This repository has been archived by the owner on Dec 16, 2022. It is now read-only.
/
bias_direction_wrappers.py
269 lines (225 loc) · 9.54 KB
/
bias_direction_wrappers.py
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
import torch
from typing import Union, Optional
from os import PathLike
from allennlp.fairness.bias_direction import (
BiasDirection,
PCABiasDirection,
PairedPCABiasDirection,
TwoMeansBiasDirection,
ClassificationNormalBiasDirection,
)
from allennlp.fairness.bias_utils import load_word_pairs, load_words
from allennlp.common import Registrable
from allennlp.data.tokenizers.tokenizer import Tokenizer
from allennlp.data import Vocabulary
class BiasDirectionWrapper(Registrable):
"""
Parent class for bias direction wrappers.
"""
def __init__(self):
self.direction: BiasDirection = None
self.noise: float = None
def __call__(self, module):
raise NotImplementedError
def train(self, mode: bool = True):
"""
# Parameters
mode : `bool`, optional (default=`True`)
Sets `requires_grad` to value of `mode` for bias direction.
"""
self.direction.requires_grad = mode
def add_noise(self, t: torch.Tensor):
"""
# Parameters
t : `torch.Tensor`
Tensor to which to add small amount of Gaussian noise.
"""
return t + self.noise * torch.randn(t.size(), device=t.device)
@BiasDirectionWrapper.register("pca")
class PCABiasDirectionWrapper(BiasDirectionWrapper):
"""
# Parameters
seed_words_file : `Union[PathLike, str]`
Path of file containing seed words.
tokenizer : `Tokenizer`
Tokenizer used to tokenize seed words.
direction_vocab : `Vocabulary`, optional (default=`None`)
Vocabulary of tokenizer. If `None`, assumes tokenizer is of
type `PreTrainedTokenizer` and uses tokenizer's `vocab` attribute.
namespace : `str`, optional (default=`"tokens"`)
Namespace of direction_vocab to use when tokenizing.
Disregarded when direction_vocab is `None`.
requires_grad : `bool`, optional (default=`False`)
Option to enable gradient calculation for bias direction.
noise : `float`, optional (default=`1e-10`)
To avoid numerical instability if embeddings are initialized uniformly.
"""
def __init__(
self,
seed_words_file: Union[PathLike, str],
tokenizer: Tokenizer,
direction_vocab: Optional[Vocabulary] = None,
namespace: str = "tokens",
requires_grad: bool = False,
noise: float = 1e-10,
):
self.ids = load_words(seed_words_file, tokenizer, direction_vocab, namespace)
self.direction = PCABiasDirection(requires_grad=requires_grad)
self.noise = noise
def __call__(self, module):
# embed subword token IDs and mean pool to get
# embedding of original word
ids_embeddings = []
for i in self.ids:
i = i.to(module.weight.device)
ids_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids_embeddings = torch.cat(ids_embeddings)
# adding trivial amount of noise
# to eliminate linear dependence amongst all embeddings
# when training first starts
ids_embeddings = self.add_noise(ids_embeddings)
return self.direction(ids_embeddings)
@BiasDirectionWrapper.register("paired_pca")
class PairedPCABiasDirectionWrapper(BiasDirectionWrapper):
"""
# Parameters
seed_word_pairs_file : `Union[PathLike, str]`
Path of file containing seed word pairs.
tokenizer : `Tokenizer`
Tokenizer used to tokenize seed words.
direction_vocab : `Vocabulary`, optional (default=`None`)
Vocabulary of tokenizer. If `None`, assumes tokenizer is of
type `PreTrainedTokenizer` and uses tokenizer's `vocab` attribute.
namespace : `str`, optional (default=`"tokens"`)
Namespace of direction_vocab to use when tokenizing.
Disregarded when direction_vocab is `None`.
requires_grad : `bool`, optional (default=`False`)
Option to enable gradient calculation for bias direction.
noise : `float`, optional (default=`1e-10`)
To avoid numerical instability if embeddings are initialized uniformly.
