Upload 6 files
Browse files- HallusionBench.json +0 -0
- HallusionBench_result_sample.json +0 -0
- evaluation.py +144 -0
- gpt4v_benchmark.py +263 -0
- tsv_to_json.py +75 -0
- utils.py +428 -0
HallusionBench.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
HallusionBench_result_sample.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
evaluation.py
ADDED
@@ -0,0 +1,144 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import csv
|
2 |
+
import json
|
3 |
+
from tqdm import tqdm
|
4 |
+
import numpy as np
|
5 |
+
from prettytable import PrettyTable
|
6 |
+
import os
|
7 |
+
import time
|
8 |
+
from utils import *
|
9 |
+
|
10 |
+
import openai
|
11 |
+
|
12 |
+
|
13 |
+
### to evaluate your method, implement and run generate_answer function!
|
14 |
+
|
15 |
+
root_dir = "."
|
16 |
+
input_file_name = "HallusionBench_result.json"
|
17 |
+
save_json_path_vd = "./hallusion_output_vd_model.json"
|
18 |
+
save_json_path_vs = "./hallusion_output_vs_model.json"
|
19 |
+
# load_json = False
|
20 |
+
load_json = True
|
21 |
+
model_output_entry = "model_prediction"
|
22 |
+
model_correctness_entry = "gpt4v_output_gpt_check"
|
23 |
+
|
24 |
+
|
25 |
+
def generate_answer(data, model_output_entry):
|
26 |
+
|
27 |
+
## TODO
|
28 |
+
## implement this section with yout model!
|
29 |
+
## your_function(img_filename, question) -> "0" (No), "1" (Yes), "2" (Uncertain)
|
30 |
+
# for r in data:
|
31 |
+
# r[model_output_entry] = your_function(r["filename"], r["question"])
|
32 |
+
|
33 |
+
return data
|
34 |
+
|
35 |
+
|
36 |
+
if __name__ == "__main__":
|
37 |
+
|
38 |
+
data_vd = []
|
39 |
+
data_vs = []
|
40 |
+
with open(input_file_name) as json_file:
|
41 |
+
datas = json.load(json_file)
|
42 |
+
|
43 |
+
for data in tqdm(datas):
|
44 |
+
if data['category'] == 'VD':
|
45 |
+
data_vd.append(data)
|
46 |
+
if data['category'] == 'VS':
|
47 |
+
data_vs.append(data)
|
48 |
+
|
49 |
+
data_vd = evaluate_by_chatgpt(data_vd, model_output_entry, model_correctness_entry, load_json=load_json, save_json_path=save_json_path_vd)
|
50 |
+
data_vd = check_same_by_chatgpt(data_vd, model_output_entry, load_json=load_json, save_json_path=save_json_path_vd)
|
51 |
+
#time.sleep(60) #
|
52 |
+
try:
|
53 |
+
data_vs = evaluate_by_chatgpt(data_vs, model_output_entry, model_correctness_entry, load_json=load_json, save_json_path=save_json_path_vs)
|
54 |
+
data_vs = check_same_by_chatgpt(data_vs, model_output_entry, load_json=load_json, save_json_path=save_json_path_vs)
|
55 |
+
except:
|
56 |
+
time.sleep(60)
|
57 |
+
data_vs = evaluate_by_chatgpt(data_vs, model_output_entry, model_correctness_entry, load_json=load_json, save_json_path=save_json_path_vs)
|
58 |
+
data_vs = check_same_by_chatgpt(data_vs, model_output_entry, load_json=load_json, save_json_path=save_json_path_vs)
|
59 |
+
|
60 |
+
print("##### GPT Evaluate #####")
|
61 |
