tokenizer.batch_encode_plus(sequences, padding=True)
|
2020-02-25 01:09:46 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
for key in encoded_sequences.keys():
|
|
|
|
pytorch_value = pytorch_tensor[key].tolist()
|
|
|
|
tensorflow_value = tensorflow_tensor[key].numpy().tolist()
|
|
|
|
encoded_value = encoded_sequences[key]
|
2020-02-25 01:09:46 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
self.assertEqual(pytorch_value, tensorflow_value, encoded_value)
|
2020-03-02 23:53:55 +08:00
|
|
|
|
|
|
|
def _check_no_pad_token_padding(self, tokenizer, sequences):
|
|
|
|
# if tokenizer does not have pad_token_id, an error should be thrown
|
|
|
|
if tokenizer.pad_token_id is None:
|
|
|
|
with self.assertRaises(ValueError):
|
|
|
|
if isinstance(sequences, list):
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizer.batch_encode_plus(sequences, padding="longest")
|
2020-03-02 23:53:55 +08:00
|
|
|
else:
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizer.encode_plus(sequences, padding=True)
|
2020-03-02 23:53:55 +08:00
|
|
|
|
|
|
|
# add pad_token_id to pass subsequent tests
|
|
|
|
tokenizer.add_special_tokens({"pad_token": "<PAD>"})
|
2020-04-09 04:22:44 +08:00
|
|
|
|
|
|
|
@require_torch
|
|
|
|
def test_torch_encode_plus_sent_to_model(self):
|
2020-05-19 22:46:55 +08:00
|
|
|
import torch
|
2020-04-09 04:22:44 +08:00
|
|
|
from transformers import MODEL_MAPPING, TOKENIZER_MAPPING
|
|
|
|
|
|
|
|
MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(MODEL_MAPPING, TOKENIZER_MAPPING)
|
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
|
|
|
|
return
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
|
|
|
|
config = config_class()
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
if config.is_encoder_decoder or config.pad_token_id is None:
|
|
|
|
return
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
model = model_class(config)
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# Make sure the model contains at least the full vocabulary size in its embedding matrix
|
|
|
|
is_using_common_embeddings = hasattr(model.get_input_embeddings(), "weight")
|
|
|
|
assert (
|
|
|
|
(model.get_input_embeddings().weight.shape[0] >= len(tokenizer))
|
|
|
|
if is_using_common_embeddings
|
|
|
|
else True
|
|
|
|
)
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# Build sequence
|
|
|
|
first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
|
|
|
|
sequence = " ".join(first_ten_tokens)
|
|
|
|
encoded_sequence = tokenizer.encode_plus(sequence, return_tensors="pt")
|
|
|
|
batch_encoded_sequence = tokenizer.batch_encode_plus([sequence, sequence], return_tensors="pt")
|
|
|
|
# This should not fail
|
2020-05-19 22:46:55 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
with torch.no_grad(): # saves some time
|
|
|
|
model(**encoded_sequence)
|
|
|
|
model(**batch_encoded_sequence)
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# if self.test_rust_tokenizer:
|
|
|
|
# fast_tokenizer = self.get_rust_tokenizer()
|
|
|
|
# encoded_sequence_fast = fast_tokenizer.encode_plus(sequence, return_tensors="pt")
|
|
|
|
# batch_encoded_sequence_fast = fast_tokenizer.batch_encode_plus([sequence, sequence], return_tensors="pt")
|
|
|
|
# # This should not fail
|
|
|
|
# model(**encoded_sequence_fast)
|
|
|
|
# model(**batch_encoded_sequence_fast)
|
2020-04-09 04:22:44 +08:00
|
|
|
|
|
|
|
@require_tf
|
|
|
|
def test_tf_encode_plus_sent_to_model(self):
|
|
|
|
from transformers import TF_MODEL_MAPPING, TOKENIZER_MAPPING
|
|
|
|
|
|
|
|
MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(TF_MODEL_MAPPING, TOKENIZER_MAPPING)
|
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
tokenizers = self.get_tokenizers(do_lower_case=False)
|
|
|
|
for tokenizer in tokenizers:
|
|
|
|
with self.subTest(f"{tokenizer.__class__.__name__}"):
|
|
|
|
if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
|
|
|
|
return
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
|
|
|
|
config = config_class()
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
if config.is_encoder_decoder or config.pad_token_id is None:
|
|
|
|
return
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
model = model_class(config)
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# Make sure the model contains at least the full vocabulary size in its embedding matrix
|
|
|
|
assert model.config.vocab_size >= len(tokenizer)
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# Build sequence
|
|
|
|
first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
|
|
|
|
sequence = " ".join(first_ten_tokens)
|
|
|
|
encoded_sequence = tokenizer.encode_plus(sequence, return_tensors="tf")
|
|
|
|
batch_encoded_sequence = tokenizer.batch_encode_plus([sequence, sequence], return_tensors="tf")
|
2020-04-09 04:22:44 +08:00
|
|
|
|
2020-06-16 05:12:51 +08:00
|
|
|
# This should not fail
|
|
|
|
model(encoded_sequence)
|
|
|
|
model(batch_encoded_sequence)
|
2020-06-04 12:57:01 +08:00
|
|
|
|
|
|
|
# TODO: Check if require_torch is the best to test for numpy here ... Maybe move to require_flax when available
|
|
|
|
@require_torch
|
|
|
|
def test_np_encode_plus_sent_to_model(self):
|
|
|
|
from transformers import MODEL_MAPPING, TOKENIZER_MAPPING
|
|
|
|
|
|
|
|
MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(MODEL_MAPPING, TOKENIZER_MAPPING)
|
|
|
|
|
|
|
|
tokenizer = self.get_tokenizer()
|
|
|
|
if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
|
|
|
|
return
|
|
|
|
|
|
|
|
config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
|
|
|
|
config = config_class()
|
|
|
|
|
|
|
|
if config.is_encoder_decoder or config.pad_token_id is None:
|
|
|
|
return
|
|
|
|
|
|
|
|
# Build sequence
|
|
|
|
first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
|
|
|
|
sequence = " ".join(first_ten_tokens)
|
|
|
|
encoded_sequence = tokenizer.encode_plus(sequence, return_tensors="np")
|
|
|
|
batch_encoded_sequence = tokenizer.batch_encode_plus([sequence, sequence], return_tensors="np")
|
|
|
|
|
|
|
|
# TODO: add forward through JAX/Flax when PR is merged
|
|
|
|
# This is currently here to make flake8 happy !
|
|
|
|
if encoded_sequence is None:
|
|
|
|
raise ValueError("Cannot convert list to numpy tensor on encode_plus()")
|
|
|
|
|
|
|
|
if batch_encoded_sequence is None:
|
|
|
|
raise ValueError("Cannot convert list to numpy tensor on batch_encode_plus()")
|
|
|
|
|
|
|
|
if self.test_rust_tokenizer:
|
|
|
|
fast_tokenizer = self.get_rust_tokenizer()
|
|
|
|
encoded_sequence_fast = fast_tokenizer.encode_plus(sequence, return_tensors="np")
|
|
|
|
batch_encoded_sequence_fast = fast_tokenizer.batch_encode_plus([sequence, sequence], return_tensors="np")
|
|
|
|
|
|
|
|
# TODO: add forward through JAX/Flax when PR is merged
|
|
|
|
# This is currently here to make flake8 happy !
|
|
|
|
if encoded_sequence_fast is None:
|
|
|
|
raise ValueError("Cannot convert list to numpy tensor on encode_plus() (fast)")
|
|
|
|
|
|
|
|
if batch_encoded_sequence_fast is None:
|
|
|
|
raise ValueError("Cannot convert list to numpy tensor on batch_encode_plus() (fast)")
|