riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-31 02:02:21 +06:00
Code Issues Actions 23 Packages Projects Releases Wiki Activity

update the arguments add_prefix_space and trim_offsets in backend_tokenizer.post_processor of RobertaTokenizerFast (#14752)

Browse Source 
* add tests

* change post-processor, pre-tokenizer and decoder (can't update decoder)

* update test (remove decoder which doesn't depend on trim and add_prefix)

* just update the post_processor

* fix change

* `trim_offsets` has no influence on `pre_tokenizer`

* remove a test that need some input from the `tokenizers` lib maintainers

* format

* add new test offsets roberta

* polish comments
This commit is contained in:
SaulLu 2021-12-22 10:51:55 +01:00 committed by GitHub
parent ec3567fe20
commit c94c1b8967
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 137 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
src/transformers/models/roberta/tokenization_roberta_fast.py
View File

@ -13,9 +13,11 @@
# See the License for the specific language governing permissions and
# limitations under the License.
"""Fast Tokenization classes for RoBERTa."""
import json
from typing import List, Optional
from tokenizers import processors
from ...tokenization_utils_base import AddedToken
from ...utils import logging
from ..gpt2.tokenization_gpt2_fast import GPT2TokenizerFast
@ -162,6 +164,7 @@ class RobertaTokenizerFast(GPT2TokenizerFast):
pad_token="<pad>",
mask_token="<mask>",
add_prefix_space=False,
trim_offsets=True,
**kwargs
):
super().__init__(
@ -177,9 +180,37 @@ class RobertaTokenizerFast(GPT2TokenizerFast):
pad_token=pad_token,
mask_token=mask_token,
add_prefix_space=add_prefix_space,
trim_offsets=trim_offsets,
**kwargs,
)
# the pre_tokenizer is already updated in the GPT2TokenizerFast `__init__`
tokenizer_component = "post_processor"
tokenizer_component_instance = getattr(self.backend_tokenizer, tokenizer_component, None)
if tokenizer_component_instance:
state = json.loads(tokenizer_component_instance.__getstate__())
# The lists 'sep' and 'cls' must be cased in tuples for the object `post_processor_class`
if "sep" in state:
state["sep"] = tuple(state["sep"])
if "cls" in state:
state["cls"] = tuple(state["cls"])
changes_to_apply = False
if state.get("add_prefix_space", add_prefix_space) != add_prefix_space:
state["add_prefix_space"] = add_prefix_space
changes_to_apply = True
if state.get("trim_offsets", trim_offsets) != trim_offsets:
state["trim_offsets"] = trim_offsets
changes_to_apply = True
if changes_to_apply:
component_class = getattr(processors, state.pop("type"))
new_value = component_class(**state)
setattr(self.backend_tokenizer, tokenizer_component, new_value)
@property
def mask_token(self) -> str:
"""

105
tests/test_tokenization_roberta.py
View File

@ -14,6 +14,7 @@
# limitations under the License.
import itertools
import json
import os
import unittest
@ -196,3 +197,107 @@ class RobertaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
self.assertSequenceEqual(
tokens_r_str, ["<s>", "A", ",", "<mask>", "ĠAllen", "N", "LP", "Ġsentence", ".", "</s>"]
)
def test_change_add_prefix_space_and_trim_offsets_args(self):
for trim_offsets, add_prefix_space in itertools.product([True, False], repeat=2):
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
self.tmpdirname, use_fast=True, add_prefix_space=add_prefix_space, trim_offsets=trim_offsets
)
pre_tokenizer_state = json.loads(tokenizer_r.backend_tokenizer.pre_tokenizer.__getstate__())
post_processor_state = json.loads(tokenizer_r.backend_tokenizer.post_processor.__getstate__())
self.assertEqual(pre_tokenizer_state["add_prefix_space"], add_prefix_space)
self.assertEqual(post_processor_state["add_prefix_space"], add_prefix_space)
self.assertEqual(post_processor_state["trim_offsets"], trim_offsets)
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_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=True
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
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)),
)
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=True
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
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)),
)
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=False
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(len(text_of_1_token), len(text_of_1_token) + 1 + len(text_of_1_token)),
)
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=False
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (0, len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(len(text_of_1_token), len(text_of_1_token) + 1 + len(text_of_1_token)),
)
text = f" {text}"
# tokenizer_r = self.rust_tokenizer_class.from_pretrained(
# pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=True
# )
# encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
# self.assertEqual(encoding.offset_mapping[0], (1, 1 + len(text_of_1_token)))
# self.assertEqual(
# encoding.offset_mapping[1],
# (1 + len(text_of_1_token) + 1, 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
# )
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=True
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (1, 1 + len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(1 + len(text_of_1_token) + 1, 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
)
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=True, trim_offsets=False
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (0, 1 + len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(1 + len(text_of_1_token), 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
)
tokenizer_r = self.rust_tokenizer_class.from_pretrained(
pretrained_name, use_fast=True, add_prefix_space=False, trim_offsets=False
)
encoding = tokenizer_r(text, return_offsets_mapping=True, add_special_tokens=False)
self.assertEqual(encoding.offset_mapping[0], (0, 1 + len(text_of_1_token)))
self.assertEqual(
encoding.offset_mapping[1],
(1 + len(text_of_1_token), 1 + len(text_of_1_token) + 1 + len(text_of_1_token)),
)