+
|
62 |
+
data_vd = assign_correctness(data_vd, correctness_entry=model_correctness_entry)
|
63 |
+
data_vs = assign_correctness(data_vs, correctness_entry=model_correctness_entry)
|
64 |
+
data = data_vd + data_vs
|
65 |
+
|
66 |
+
all_data = get_eval_all(data, model_correctness_entry)
|
67 |
+
all_vd = get_eval_all(data_vd, model_correctness_entry)
|
68 |
+
all_vs = get_eval_all(data_vs, model_correctness_entry)
|
69 |
+
|
70 |
+
table1 = [["per question", "Total"],
|
71 |
+
["VD", round(100 * all_vd["correct"]/all_vd["total"], 4)],
|
72 |
+
["VS", round(100 * all_vs["correct"]/all_vs["total"], 4)],
|
73 |
+
["Overall", round(100 * all_data["correct"]/all_data["total"], 4)]]
|
74 |
+
tab1 = PrettyTable(table1[0])
|
75 |
+
tab1.add_rows(table1[1:])
|
76 |
+
|
77 |
+
|
78 |
+
q_acc_gpt = round(100 * all_data["correct"]/all_data["total"], 4)
|
79 |
+
|
80 |
+
all_data = get_eval_pair_all(data, model_correctness_entry)
|
81 |
+
easy = get_eval_pair_easy(data)
|
82 |
+
hard = get_eval_pair_hard(data)
|
83 |
+
all_vd = get_eval_pair_all(data_vd, model_correctness_entry)
|
84 |
+
easy_vd = get_eval_pair_easy(data_vd)
|
85 |
+
hard_vd = get_eval_pair_hard(data_vd)
|
86 |
+
all_vs = get_eval_pair_all(data_vs, model_correctness_entry)
|
87 |
+
easy_vs = get_eval_pair_easy(data_vs)
|
88 |
+
hard_vs = get_eval_pair_hard(data_vs)
|
89 |
+
# question pair level
|
90 |
+
table3 = [["per question pair", "Easy", "Hard", "Total"],
|
91 |
+
["VD", round(100 * easy_vd["correct"]/easy_vd["total"], 4), round(100 * hard_vd["correct"]/hard_vd["total"], 4), round(100 * all_vd["correct"]/all_vd["total"], 4)],
|
92 |
+
["VS", round(100 * easy_vs["correct"]/easy_vs["total"], 4), round(100 * hard_vs["correct"]/hard_vs["total"], 4), round(100 * all_vs["correct"]/all_vs["total"], 4)],
|
93 |
+
["Overall", round(100 * easy["correct"]/easy["total"], 4), round(100 * hard["correct"]/hard["total"], 4), round(100 * all_data["correct"]/all_data["total"], 4)]]
|
94 |
+
tab3 = PrettyTable(table3[0])
|
95 |
+
tab3.add_rows(table3[1:])
|
96 |
+
#print(tab3)
|
97 |
+
|
98 |
+
|
99 |
+
fig_all = get_eval_fig(data)
|
100 |
+
fig_vd = get_eval_fig(data_vd)
|
101 |
+
fig_vs = get_eval_fig(data_vs)
|
102 |
+
|
103 |
+
# image level
|
104 |
+
table2 = [["per figure", "Correct", "Wrong", "Score"],
|
105 |
+
["VD", round(100 * fig_vd["correct"]/fig_vd["total"], 4), round(100 * fig_vd["inconsistent"]/fig_vd["total"], 4) + round(100 * fig_vd["wrong"]/fig_vd["total"], 4), round(fig_vd["score"], 4)],
|
106 |
+
["VS", round(100 * fig_vs["correct"]/fig_vs["total"], 4), round(100 * fig_vs["inconsistent"]/fig_vs["total"], 4) + round(100 * fig_vs["wrong"]/fig_vs["total"], 4), round(fig_vs["score"], 4)],
|
107 |
+
["Overall", round(100 * fig_all["correct"]/fig_all["total"], 4), round(100 * fig_all["inconsistent"]/fig_all["total"], 4) + round(100 * fig_all["wrong"]/fig_all["total"], 4), round(fig_all["score"], 4)]]
