-
Notifications
You must be signed in to change notification settings - Fork 300
/
seq_aligner.py
196 lines (168 loc) · 6.41 KB
/
seq_aligner.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
# Copyright 2022 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
import numpy as np
class ScoreParams:
def __init__(self, gap, match, mismatch):
self.gap = gap
self.match = match
self.mismatch = mismatch
def mis_match_char(self, x, y):
if x != y:
return self.mismatch
else:
return self.match
def get_matrix(size_x, size_y, gap):
matrix = []
for i in range(len(size_x) + 1):
sub_matrix = []
for j in range(len(size_y) + 1):
sub_matrix.append(0)
matrix.append(sub_matrix)
for j in range(1, len(size_y) + 1):
matrix[0][j] = j*gap
for i in range(1, len(size_x) + 1):
matrix[i][0] = i*gap
return matrix
def get_matrix(size_x, size_y, gap):
matrix = np.zeros((size_x + 1, size_y + 1), dtype=np.int32)
matrix[0, 1:] = (np.arange(size_y) + 1) * gap
matrix[1:, 0] = (np.arange(size_x) + 1) * gap
return matrix
def get_traceback_matrix(size_x, size_y):
matrix = np.zeros((size_x + 1, size_y +1), dtype=np.int32)
matrix[0, 1:] = 1
matrix[1:, 0] = 2
matrix[0, 0] = 4
return matrix
def global_align(x, y, score):
matrix = get_matrix(len(x), len(y), score.gap)
trace_back = get_traceback_matrix(len(x), len(y))
for i in range(1, len(x) + 1):
for j in range(1, len(y) + 1):
left = matrix[i, j - 1] + score.gap
up = matrix[i - 1, j] + score.gap
diag = matrix[i - 1, j - 1] + score.mis_match_char(x[i - 1], y[j - 1])
matrix[i, j] = max(left, up, diag)
if matrix[i, j] == left:
trace_back[i, j] = 1
elif matrix[i, j] == up:
trace_back[i, j] = 2
else:
trace_back[i, j] = 3
return matrix, trace_back
def get_aligned_sequences(x, y, trace_back):
x_seq = []
y_seq = []
i = len(x)
j = len(y)
mapper_y_to_x = []
while i > 0 or j > 0:
if trace_back[i, j] == 3:
x_seq.append(x[i-1])
y_seq.append(y[j-1])
i = i-1
j = j-1
mapper_y_to_x.append((j, i))
elif trace_back[i][j] == 1:
x_seq.append('-')
y_seq.append(y[j-1])
j = j-1
mapper_y_to_x.append((j, -1))
elif trace_back[i][j] == 2:
x_seq.append(x[i-1])
y_seq.append('-')
i = i-1
elif trace_back[i][j] == 4:
break
mapper_y_to_x.reverse()
return x_seq, y_seq, torch.tensor(mapper_y_to_x, dtype=torch.int64)
def get_mapper(x: str, y: str, tokenizer, max_len=77):
x_seq = tokenizer.encode(x)
y_seq = tokenizer.encode(y)
score = ScoreParams(0, 1, -1)
matrix, trace_back = global_align(x_seq, y_seq, score)
mapper_base = get_aligned_sequences(x_seq, y_seq, trace_back)[-1]
alphas = torch.ones(max_len)
alphas[: mapper_base.shape[0]] = mapper_base[:, 1].ne(-1).float()
mapper = torch.zeros(max_len, dtype=torch.int64)
mapper[:mapper_base.shape[0]] = mapper_base[:, 1]
mapper[mapper_base.shape[0]:] = len(y_seq) + torch.arange(max_len - len(y_seq))
return mapper, alphas
def get_refinement_mapper(prompts, tokenizer, max_len=77):
x_seq = prompts[0]
mappers, alphas = [], []
for i in range(1, len(prompts)):
mapper, alpha = get_mapper(x_seq, prompts[i], tokenizer, max_len)
mappers.append(mapper)
alphas.append(alpha)
return torch.stack(mappers), torch.stack(alphas)
def get_word_inds(text: str, word_place: int, tokenizer):
split_text = text.split(" ")
if type(word_place) is str:
word_place = [i for i, word in enumerate(split_text) if word_place == word]
elif type(word_place) is int:
word_place = [word_place]
out = []
if len(word_place) > 0:
words_encode = [tokenizer.decode([item]).strip("#") for item in tokenizer.encode(text)][1:-1]
cur_len, ptr = 0, 0
for i in range(len(words_encode)):
cur_len += len(words_encode[i])
if ptr in word_place:
out.append(i + 1)
if cur_len >= len(split_text[ptr]):
ptr += 1
cur_len = 0
return np.array(out)
def get_replacement_mapper_(x: str, y: str, tokenizer, max_len=77):
words_x = x.split(' ')
words_y = y.split(' ')
if len(words_x) != len(words_y):
raise ValueError(f"attention replacement edit can only be applied on prompts with the same length"
f" but prompt A has {len(words_x)} words and prompt B has {len(words_y)} words.")
inds_replace = [i for i in range(len(words_y)) if words_y[i] != words_x[i]]
inds_source = [get_word_inds(x, i, tokenizer) for i in inds_replace]
inds_target = [get_word_inds(y, i, tokenizer) for i in inds_replace]
mapper = np.zeros((max_len, max_len))
i = j = 0
cur_inds = 0
while i < max_len and j < max_len:
if cur_inds < len(inds_source) and inds_source[cur_inds][0] == i:
inds_source_, inds_target_ = inds_source[cur_inds], inds_target[cur_inds]
if len(inds_source_) == len(inds_target_):
mapper[inds_source_, inds_target_] = 1
else:
ratio = 1 / len(inds_target_)
for i_t in inds_target_:
mapper[inds_source_, i_t] = ratio
cur_inds += 1
i += len(inds_source_)
j += len(inds_target_)
elif cur_inds < len(inds_source):
mapper[i, j] = 1
i += 1
j += 1
else:
mapper[j, j] = 1
i += 1
j += 1
return torch.from_numpy(mapper).float()
def get_replacement_mapper(prompts, tokenizer, max_len=77):
x_seq = prompts[0]
mappers = []
for i in range(1, len(prompts)):
mapper = get_replacement_mapper_(x_seq, prompts[i], tokenizer, max_len)
mappers.append(mapper)
return torch.stack(mappers)