Skip to content
Navigation Menu
{{ message }}
-
Notifications
You must be signed in to change notification settings - Fork 377
Expand file tree
/
Copy pathsplit_merge_tokenizer.py
More file actions
273 lines (233 loc) · 10.4 KB
/
Copy pathsplit_merge_tokenizer.py
File metadata and controls
273 lines (233 loc) · 10.4 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
# coding=utf-8
# Copyright 2026 TF.Text Authors.
#
# 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.
"""Ops to tokenize words into subwords."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from tensorflow.python.eager import monitoring
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import ops
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import array_ops_stack
from tensorflow.python.ops import math_ops
from tensorflow.python.ops.ragged import ragged_tensor
from tensorflow.python.ops.ragged.ragged_tensor import RaggedTensor
from tensorflow_text.python.ops.tokenization import TokenizerWithOffsets
# pylint: disable=g-bad-import-order
from tensorflow.python.framework import load_library
from tensorflow.python.platform import resource_loader
gen_split_merge_tokenizer = load_library.load_op_library(resource_loader.get_path_to_datafile('_split_merge_tokenizer.so'))
_tf_text_split_merge_tokenizer_op_create_counter = monitoring.Counter(
'/nlx/api/python/split_merge_tokenizer_create_counter',
'Counter for number of SplitMergeTokenizers created in Python.')
class SplitMergeTokenizer(TokenizerWithOffsets):
"""Tokenizes a tensor of UTF-8 string into words according to labels."""
def __init__(self):
"""Initializes a new instance.
"""
super(SplitMergeTokenizer, self).__init__()
_tf_text_split_merge_tokenizer_op_create_counter.get_cell().increase_by(1)
def tokenize(self,
input, # pylint: disable=redefined-builtin
labels,
force_split_at_break_character=True):
"""Tokenizes a tensor of UTF-8 strings according to labels.
### Example:
>>> strings = ["HelloMonday", "DearFriday"]
>>> labels = [[0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1],
... [0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0]]
>>> tokenizer = SplitMergeTokenizer()
>>> tokenizer.tokenize(strings, labels)
<tf.RaggedTensor [[b'Hello', b'Monday'], [b'Dear', b'Friday']]>
Args:
input: An N-dimensional `Tensor` or `RaggedTensor` of UTF-8 strings.
labels: An (N+1)-dimensional `Tensor` or `RaggedTensor` of `int32`, with
`labels[i1...iN, j]` being the split(0)/merge(1) label of the j-th
character for `input[i1...iN]`. Here split means create a new word with
this character and merge means adding this character to the previous
word.
force_split_at_break_character: bool indicates whether to force start a
new word after seeing a ICU defined whitespace character. When seeing
one or more ICU defined whitespace character:
* if `force_split_at_break_character` is set true, then create a new
word at the first non-space character, regardless of the label of that
character, for instance:
```python
input="New York"
labels=[0, 1, 1, 0, 1, 1, 1, 1]
output tokens=["New", "York"]
```
```python
input="New York"
labels=[0, 1, 1, 1, 1, 1, 1, 1]
output tokens=["New", "York"]
```
```python
input="New York",
labels=[0, 1, 1, 1, 0, 1, 1, 1]
output tokens=["New", "York"]
```
* otherwise, whether to create a new word or not for the first non-space
character depends on the label of that character, for instance:
```python
input="New York",
labels=[0, 1, 1, 0, 1, 1, 1, 1]
output tokens=["NewYork"]
```
```python
input="New York",
labels=[0, 1, 1, 1, 1, 1, 1, 1]
output tokens=["NewYork"]
```
```python
input="New York",
labels=[0, 1, 1, 1, 0, 1, 1, 1]
output tokens=["New", "York"]
```
Returns:
A `RaggedTensor` of strings where `tokens[i1...iN, j]` is the string
content of the `j-th` token in `input[i1...iN]`
"""
subword, _, _ = self.tokenize_with_offsets(input, labels,
force_split_at_break_character)
return subword
def tokenize_with_offsets(self,
input, # pylint: disable=redefined-builtin
labels,
force_split_at_break_character=True):
"""Tokenizes a tensor of UTF-8 strings into tokens with [start,end) offsets.