|
108 |
+
tab2 = PrettyTable(table2[0])
|
109 |
+
tab2.add_rows(table2[1:])
|
110 |
+
|
111 |
+
pair_acc_gpt = round(100 * all_data["correct"]/all_data["total"], 4)
|
112 |
+
figure_acc_gpt = round(100 * fig_all["correct"]/fig_all["total"], 4)
|
113 |
+
easy_acc_gpt = round(100 * easy["correct"]/easy["total"], 4)
|
114 |
+
hard_acc_gpt = round(100 * hard["correct"]/hard["total"], 4)
|
115 |
+
|
116 |
+
|
117 |
+
|
118 |
+
print("##### Question Stats #####")
|
119 |
+
print("Easy Questions: " + str(easy_vd["total_q"]) + "(Visual Dependent) + " + str(easy_vs["total_q"]) + "(Visual Supplement)")
|
120 |
+
print("Hard Questions: " + str(hard_vd["total_q"]) + "(Visual Dependent) + " + str(hard_vs["total_q"]) + "(Visual Supplement)")
|
121 |
+
print("Total Questions: " + str(all_data["total_q"]))
|
122 |
+
|
123 |
+
|
124 |
+
print("##### Figure Stats #####")
|
125 |
+
print("Visual Dependent Figures: " + str(fig_vd["total"]))
|
126 |
+
print("Visual Supplement Figures: " + str(fig_vs["total"]))
|
127 |
+
print("Total Figures: " + str(fig_all["total"]))
|
128 |
+
|
129 |
+
print("##### Leaderboard Stats #####")
|
130 |
+
|
131 |
+
table = [["", "Acc per question pair (qAcc)", "Acc per figure (fAcc)", "Acc per easy question (easy aAcc)", "Acc per hard question (hard aAcc)", "Acc per question (aAcc)"],
|
132 |
+
["GPT Eval", pair_acc_gpt, figure_acc_gpt, easy_acc_gpt, hard_acc_gpt, q_acc_gpt]]
|
133 |
+
leaderboard = PrettyTable(table[0])
|
134 |
+
leaderboard.add_rows(table[1:])
|
135 |
+
print(leaderboard)
|
136 |
+
|
137 |
+
|
138 |
+
stats = yes_ratio_stats(data)
|
139 |
+
|
140 |
+
table = [["", "Yes/No Bias (Pct Diff)", "Yes/No Bias (FP Ratio)", "Consistency Test (correct)", "Consistency Test (inconsistent)", "Consistency Test (wrong)", "LH", "VI", "Mixed"],
|
141 |
+
["GPT Eval", stats["diff"], stats["fp"], round(100 * fig_all["correct"]/fig_all["total"], 4), round(100 * fig_all["inconsistent"]/fig_all["total"], 4), round(100 * fig_all["wrong"]/fig_all["total"], 4), round(100 * all_data["LH_cg"]/(all_data["LH_cg"] + all_data["VI_cg"] + all_data["Mix_cg"]), 4), round(100 * all_data["VI_cg"]/(all_data["LH_cg"] + all_data["VI_cg"] + all_data["Mix_cg"]), 4), round(100 * all_data["Mix_cg"]/(all_data["LH_cg"] + all_data["VI_cg"] + all_data["Mix_cg"]), 4)]]
|
142 |
+
test = PrettyTable(table[0])
|
143 |
+
test.add_rows(table[1:])
|
144 |
+
print(test)
|
gpt4v_benchmark.py
ADDED
@@ -0,0 +1,263 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import csv
|
2 |
+
import json
|
3 |
+
from tqdm import tqdm
|
4 |
+
import numpy as np
|
5 |
+
from prettytable import PrettyTable
|
6 |
+
import os
|
7 |
+
from utils import *
|
8 |
+
|
9 |
+
import openai
|
10 |
+
|
11 |
+
### to evaluate your method, implement and run generate_answer function!
|
12 |
+
|
13 |
+
root_dir = "."
