-
Notifications
You must be signed in to change notification settings - Fork 13
/
module.py
167 lines (149 loc) · 8.41 KB
/
module.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
import torch
import alignment_C as Aln_C
from torch import nn
import torch.nn.functional as F
import torch.multiprocessing as mp
from dataload import num_to_base
import time
import numpy as np
class Train_Module:
def __init__(self, config):
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.config = config
def train_MLM(self, low_seq_0, masked_seq_0, prediction_scores):
criterion = nn.CrossEntropyLoss()
mask = masked_seq_0 - low_seq_0 != 0
length_mask = masked_seq_0 != 0
same_base_mask = torch.bernoulli(torch.ones(mask.shape)*0.05).byte().to(self.device)
mask = mask + same_base_mask
index = torch.nonzero(mask).split(1, dim=1)
prediction_scores = torch.squeeze(prediction_scores[index])
new_low_seq_0 = torch.squeeze(low_seq_0[index])
loss = criterion(prediction_scores, new_low_seq_0)
_, preds = torch.max(prediction_scores, 1)
correct = torch.sum(preds == new_low_seq_0.data).double()/len(new_low_seq_0)
return loss, correct
def train_SSL(self, SS, prediction_scores):
criterion = nn.CrossEntropyLoss()
index = torch.nonzero(SS).split(1, dim=1)
prediction_scores = torch.squeeze(prediction_scores[index])
SS_answer = torch.squeeze(SS[index])
loss = criterion(prediction_scores, SS_answer)
_, preds = torch.max(prediction_scores, 1)
correct = torch.sum(preds == SS_answer.data).double()/len(SS_answer)
return loss, correct
def train_SFP(self, low_seq_0, seq_len_0, low_seq_1, seq_len_1, family_0, family_1, z0_list, z1_list):
distance_label = torch.where((family_0 - family_1) == 0, torch.ones(family_0.size()[0]), torch.full((family_0.size()[0],), -1) ).to("cuda")
distance_label_same = torch.where((family_0 - family_1) == 0, torch.ones(family_0.size()[0]), torch.zeros(family_0.size()[0]) ).to("cuda")
distance_label_diff = torch.where((family_0 - family_1) != 0, torch.ones(family_0.size()[0]), torch.zeros(family_0.size()[0]) ).to("cuda")
_, logits = self.match(z0_list, z1_list)
distance = -logits
distance_same = torch.dot(distance_label_same , distance.view(-1)) / torch.sum(distance_label_same)
distance_diff = torch.dot(distance_label_diff , distance.view(-1)) / torch.sum(distance_label_diff)
margin = 600
loss = F.relu( distance_same - distance_diff + margin )
# accuracy
predicted_label = torch.where(distance.view(-1) < 660, torch.ones(family_0.size()[0]).to("cuda"), torch.full((family_0.size()[0],), -1).to("cuda"))
match = torch.where((distance_label * predicted_label) == 1, torch.ones(family_0.size()[0]).to("cuda"), torch.zeros(family_0.size()[0]).to("cuda"))
correct = torch.sum(match)
return loss, correct
def train_MUL(self, z0_list, z1_list, common_index_0, common_index_1, seq_len_0, seq_len_1):
bert_scores, _ = self.match(z0_list, z1_list)
loss = 0.0
for i, bert_score in enumerate(bert_scores):
loss += self.structural_learning(bert_score, common_index_0[i], seq_len_0[i], common_index_1[i], seq_len_1[i])
return loss
def test_align(self, low_seq_0, low_seq_1, z0_list, z1_list, common_index_0, common_index_1, seq_len_0, seq_len_1, show_aln):
bert_scores, _ = self.match(z0_list, z1_list)
sequence_a = num_to_base(low_seq_0)
sequence_b = num_to_base(low_seq_1)
len_pred_match = 0
len_ref_match = 0
len_TP = 0
for i, bert_score in enumerate(bert_scores):
bert_score = torch.flatten(bert_score.T).tolist()
x = 1 if show_aln == True else 0
common_index_A_B = Aln_C.global_aln(bert_score, [0] * len(bert_score), [0] * len(bert_score), sequence_a[i], sequence_b[i], seq_len_0[i], seq_len_1[i], self.config.gap_opening, self.config.gap_extension, x, 0)
common_index_A_B = torch.tensor(common_index_A_B).to(self.device).view(2, -1)
len_pred_match += int(torch.sum(common_index_A_B[0]))
len_ref_match += int(torch.sum(common_index_0[i]))
a = torch.flatten( (common_index_A_B[0] == 1).nonzero()*10000 + (common_index_A_B[1] == 1).nonzero()).tolist()
b = torch.flatten( (common_index_0[i] == 1).nonzero()*10000 + (common_index_1[i] == 1).nonzero()).tolist()
len_TP += len(set(a) & set(b))
return len_TP, len_pred_match, len_ref_match