### Example:
>>> strings = ["HelloMonday", "DearFriday"]
>>> labels = [[0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1],
... [0, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0]]
>>> tokenizer = SplitMergeTokenizer()
>>> tokens, starts, ends = tokenizer.tokenize_with_offsets(strings, labels)
>>> tokens
<tf.RaggedTensor [[b'Hello', b'Monday'], [b'Dear', b'Friday']]>
>>> starts
<tf.RaggedTensor [[0, 5], [0, 4]]>
>>> ends
<tf.RaggedTensor [[5, 11], [4, 10]]>
Args:
input: An N-dimensional `Tensor` or `RaggedTensor` of UTF-8 strings.
labels: An (N+1)-dimensional `Tensor` or `RaggedTensor` of int32, with
labels[i1...iN, j] being the split(0)/merge(1) label of the j-th
character for input[i1...iN]. Here split means create a new word with
this character and merge means adding this character to the previous
word.
force_split_at_break_character: bool indicates whether to force start a
new word after seeing a ICU defined whitespace character. When seeing
one or more ICU defined whitespace character:
* if `force_split_at_break_character` is set true, then create a new
word at the first non-space character, regardless of the label of
that character, for instance:
```python
input="New York"
labels=[0, 1, 1, 0, 1, 1, 1, 1]
output tokens=["New", "York"]
```
```python
input="New York"
labels=[0, 1, 1, 1, 1, 1, 1, 1]
output tokens=["New", "York"]
```
```python
input="New York",
labels=[0, 1, 1, 1, 0, 1, 1, 1]
output tokens=["New", "York"]
```
* otherwise, whether to create a new word or not for the first non-space
character depends on the label of that character, for instance:
```python
input="New York",
labels=[0, 1, 1, 0, 1, 1, 1, 1]
output tokens=["NewYork"]
```
```python
input="New York",
labels=[0, 1, 1, 1, 1, 1, 1, 1]
output tokens=["NewYork"]
```
```python
input="New York",
labels=[0, 1, 1, 1, 0, 1, 1, 1]
output tokens=["New", "York"]
```
Returns:
A tuple `(tokens, start_offsets, end_offsets)` where:
tokens: is a `RaggedTensor` of strings where `tokens[i1...iN, j]` is
the string content of the `j-th` token in `input[i1...iN]`
start_offsets: is a `RaggedTensor` of int64s where
`start_offsets[i1...iN, j]` is the byte offset for the start of the
`j-th` token in `input[i1...iN]`.
end_offsets: is a `RaggedTensor` of int64s where
`end_offsets[i1...iN, j]` is the byte offset immediately after the
end of the `j-th` token in `input[i...iN]`.
"""
name = None
with ops.name_scope(
name, 'SplitMergeTokenizeWithOffsets',
[input, labels, force_split_at_break_character]):
# Check that the types are expected and the ragged rank is appropriate.
tokens = ragged_tensor.convert_to_tensor_or_ragged_tensor(input)
labels = ragged_tensor.convert_to_tensor_or_ragged_tensor(labels)
rank = tokens.shape.ndims
if rank is None:
raise ValueError('input must have a known rank.')
if rank == 0:
words, starts, ends = self.tokenize_with_offsets(
array_ops_stack.stack([tokens]),
array_ops_stack.stack([labels]),
force_split_at_break_character)
return words.values, starts.values, ends.values
elif rank > 1:
if not ragged_tensor.is_ragged(tokens):
tokens = ragged_tensor.RaggedTensor.from_tensor(
tokens, ragged_rank=rank - 1)
# Convert to a 2D ragged tensor from labels of shape
# [#input_string, (labels per string)]
if not ragged_tensor.is_ragged(labels):
labels2d = array_ops.reshape(labels, [-1, labels.shape[-1]])
labels_unpack = ragged_tensor.RaggedTensor.from_tensor(labels2d)
else:
labels_unpack = ragged_tensor.RaggedTensor.from_row_splits(
values=labels.flat_values,
row_splits=labels.nested_row_splits[-1])
words, starts, ends = self.tokenize_with_offsets(
tokens.flat_values,
labels_unpack,
force_split_at_break_character)
words = words.with_row_splits_dtype(tokens.row_splits.dtype)
starts = starts.with_row_splits_dtype(tokens.row_splits.dtype)
ends = ends.with_row_splits_dtype(tokens.row_splits.dtype)
return (tokens.with_flat_values(words),
tokens.with_flat_values(starts),
tokens.with_flat_values(ends))
if not ragged_tensor.is_ragged(labels):
ragged_labels = ragged_tensor.RaggedTensor.from_tensor(labels)
else:
ragged_labels = labels
row_splits = math_ops.cast(ragged_labels.row_splits, dtypes.int32)
# Tokenize the strings into tokens.
values, row_splits, starts, ends = (
gen_split_merge_tokenizer.split_merge_tokenize_with_offsets(
input_values=tokens,
labels=ragged_labels.flat_values,
row_splits=row_splits,
force_split_at_break_character=force_split_at_break_character))
words = RaggedTensor.from_row_splits(values, row_splits, validate=False)
starts = RaggedTensor.from_row_splits(starts, row_splits, validate=False)
ends = RaggedTensor.from_row_splits(ends, row_splits, validate=False)
return words, starts, ends
You can’t perform that action at this time.