/
comput_vqavs_score.py
172 lines (139 loc) · 6.19 KB
/
comput_vqavs_score.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
from os import path as osp
import json
import os
import torch
import argparse
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--input', type=str, default='/root/VQA/baselines/saved_models/LXMERT/test1LXMERT_epoch40.json')
parser.add_argument('--name', type=str, default='test')
parser.add_argument('--dataroot', type=str, default='/root/VQA/baselines/cache')
args = parser.parse_args()
return args
def get_scores(annotations, predictions):
score = 0
count = 0
other_score = 0
yes_no_score = 0
num_score = 0
yes_count = 0
other_count = 0
num_count = 0
upper_bound = 0
upper_bound_num = 0
upper_bound_yes_no = 0
upper_bound_other = 0
for pred, anno in zip(predictions, annotations):
if pred['question_id'] == anno['question_id']:
G_T= max(anno['answer_count'].values())
upper_bound += min(1, G_T / 3)
if pred['answer'] in anno['answers_word']:
proba = anno['answer_count'][pred['answer']]
score += min(1, proba / 3)
count +=1
if anno['answer_type'] == 'yes/no':
yes_no_score += min(1, proba / 3)
upper_bound_yes_no += min(1, G_T / 3)
yes_count +=1
if anno['answer_type'] == 'other':
other_score += min(1, proba / 3)
upper_bound_other += min(1, G_T / 3)
other_count +=1
if anno['answer_type'] == 'number':
num_score += min(1, proba / 3)
upper_bound_num += min(1, G_T / 3)
num_count +=1
else:
score += 0
yes_no_score +=0
other_score +=0
num_score +=0
if anno['answer_type'] == 'yes/no':
upper_bound_yes_no += min(1, G_T / 3)
yes_count +=1
if anno['answer_type'] == 'other':
upper_bound_other += min(1, G_T / 3)
other_count +=1
if anno['answer_type'] == 'number':
upper_bound_num += min(1, G_T / 3)
num_count +=1
return round(score*100/len(annotations),2)
def get_OOD_ans_pred(annotations, predictions, QT_qid, KW_qid, KWP_qid, QTKW_qid, KO_qid, KOP_qid, QTKO_qid, KWKO_qid, QTKWKO_qid):
QT_annotations=[]
KW_annotations=[]
KWP_annotations=[]
QTKW_annotations=[]
KO_annotations=[]
KOP_annotations=[]
QTKO_annotations=[]
KWKO_annotations=[]
QTKWKO_annotations=[]
QT_predictions=[]
KW_predictions=[]
KWP_predictions=[]
QTKW_predictions=[]
KO_predictions=[]
KOP_predictions=[]
QTKO_predictions=[]
KWKO_predictions=[]
QTKWKO_predictions=[]
for x, y in zip(annotations, predictions):
assert x['question_id'] == y['question_id']
if x['question_id'] in QT_qid:
QT_annotations.append(x)
QT_predictions.append(y)
if x['question_id'] in KW_qid:
KW_annotations.append(x)
KW_predictions.append(y)
if x['question_id'] in KWP_qid:
KWP_annotations.append(x)
KWP_predictions.append(y)
if x['question_id'] in QTKW_qid:
QTKW_annotations.append(x)
QTKW_predictions.append(y)
if x['question_id'] in KO_qid:
KO_annotations.append(x)
KO_predictions.append(y)
if x['question_id'] in KOP_qid:
KOP_annotations.append(x)
KOP_predictions.append(y)
if x['question_id'] in QTKO_qid:
QTKO_annotations.append(x)
QTKO_predictions.append(y)
if x['question_id'] in KWKO_qid:
KWKO_annotations.append(x)
KWKO_predictions.append(y)
if x['question_id'] in QTKWKO_qid:
QTKWKO_annotations.append(x)
QTKWKO_predictions.append(y)
return (QT_annotations,KW_annotations,KWP_annotations,QTKW_annotations,KO_annotations,KOP_annotations,QTKO_annotations,KWKO_annotations,QTKWKO_annotations),(QT_predictions,KW_predictions,KWP_predictions,QTKW_predictions,KO_predictions,KOP_predictions,QTKO_predictions,KWKO_predictions,QTKWKO_predictions)
if __name__ == '__main__':
args = parse_args()
#加载标注数据
with open('VQAvs_test_annotations.json', 'r') as f:
test_anno = json.load(f)
annotations = test_anno['annotations']
#加载选手预测结果
predictions = sorted(json.load(open(args.input)), key=lambda x: x['question_id'])
#子指标1(IID score)
iid_score=get_scores(annotations, predictions)
QT_qid, KW_qid, KWP_qid, QTKW_qid, KO_qid, KOP_qid, QTKO_qid, KWKO_qid, QTKWKO_qid = test_anno["QT_qid"] ,test_anno["KW_qid"], test_anno["KWP_qid"], test_anno["QTKW_qid"], test_anno["KO_qid"], test_anno["KOP_qid"], test_anno["QTKO_qid"], test_anno["KWKO_qid"], test_anno["QTKWKO_qid"]
(QT_annotations,KW_annotations,KWP_annotations,QTKW_annotations,KO_annotations,KOP_annotations,QTKO_annotations,KWKO_annotations,QTKWKO_annotations),(QT_predictions,KW_predictions,KWP_predictions,QTKW_predictions,KO_predictions,KOP_predictions,QTKO_predictions,KWKO_predictions,QTKWKO_predictions)=get_OOD_ans_pred(annotations, predictions, QT_qid, KW_qid, KWP_qid, QTKW_qid, KO_qid, KOP_qid, QTKO_qid, KWKO_qid, QTKWKO_qid)
#language-based modality OOD scores, 这4个成绩的均值是子指标2 (OOD score on language-based modality sets)
QT_score=get_scores(QT_annotations, QT_predictions)
KW_score=get_scores(KW_annotations, KW_predictions)
KWP_score=get_scores(KWP_annotations, KWP_predictions)
QTKW_score=get_scores(QTKW_annotations, QTKW_predictions)
#visual-based modality OOD scores, 这2个成绩的均值是子指标3 (OOD score on visual-based modality sets)
KO_score=get_scores(KO_annotations, KO_predictions)
KOP_score=get_scores(KOP_annotations, KOP_predictions)
#cross-modality OOD scores, 这3个成绩的均值是子指标4 (OOD score on cross-modality sets)
QTKO_score=get_scores(QTKO_annotations, QTKO_predictions)
KWKO_score=get_scores(KWKO_annotations, KWKO_predictions)
QTKWKO_score=get_scores(QTKWKO_annotations, QTKWKO_predictions)
print('Final_Score: average score on all OOD test sets\t', (QT_score+KW_score+KWP_score+QTKW_score+KO_score+KOP_score+QTKO_score+KWKO_score+QTKWKO_score)/9)
print('sub-metric 1: IID score', iid_score)
#print('sub-metric 2: average OOD score on language-based modality sets', (QT_score+KW_score+KWP_score+QTKW_score)/4)
#print('sub-metric 3: average OOD score on visual-based modality sets', (KO_score+KOP_score)/2)
#print('sub-metric 4: average OOD score on cross-modality sets', (QTKO_score+KWKO_score+QTKWKO_score)/3)
print("iid_score", iid_score, "QT_score", QT_score, "KW_score", KW_score, "KWP_score", KWP_score, "QTKW_score", QTKW_score, "KO_score", KO_score, "KOP_score", KOP_score, "QTKO_score", QTKO_score, "KWKO_score", KWKO_score, "QTKWKO_score", QTKWKO_score)