|
14 |
+
llava = False
|
15 |
+
# llava = True
|
16 |
+
# load_json = False
|
17 |
+
load_json = True
|
18 |
+
input_file_name = "HallusionBench.tsv"
|
19 |
+
|
20 |
+
save_json_path_vd = "./hallusion_output_vd.json"
|
21 |
+
save_json_path_vs = "./hallusion_output_vs.json"
|
22 |
+
model_output_entry = "gpt4v_output"
|
23 |
+
model_correctness_entry = "gpt4v_output_gpt_check"
|
24 |
+
model_correctness_entry_human = "gpt4v_output_human_check"
|
25 |
+
|
26 |
+
if llava:
|
27 |
+
save_json_path_vd = "./hallusion_output_vd_llava.json"
|
28 |
+
save_json_path_vs = "./hallusion_output_vs_llava.json"
|
29 |
+
model_output_entry = "llava_1_5_output"
|
30 |
+
model_correctness_entry = "llava_1_5_output_gpt_check"
|
31 |
+
model_correctness_entry_human = "llava_1_5_output_human_check"
|
32 |
+
|
33 |
+
|
34 |
+
col_idx = {
|
35 |
+
'category':0,
|
36 |
+
'subcategory':1,
|
37 |
+
'visual_input':2,
|
38 |
+
'set_id':3,
|
39 |
+
'figure_id':4,
|
40 |
+
'sample_note':5,
|
41 |
+
'question_id':6,
|
42 |
+
'question':7,
|
43 |
+
'gt_answer_details':8,
|
44 |
+
'gt_answer':9,
|
45 |
+
'gpt4v_output':10,
|
46 |
+
'gpt4v_output_human_check': 11,
|
47 |
+
'llava_1_5_output':12,
|
48 |
+
'llava_1_5_output_human_check': 13,
|
49 |
+
}
|
50 |
+
|
51 |
+
|
52 |
+
|
53 |
+
def generate_answer(data, model_output_entry):
|
54 |
+
|
55 |
+
## TODO
|
56 |
+
## implement this section with yout model!
|
57 |
+
## your_function(img_filename, question) -> "0" (No), "1" (Yes), "2" (Uncertain)
|
58 |
+
# for r in data:
|
59 |
+
# r[model_output_entry] = your_function(r["filename"], r["question"])
|
60 |
+
|
61 |
+
return data
|
62 |
+
|
63 |
+
|
64 |
+
|
65 |
+
if __name__ == "__main__":
|
66 |
+
|
67 |
+
data_vd = []
|
68 |
+
data_vs = []
|
69 |
+
with open(input_file_name) as file:
|
70 |
+
tsv_file = csv.reader(file, delimiter="\t")
|
71 |
+
flag = 0
|
72 |
+
for line in tsv_file:
|
73 |
+
if line[0] not in ["VD", "VS"]:
|
74 |
+
continue
|
75 |
+
data_dict = {}
|
76 |
+
for k, v in col_idx.items():
|
77 |
+
data_dict[k] = line[v]
|
78 |
+
|
79 |
+
data_dict["filename"] = get_image_file_location(root_dir, data_dict)
|
80 |
+
if line[0] == "VD":
|
81 |
+
data_vd.append(data_dict)
|
82 |
+
else:
|
83 |
+
data_vs.append(data_dict)
|
84 |
+
|
85 |
+
## TODO
|
86 |
+
data_vd = generate_answer(data_vd, model_output_entry)
|
87 |
+
data_vs = generate_answer(data_vs, model_output_entry)
|
88 |
+
## END
|
89 |
+
|
90 |
+
data_vd = evaluate_by_chatgpt(data_vd, model_output_entry, model_correctness_entry, load_json=load_json, save_json_path=save_json_path_vd)
|
91 |
+
data_vd = check_same_by_chatgpt(data_vd, model_output_entry, load_json=load_json, save_json_path=save_json_path_vd)
|
92 |
+
data_vs = evaluate_by_chatgpt(data_vs, model_output_entry, model_correctness_entry, load_json=load_json, save_json_path=save_json_path_vs)
|
93 |
+
data_vs = check_same_by_chatgpt(data_vs, model_output_entry, load_json=load_json, save_json_path=save_json_path_vs)