"""
def __init__(
self,
seed_word_pairs_file: Union[PathLike, str],
tokenizer: Tokenizer,
direction_vocab: Optional[Vocabulary] = None,
namespace: str = "tokens",
requires_grad: bool = False,
noise: float = 1e-10,
):
self.ids1, self.ids2 = load_word_pairs(
seed_word_pairs_file, tokenizer, direction_vocab, namespace
)
self.direction = PairedPCABiasDirection(requires_grad=requires_grad)
self.noise = noise
def __call__(self, module):
# embed subword token IDs and mean pool to get
# embedding of original word
ids1_embeddings = []
for i in self.ids1:
i = i.to(module.weight.device)
ids1_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids2_embeddings = []
for i in self.ids2:
i = i.to(module.weight.device)
ids2_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids1_embeddings = torch.cat(ids1_embeddings)
ids2_embeddings = torch.cat(ids2_embeddings)
ids1_embeddings = self.add_noise(ids1_embeddings)
ids2_embeddings = self.add_noise(ids2_embeddings)
return self.direction(ids1_embeddings, ids2_embeddings)
@BiasDirectionWrapper.register("two_means")
class TwoMeansBiasDirectionWrapper(BiasDirectionWrapper):
"""
# Parameters
seed_word_pairs_file : `Union[PathLike, str]`
Path of file containing seed word pairs.
tokenizer : `Tokenizer`
Tokenizer used to tokenize seed words.
direction_vocab : `Vocabulary`, optional (default=`None`)
Vocabulary of tokenizer. If `None`, assumes tokenizer is of
type `PreTrainedTokenizer` and uses tokenizer's `vocab` attribute.
namespace : `str`, optional (default=`"tokens"`)
Namespace of direction_vocab to use when tokenizing.
Disregarded when direction_vocab is `None`.
requires_grad : `bool`, optional (default=`False`)
Option to enable gradient calculation for bias direction.
noise : `float`, optional (default=`1e-10`)
To avoid numerical instability if embeddings are initialized uniformly.
"""
def __init__(
self,
seed_word_pairs_file: Union[PathLike, str],
tokenizer: Tokenizer,
direction_vocab: Optional[Vocabulary] = None,
namespace: str = "tokens",
requires_grad: bool = False,
noise: float = 1e-10,
):
self.ids1, self.ids2 = load_word_pairs(
seed_word_pairs_file, tokenizer, direction_vocab, namespace
)
self.direction = TwoMeansBiasDirection(requires_grad=requires_grad)
self.noise = noise
def __call__(self, module):
# embed subword token IDs and mean pool to get
# embedding of original word
ids1_embeddings = []
for i in self.ids1:
i = i.to(module.weight.device)
ids1_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids2_embeddings = []
for i in self.ids2:
i = i.to(module.weight.device)
ids2_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids1_embeddings = torch.cat(ids1_embeddings)
ids2_embeddings = torch.cat(ids2_embeddings)
ids1_embeddings = self.add_noise(ids1_embeddings)
ids2_embeddings = self.add_noise(ids2_embeddings)
return self.direction(ids1_embeddings, ids2_embeddings)
@BiasDirectionWrapper.register("classification_normal")
class ClassificationNormalBiasDirectionWrapper(BiasDirectionWrapper):
"""
# Parameters
seed_word_pairs_file : `Union[PathLike, str]`
Path of file containing seed word pairs.
tokenizer : `Tokenizer`
Tokenizer used to tokenize seed words.
direction_vocab : `Vocabulary`, optional (default=`None`)
Vocabulary of tokenizer. If `None`, assumes tokenizer is of
type `PreTrainedTokenizer` and uses tokenizer's `vocab` attribute.
namespace : `str`, optional (default=`"tokens"`)
Namespace of direction_vocab to use when tokenizing.
Disregarded when direction_vocab is `None`.
noise : `float`, optional (default=`1e-10`)
To avoid numerical instability if embeddings are initialized uniformly.
"""
def __init__(
self,
seed_word_pairs_file: Union[PathLike, str],
tokenizer: Tokenizer,
direction_vocab: Optional[Vocabulary] = None,
namespace: str = "tokens",
noise: float = 1e-10,
):
self.ids1, self.ids2 = load_word_pairs(
seed_word_pairs_file, tokenizer, direction_vocab, namespace
)
self.direction = ClassificationNormalBiasDirection()
self.noise = noise
def __call__(self, module):
# embed subword token IDs and mean pool to get
# embedding of original word
ids1_embeddings = []
for i in self.ids1:
i = i.to(module.weight.device)
ids1_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids2_embeddings = []
for i in self.ids2:
i = i.to(module.weight.device)
ids2_embeddings.append(torch.mean(module.forward(i), dim=0, keepdim=True))
ids1_embeddings = torch.cat(ids1_embeddings)
ids2_embeddings = torch.cat(ids2_embeddings)
ids1_embeddings = self.add_noise(ids1_embeddings)
ids2_embeddings = self.add_noise(ids2_embeddings)
return self.direction(ids1_embeddings, ids2_embeddings)