-
Notifications
You must be signed in to change notification settings - Fork 0
/
quick_sens_v2.py
152 lines (132 loc) · 6.02 KB
/
quick_sens_v2.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
# mark
from transformers import RobertaTokenizer, RobertaForMaskedLM
from transformers import BertTokenizer, BertForMaskedLM
from transformers import AutoTokenizer, AutoModelForMaskedLM
from transformers import AdamW
import torch
from torch.nn import functional as F
import numpy as np
from numpy import linalg as LA
import pandas as pd
import math
import statistics
from collections import defaultdict
import csv
import random
import os
from multiclassUpdate import multiclass_update
random.seed(10)
def equal(l1,l2):
assert len(l1) == len(l2)
for x,y in zip(l1,l2):
if x!=y: return False
return True
def get_lognorm_score(model, tokenizer, sentence, trait):
# print(trait)
# print(sentence)
trait_ids = tokenizer.encode(trait, return_tensors='pt').squeeze().tolist()
# print(trait_ids)
trait_ids = trait_ids[1:len(trait_ids)-1]
# print(trait_ids)
scores = []
input_ids = tokenizer.encode(sentence, return_tensors='pt')
masked_position = (input_ids.squeeze() == tokenizer.mask_token_id).nonzero()
masked_poss = [mask.item() for mask in masked_position]
# print(masked_poss)
for i in range(len(trait_ids)):
# print(input_ids)
# print(masked_pos)
# print(i)
first_masked_pos = masked_poss[i]
target_id = trait_ids[i]
# print("Masked position, target id: ", first_masked_pos, target_id)
output = model(input_ids,output_hidden_states=True)
logits = output.logits.squeeze()[first_masked_pos].cpu().detach().numpy() # y
lhs = output.hidden_states[-1].squeeze()[first_masked_pos]
# print('X: ', lhs.shape)
lhs = model.cls.predictions.transform(lhs).cpu().detach().numpy() # x
# print('X: ', lhs.shape)
# print('Y: ', logits.shape, logits)
weights = model.cls.predictions.decoder.weight.squeeze().cpu().detach().numpy()
bias = model.cls.predictions.decoder.bias.reshape(-1,1).cpu().detach().numpy()
# print('A: ',weights.shape)
# print('Bias: ',bias.shape, bias)
lhs = np.append(lhs, 1)
# print('Final X: ', lhs.shape)
W = np.append(weights, bias, axis=1)
# print('Final A: ', W.shape)
newW = multiclass_update(W, lhs, target_id)
score = np.log(1-LA.norm(W-newW))
scores = np.append(scores, score)
input_ids[0][first_masked_pos] = target_id
# print('#',scores)
fscore = np.sum(scores)
# print(fscore)
return fscore
def wordasso(model, tokenizer, groups, traits, prior_group, tmplt="The <group> person is <mask>.",traits_for_prior=None, tplt_for_prior=None):
scores = {'traits': traits}
for group in groups:
group_scores = []
if traits_for_prior == None:
traits_for_prior = [None]*len(traits)
for trait, trait_fp in zip(traits, traits_for_prior):
if trait_fp == None and tplt_for_prior == None:
trait_ids = tokenizer.encode(trait, return_tensors='pt').squeeze()
trait_len = list(trait_ids.shape)[0]-2
# print(trait,trait_ids, trait_len)
input_txt = tmplt.replace(' <mask>', (' '+tokenizer.mask_token)*trait_len).replace('<group>',prior_group)
# print(input_txt)
prior = get_lognorm_score(model, tokenizer, input_txt, trait)
elif trait_fp != None:
trait_ids = tokenizer.encode(trait_fp, return_tensors='pt').squeeze()
trait_len = list(trait_ids.shape)[0]-2
# print(trait,trait_ids, trait_len)
input_txt = tmplt.replace(' <mask>', (' '+tokenizer.mask_token)*trait_len).replace('<group>',prior_group)
# print(input_txt)
prior = get_lognorm_score(model, tokenizer, input_txt, trait_fp)
trait_ids = tokenizer.encode(trait, return_tensors='pt').squeeze()
trait_len = list(trait_ids.shape)[0]-2
# print(trait,trait_ids, trait_len)
elif tplt_for_prior != None:
trait_ids = tokenizer.encode(trait, return_tensors='pt').squeeze()
trait_len = list(trait_ids.shape)[0]-2
# print(trait,trait_ids, trait_len)
input_txt = tplt_for_prior.replace(' <mask>', (' '+tokenizer.mask_token)*trait_len).replace('<group>',prior_group)
# print(input_txt)
prior = get_lognorm_score(model, tokenizer, input_txt, trait)
input_txt = tmplt.replace(' <mask>', (' '+tokenizer.mask_token)*trait_len).replace('<group>',group)
# print(input_txt)
target = get_lognorm_score(model, tokenizer, input_txt, trait)
lps_score = target-prior
# p_scores.append(prior)#
# t_scores.append(target)#
group_scores.append(lps_score)
scores[group] = group_scores
df = pd.DataFrame(data=scores)
return df
# no space in sentences
def wordasso_zh(model, tokenizer, groups, traits, prior_group, tmplt="The <group> person is <mask>."):
scores = {'traits': traits}
for group in groups:
group_scores = []
for trait in traits:
trait_ids = tokenizer.encode(trait, return_tensors='pt').squeeze()
trait_len = list(trait_ids.shape)[0]-2
# print(trait,trait_ids, trait_len)
input_txt = tmplt.replace('<mask>', (tokenizer.mask_token)*trait_len).replace('<group>',prior_group)
# print(input_txt)
prior = get_lognorm_score(model, tokenizer, input_txt, trait)
input_txt = tmplt.replace('<mask>', (tokenizer.mask_token)*trait_len).replace('<group>',group)
# print(input_txt)
target = get_lognorm_score(model, tokenizer, input_txt, trait)
lps_score = target-prior
# p_scores.append(prior)#
# t_scores.append(target)#
group_scores.append(lps_score)
scores[group] = group_scores
df = pd.DataFrame(data=scores)
return df
def main():
pass
if __name__ == "__main__":
main()