|
94 |
+
|
95 |
+
data_vd = assign_correctness(data_vd, correctness_entry=model_correctness_entry_human)
|
96 |
+
data_vs = assign_correctness(data_vs, correctness_entry=model_correctness_entry_human)
|
97 |
+
data = data_vd + data_vs
|
98 |
+
|
99 |
+
all_data = get_eval_all(data, model_correctness_entry_human)
|
100 |
+
all_vd = get_eval_all(data_vd, model_correctness_entry_human)
|
101 |
+
all_vs = get_eval_all(data_vs, model_correctness_entry_human)
|
102 |
+
human_check_correctness = [i["correct"] for i in data]
|
103 |
+
|
104 |
+
print("##### Human Evaluate #####")
|
105 |
+
|
106 |
+
# question level
|
107 |
+
table1 = [["per question", "Total"],
|
108 |
+
["VD", round(100 * all_vd["correct"]/all_vd["total"], 4)],
|
109 |
+
["VS", round(100 * all_vs["correct"]/all_vs["total"], 4)],
|
110 |
+
["Overall", round(100 * all_data["correct"]/all_data["total"], 4)]]
|
111 |
+
tab1 = PrettyTable(table1[0])
|
112 |
+
tab1.add_rows(table1[1:])
|
113 |
+
print(tab1)
|
114 |
+
q_acc_human = round(100 * all_data["correct"]/all_data["total"], 4)
|
115 |
+
|
116 |
+
|
117 |
+
all_data = get_eval_pair_all(data, model_correctness_entry_human)
|
118 |
+
easy = get_eval_pair_easy(data)
|
119 |
+
hard = get_eval_pair_hard(data)
|
120 |
+
all_vd = get_eval_pair_all(data_vd, model_correctness_entry_human)
|
121 |
+
easy_vd = get_eval_pair_easy(data_vd)
|
122 |
+
hard_vd = get_eval_pair_hard(data_vd)
|
123 |
+
all_vs = get_eval_pair_all(data_vs, model_correctness_entry_human)
|
124 |
+
easy_vs = get_eval_pair_easy(data_vs)
|
125 |
+
hard_vs = get_eval_pair_hard(data_vs)
|
126 |
+
# question pair level
|
127 |
+
table3 = [["per question pair", "Easy", "Hard", "Total"],
|
128 |
+
["VD", round(100 * easy_vd["correct"]/easy_vd["total"], 4), round(100 * hard_vd["correct"]/hard_vd["total"], 4), round(100 * all_vd["correct"]/all_vd["total"], 4)],
|
129 |
+
["VS", round(100 * easy_vs["correct"]/easy_vs["total"], 4), round(100 * hard_vs["correct"]/hard_vs["total"], 4), round(100 * all_vs["correct"]/all_vs["total"], 4)],
|
130 |
+
["Overall", round(100 * easy["correct"]/easy["total"], 4), round(100 * hard["correct"]/hard["total"], 4), round(100 * all_data["correct"]/all_data["total"], 4)]]
|
131 |
+
tab3 = PrettyTable(table3[0])
|
132 |
+
tab3.add_rows(table3[1:])
|
133 |
+
print(tab3)
|
134 |
+
|
135 |
+
|
136 |
+
fig_all = get_eval_fig(data)
|
137 |
+
fig_vd = get_eval_fig(data_vd)
|
138 |
+
fig_vs = get_eval_fig(data_vs)
|
139 |
+
fig_all_human = fig_all
|
140 |
+
all_data_human = all_data
|
141 |
+
# image level
|
142 |
+
table2 = [["per figure", "Correct", "Inconsistant", "Wrong", "Score"],
|
143 |
+
["VD", round(100 * fig_vd["correct"]/fig_vd["total"], 4), round(100 * fig_vd["inconsistent"]/fig_vd["total"], 4), round(100 * fig_vd["wrong"]/fig_vd["total"], 4), round(100 * fig_vd["score"], 4)],
|
144 |
+
["VS", round(100 * fig_vs["correct"]/fig_vs["total"], 4), round(100 * fig_vs["inconsistent"]/fig_vs["total"], 4), round(100 * fig_vs["wrong"]/fig_vs["total"], 4), round(100 * fig_vs["score"], 4)],
