-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy patheval.py
executable file
·193 lines (181 loc) · 6.51 KB
/
eval.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
######最後的測試，有加入投票機制######
from sentence_transformers import SentenceTransformer, util
import json
import torch
from transformers import *
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
import math
from tqdm import tqdm
import sys
from sklearn.metrics import f1_score, precision_score, recall_score,accuracy_score
from collections import Counter
#參數設定
faq_len=6331
test_len=633 #測試資料筆數
once_data_len=500
faq_file="Data/faq.txt"
test_file="Data/test_set.txt"
idf_file='Data/idf.txt'
choice_type='cls'
output_log_file='Data/result.txt'
model_dir = 'model_cls/'
output_dir="result.txt"
bert_token='bert-base-chinese'#中文:bert-base-chinese;英文:bert-base-uncased
idf={}
with open(idf_file,'r') as f:
while 1:
line=f.readline()
if len(line)==0 or line=='\n':
break
line=[float(e) for e in line.split()]
idf[line[0]]=line[1]
class Sbert(nn.Module):
def __init__(self,idf):
super(Sbert, self).__init__()
self.bert= BertModel.from_pretrained(model_dir)
self.idf=idf
def forward(self, in1,in1m,pooling='idf'):
loss1, a = self.bert(in1,
token_type_ids=None,
attention_mask=in1m)
#################pooling###########################
if pooling=='idf':
for i in range(len(in1)):
for j in range(100):
if in1m[i][j]==1:
idf_weight=0.0
if int( in1[i][j]) in self.idf:
idf_weight=math.log(719/(1+self.idf[int(in1[i][j])]),2)
else:
idf_weight=math.log(719/1,2)
loss1[i][j]*=idf_weight
input_mask_expanded1 = in1m.unsqueeze(-1).expand(loss1.size()).float()
sum_embeddings1 = torch.sum(loss1 * input_mask_expanded1, 1)
sum_mask1 = torch.clamp(input_mask_expanded1.sum(1), min=1e-9)
output_vector1 = sum_embeddings1 / sum_mask1
if pooling=='avg':
input_mask_expanded1 = in1m.unsqueeze(-1).expand(loss1.size()).float()
sum_embeddings1 = torch.sum(loss1 * input_mask_expanded1, 1)
sum_mask1 = torch.clamp(input_mask_expanded1.sum(1), min=1e-9)
output_vector1 = sum_embeddings1 / sum_mask1
#[cls]token#
if pooling=='cls':
output_vector1=loss1[:, 0, :].float()
#max#
if pooling=='max':
input_mask_expanded1 = in1m.unsqueeze(-1).expand(loss1.size()).float()
loss1[input_mask_expanded1 == 0] = -1e9
output_vector1 = torch.max(loss1, 1)[0]
return output_vector1
tokenizer=BertTokenizer.from_pretrained(bert_token)#中文
model=Sbert(idf)
FAQ=[]
##load faq and testset
with open(faq_file,"r") as f:
for i in range(faq_len):
line=f.readline()
line=[e for e in line.split()]
question=" ".join(line[1:])
answer=f.readline()
FAQ.append([i,int(line[0]),question,answer])
testdata=[]
gold_ls=[]
with open(test_file,"r") as f:
for i in range(test_len):
line=f.readline()
line=[e for e in line.split()]
query="".join(line[-1])
if(line==[]):
break
label=line[0].split(',')
testdata.append([label,query])
gold_ls.append(label)
##make score board
scoreboard=[]
for i in tqdm(range(test_len)):
#print(testdata[i][1]) #測試資料的句子
encoded_dict1=tokenizer.encode_plus(
testdata[i][1],
add_special_tokens = True,
max_length = 100,
pad_to_max_length = True,
return_attention_mask = True,
return_tensors = 'pt',
truncation=True)
invector=model(encoded_dict1['input_ids'],encoded_dict1['attention_mask'],choice_type)
faq_score=[] #存一個Query的排名
k=0
for idx,j in tqdm(enumerate(range(0,faq_len,once_data_len))):
faqvectorls = torch.load('Tensor/faqvector'+str(idx)+'.pt')
for faqvector in faqvectorls:
score=float(torch.cosine_similarity(faqvector,invector))
# if(k==6330):
# break
faq_score.append([FAQ[k][1],score])
k+=1
scoreboard.append(faq_score)#存所有Query的排名
predict_ls=[] #預測結果
##sort each score 以下做投票機制
for i in range(test_len):
Top5_ls=[]
scoreboard[i].sort(key=lambda s: s[1],reverse=True)
Top5=scoreboard[i][:5]#儲存第五名結果
Top1=scoreboard[i][0][0]#儲存第一名結果
for t in Top5:
Top5_ls.append(t[0])
Top5_dict=dict(Counter(Top5_ls))
time=max(zip(Top5_dict.values(),Top5_dict.keys()))[0]
label=max(zip(Top5_dict.values(),Top5_dict.keys()))[1]
if(time>2):
predict_ls.append(label)
else:
predict_ls.append(Top1)
#以下計算準確率
wp = open(output_dir, "w")
count=0
for p,g,test in zip(predict_ls,gold_ls,testdata):
#wp.writelines(str(str(p) in g)+str(p)+','+' '.join(g)+'\n')
if(str(p) in g):
count+=1
else: #輸出預測錯誤的資料
wp.writelines(str(p)+' '+','.join(test[0])+' '+test[1]+'\n')
wp.close()
print("準確率:"+str(count/test_len))
##評估Precision、Recall、F1
# precision = precision_score(gold_ls, predict_ls, average='macro')
# recall = recall_score(gold_ls, predict_ls, average='macro')
# f1 = f1_score(gold_ls,predict_ls, average='macro')
# acc=accuracy_score(gold_ls, predict_ls)
# ##評估mrr map p@5
# mrr=0.0
# p5=0.0
# map=0.0
# for i in range(test_len):
# count=0.0
# ap=0.0
# target_faq=testdata[i][0] #testdata的label資料
# first=True
# for j in range(faq_len):
# if scoreboard[i][j][0]==target_faq:#第i個Query，第j名次，的label資料，如果等於正確標記
# count+=1
# ap+=count/(j+1)
# if first:
# mrr+=1/(j+1)#第一次出現的位置
# first=False#出現過後了把機制關掉
# if j==4:
# p5+=count/5
# map+=(ap/count)
# map/=test_len
# p5/=test_len
# mrr/=test_len
# print(map,p5,mrr)
# with open(output_log_file,'w') as f:
# f.writelines("mAP:"+str(map)+'\n')
# f.writelines("P@5:"+str(p5)+'\n')
# f.writelines("MRR:"+str(mrr)+'\n')
# f.writelines("precision:"+str(precision)+'\n')
# f.writelines("recall:"+str(recall)+'\n')
# f.writelines("f1:"+str(f1)+'\n')
# f.writelines("acc:"+str(acc)+'\n')