def structural_learning(self, bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1):
reference_alignment_score = self.calc_reference_alignment_score(bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1)
prediction_alignment_score, _ = self.calc_prediction_alignment_score(bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1)
return prediction_alignment_score - reference_alignment_score
def match_bert_score(self, bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1):
index_0 = (common_index_0 == 1).nonzero()
index_1 = (common_index_1 == 1).nonzero()
index = torch.cat([index_0, index_1], axis=1).T
index = tuple(index)
omega = bert_score[index]
return omega
def margin_matrix(self, bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1):
index_0 = (common_index_0 == 1).nonzero()
index_1 = (common_index_1 == 1).nonzero()
index = torch.cat([index_0, index_1], axis=1).T
index = tuple(index)
margin_mat_FP = torch.ones(bert_score.shape).to(self.device) * self.config.margin_FP
margin_mat_FP[index] = 0.0
margin_mat_FN = torch.zeros(bert_score.shape).to(self.device)
margin_mat_FN[index] = self.config.margin_FN
return margin_mat_FP, margin_mat_FN
def margin_score(self, common_index_A_B, common_index_0, common_index_1):
len_pred_match = int(torch.sum(common_index_A_B[0]))
len_ref_match = int(torch.sum(common_index_0))
a = torch.flatten( (common_index_A_B[0] == 1).nonzero()*10000 + (common_index_A_B[1] == 1).nonzero()).tolist()
b = torch.flatten( (common_index_0 == 1).nonzero()*10000 + (common_index_1 == 1).nonzero()).tolist()
len_TP = len(set(a) & set(b))
len_FP = len_pred_match - len_TP
len_FN = len_ref_match - len_TP
return len_FP * self.config.margin_FP + len_FN * self.config.margin_FN
def calc_reference_alignment_score(self, bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1):
reference_alignment_score = 0
reference_alignment_score += self.gapscore(common_index_0, seq_len_0)
reference_alignment_score += self.gapscore(common_index_1, seq_len_1)
omega = self.match_bert_score(bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1)
reference_alignment_score += torch.sum(omega)
return reference_alignment_score
def calc_prediction_alignment_score(self, bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1):
sequence_a = "N"*seq_len_0
sequence_b = "N"*seq_len_1
margin_mat_FP, margin_mat_FN = self.margin_matrix(bert_score, common_index_0, seq_len_0, common_index_1, seq_len_1)
common_index_A_B = Aln_C.global_aln(torch.flatten(bert_score.T).tolist(), torch.flatten(margin_mat_FP.T).tolist(), torch.flatten(margin_mat_FN.T).tolist(), sequence_a, sequence_b, seq_len_0, seq_len_1, self.config.gap_opening, self.config.gap_extension, 0, 0)
common_index_A_B = torch.tensor(common_index_A_B).to(self.device).view(2, -1)
prediction_alignment_score = self.calc_reference_alignment_score(bert_score, common_index_A_B[0], seq_len_0, common_index_A_B[1], seq_len_1) + self.margin_score(common_index_A_B, common_index_0, common_index_1)
return prediction_alignment_score, common_index_A_B
def gapscore(self, index, seq_len):
seq_len = int(seq_len)
index = index[:seq_len].to('cpu').detach().numpy().copy()
all_zeros = seq_len - np.count_nonzero(index)
extend_zeros = seq_len - np.count_nonzero(np.insert(index[:-1], 0, 1) + index)
open_zeros = all_zeros - extend_zeros
return (-1*self.config.gap_opening + -1*self.config.gap_extension ) * open_zeros + -1*self.config.gap_extension * extend_zeros
def match(self,z0_list, z1_list):
match_scores = []
logits = []
for z0, z1 in zip(z0_list, z1_list):
# cos similarity
match_score = nn.CosineSimilarity(dim=2, eps=1e-6)(z0.unsqueeze(1).repeat(1, z1.shape[0],1) , z1.unsqueeze(0).repeat(z0.shape[0], 1,1))
s = 1.3 * match_score
# L1 distance
# s = -torch.sum(torch.abs(z0.unsqueeze(1)-z1), -1)
# match_score = torch.exp(s)
# soft align
a, b = F.softmax(s, 1), F.softmax(s, 0)
c = a + b - a*b
c = torch.sum(c*s)/torch.sum(c)
match_scores.append(match_score)
logits.append(c.view(-1))
logits = torch.stack(logits, 0)
return match_scores, logits
def em(self, h, lengths):
# get representations with different lengths from the collated single matrix
e = [None] * len(lengths)
for i in range(len(lengths)):
e[i] = h[i, :lengths[i]]
return e