|
145 |
+
["Overall", round(100 * fig_all["correct"]/fig_all["total"], 4), round(100 * fig_all["inconsistent"]/fig_all["total"], 4), round(100 * fig_all["wrong"]/fig_all["total"], 4), round(100 * fig_all["score"], 4)]]
|
146 |
+
tab2 = PrettyTable(table2[0])
|
147 |
+
tab2.add_rows(table2[1:])
|
148 |
+
print(tab2)
|
149 |
+
|
150 |
+
pair_acc_human = round(100 * all_data["correct"]/all_data["total"], 4)
|
151 |
+
figure_acc_human = round(100 * fig_all["correct"]/fig_all["total"], 4)
|
152 |
+
easy_acc_human = round(100 * easy["correct"]/easy["total"], 4)
|
153 |
+
hard_acc_human = round(100 * hard["correct"]/hard["total"], 4)
|
154 |
+
|
155 |
+
stats_human = yes_ratio_stats(data)
|
156 |
+
|
157 |
+
# from IPython import embed;embed()
|
158 |
+
############################################
|
159 |
+
|
160 |
+
print("##### GPT Evaluate #####")
|
161 |
+
|
162 |
+
data_vd = assign_correctness(data_vd, correctness_entry=model_correctness_entry)
|
163 |
+
data_vs = assign_correctness(data_vs, correctness_entry=model_correctness_entry)
|
164 |
+
data = data_vd + data_vs
|
165 |
+
|
166 |
+
all_data = get_eval_all(data, model_correctness_entry)
|
167 |
+
all_vd = get_eval_all(data_vd, model_correctness_entry)
|
168 |
+
all_vs = get_eval_all(data_vs, model_correctness_entry)
|
169 |
+
gpt_check_correctness = [i["correct"] for i in data]
|
170 |
+
|
171 |
+
# question level
|
172 |
+
table1 = [["per question", "Total"],
|
173 |
+
["VD", round(100 * all_vd["correct"]/all_vd["total"], 4)],
|
174 |
+
["VS", round(100 * all_vs["correct"]/all_vs["total"], 4)],
|
175 |
+
["Overall", round(100 * all_data["correct"]/all_data["total"], 4)]]
|
176 |
+
tab1 = PrettyTable(table1[0])
|
177 |
+
tab1.add_rows(table1[1:])
|
178 |
+
print(tab1)
|
179 |
+
|
180 |
+
q_acc_gpt = round(100 * all_data["correct"]/all_data["total"], 4)
|
181 |
+
|
182 |
+
all_data = get_eval_pair_all(data, model_correctness_entry)
|
183 |
+
easy = get_eval_pair_easy(data)
|
184 |
+
hard = get_eval_pair_hard(data)
|
185 |
+
all_vd = get_eval_pair_all(data_vd, model_correctness_entry)
|
186 |
+
easy_vd = get_eval_pair_easy(data_vd)
|
187 |
+
hard_vd = get_eval_pair_hard(data_vd)
|
188 |
+
all_vs = get_eval_pair_all(data_vs, model_correctness_entry)
|
189 |
+
easy_vs = get_eval_pair_easy(data_vs)
|
190 |
+
hard_vs = get_eval_pair_hard(data_vs)
|
191 |
+
# question pair level
|
192 |
+
table3 = [["per question pair", "Easy", "Hard", "Total"],
|
193 |
+
["VD", round(100 * easy_vd["correct"]/easy_vd["total"], 4), round(100 * hard_vd["correct"]/hard_vd["total"], 4), round(100 * all_vd["correct"]/all_vd["total"], 4)],
|
194 |
+
["VS", round(100 * easy_vs["correct"]/easy_vs["total"], 4), round(100 * hard_vs["correct"]/hard_vs["total"], 4), round(100 * all_vs["correct"]/all_vs["total"], 4)],
|
195 |
+
["Overall", round(100 * easy["correct"]/easy["total"], 4), round(100 * hard["correct"]/hard["total"], 4), round(100 * all_data["correct"]/all_data["total"], 4)]]
|
196 |
+
tab3 = PrettyTable(table3[0])
|
197 |
+
tab3.add_rows(table3[1:])
|
198 |
+
print(tab3)
|
199 |
+
|
200 |
+
|
201 |
+
fig_all = get_eval_fig(data)
|
202 |
+
fig_vd = get_eval_fig(data_vd)
|
203 |
+
fig_vs = get_eval_fig(data_vs)
|
204 |
+
|
205 |
+
# image level
|
206 |
+
table2 = [["per figure", "Correct", "Wrong", "Score"],
|
207 |
+
["VD", round(100 * fig_vd["correct"]/fig_vd["total"], 4), round(100 * fig_vd["inconsistent"]/fig_vd["total"], 4) + round(100 * fig_vd["wrong"]/fig_vd["total"], 4), round(fig_vd["score"], 4)],
|
208 |
+
["VS", round(100 * fig_vs["correct"]/fig_vs["total"], 4), round(100 * fig_vs["inconsistent"]/fig_vs["total"], 4) + round(100 * fig_vs["wrong"]/fig_vs["total"], 4), round(fig_vs["score"], 4)],
|
209 |
+
["Overall", round(100 * fig_all["correct"]/fig_all["total"], 4), round(100 * fig_all["inconsistent"]/fig_all["total"], 4) + round(100 * fig_all["wrong"]/fig_all["total"], 4), round(fig_all["score"], 4)]]
|
210 |
+
tab2 = PrettyTable(table2[0])
|
211 |
+
tab2.add_rows(table2[1:])
|
212 |
+
print(tab2)
|
213 |
+
|
214 |
+
pair_acc_gpt = round(100 * all_data["correct"]/all_data["total"], 4)
|
215 |
+
figure_acc_gpt = round(100 * fig_all["correct"]/fig_all["total"], 4)
|
216 |
+
easy_acc_gpt = round(100 * easy["correct"]/easy["total"], 4)
|
217 |
+
hard_acc_gpt = round(100 * hard["correct"]/hard["total"], 4)
|
218 |
+
|
219 |
+
##############################
|
220 |
+
|
221 |
+
|
222 |
+
print("##### Question Stats #####")
|
223 |
+
print("Easy Questions: " + str(easy_vd["total_q"]) + "(Visual Dependent) + " + str(easy_vs["total_q"]) + "(Visual Supplement)")
|
224 |
+
print("Hard Questions: " + str(hard_vd["total_q"]) + "(Visual Dependent) + " + str(hard_vs["total_q"]) + "(Visual Supplement)")
|
225 |
+
print("Total Questions: " + str(all_data["total_q"]))
|
226 |
+
|
227 |
+
|
228 |
+
print("##### Figure Stats #####")
|
229 |
+
print("Visual Dependent Figures: " + str(fig_vd["total"]))
|
230 |
+
print("Visual Supplement Figures: " + str(fig_vs["total"]))
|
231 |
+
print("Total Figures: " + str(fig_all["total"]))
|
232 |
+
|
233 |
+
print("##### Leaderboard Stats #####")
|
234 |
+
|
235 |
+
table = [["", "Acc per question pair (qAcc)", "Acc per figure (fAcc)", "Acc per easy question (easy aAcc)", "Acc per hard question (hard aAcc)", "Acc per question (aAcc)"],
|
236 |
+
["Human Eval", pair_acc_human, figure_acc_human, easy_acc_human, hard_acc_human, q_acc_human],
|
237 |
+
["GPT Eval", pair_acc_gpt, figure_acc_gpt, easy_acc_gpt, hard_acc_gpt, q_acc_gpt]]
|
238 |
+
leaderboard = PrettyTable(table[0])
|
239 |
+
leaderboard.add_rows(table[1:])
|
240 |
+
print(leaderboard)
|
241 |
+
|
242 |
+
print(all_data["total"], all_data["wrong"], all_data["LH"], all_data["VI"], all_data["Mix"])
|
243 |
+
print(all_data["total_q"], all_data["LH_cg"], all_data["VI_cg"], all_data["Mix_cg"])
|
244 |
+
|
245 |
+
print(len(gpt_check_correctness))
|
246 |
+
print(len(human_check_correctness))
|
247 |
+
print(sum(np.array(human_check_correctness) == np.array(gpt_check_correctness)))
|
248 |
+
print(sum(np.array(human_check_correctness) == np.array(gpt_check_correctness)) / len(gpt_check_correctness))
|
249 |
+
|
250 |
+
|
251 |
+
yes = [int(i["gt_answer"]) for i in data]
|
252 |
+
print(sum(yes))
|
253 |
+
print(len(yes))
|
254 |
+
print(sum(yes)/len(yes))
|
255 |
+
|
256 |
+
stats_gpt = yes_ratio_stats(data)
|
257 |
+
|
258 |
+
table = [["", "Yes/No Bias (Pct Diff)", "Yes/No Bias (FP Ratio)", "Consistency Test (correct)", "Consistency Test (inconsistent)", "Consistency Test (wrong)", "LH", "VI", "Mixed"],
|
259 |
+
["Human Eval", stats_human["diff"], stats_human["fp"], round(100 * fig_all_human["correct"]/fig_all_human["total"], 4), round(100 * fig_all_human["inconsistent"]/fig_all_human["total"], 4), round(100 * fig_all_human["wrong"]/fig_all_human["total"], 4), round(100 * all_data_human["LH_cg"]/(all_data_human["LH_cg"] + all_data_human["VI_cg"] + all_data_human["Mix_cg"]), 4), round(100 * all_data_human["VI_cg"]/(all_data_human["LH_cg"] + all_data_human["VI_cg"] + all_data_human["Mix_cg"]), 4), round(100 * all_data_human["Mix_cg"]/(all_data_human["LH_cg"] + all_data_human["VI_cg"] + all_data_human["Mix_cg"]), 4)],
|
260 |
+
["GPT Eval", stats_gpt["diff"], stats_gpt["fp"], round(100 * fig_all["correct"]/fig_all["total"], 4), round(100 * fig_all["inconsistent"]/fig_all["total"], 4), round(100 * fig_all["wrong"]/fig_all["total"], 4), round(100 * all_data["LH_cg"]/(all_data["LH_cg"] + all_data["VI_cg"] + all_data["Mix_cg"]), 4), round(100 * all_data["VI_cg"]/(all_data["LH_cg"] + all_data["VI_cg"] + all_data["Mix_cg"]), 4), round(100 * all_data["Mix_cg"]/(all_data["LH_cg"] + all_data["VI_cg"] + all_data["Mix_cg"]), 4)]]
|
261 |
+
test = PrettyTable(table[0])
|
262 |
+
test.add_rows(table[1:])
|
263 |
+
print(test)
|
tsv_to_json.py
ADDED
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
from tqdm import tqdm
|
3 |
+
import os
|
4 |
+
|
5 |
+
def get_image_file_location(root, row):
|
6 |
+
if int(row['visual_input']) == 0:
|
7 |
+
return None
|
8 |
+
img_file = row['set_id'] + "_" + row['figure_id'] + ".png"
|
9 |
+
return os.path.join(root, row['category'], row['subcategory'], img_file)
|
10 |
+
col_idx = {
|
11 |
+
'category':0,
|
12 |
+
'subcategory':1,
|
13 |
+
'visual_input':2,
|
14 |
+
'set_id':3,
|
15 |
+
'figure_id':4,
|
16 |
+
'sample_note':5,
|
17 |
+
'question_id':6,
|
18 |
+
'question':7,
|
19 |
+
'gt_answer_details':8,
|
20 |
+
'gt_answer':9,
|
21 |
+
# 'gpt4v_output':10,
|
22 |
+
# 'gpt4v_output_human_check': 11,
|
23 |
+
# 'llava_1_5_output':12,
|
24 |
+
# 'llava_1_5_output_human_check': 13,
|
25 |
+
}
|
26 |
+
import csv
|
27 |
+
import json
|
28 |
+
data_vd = []
|
29 |
+
data_vs = []
|
30 |
+
root_dir = "."
|
31 |
+
input_file_name = 'HallusionBench.tsv'
|
32 |
+
with open(input_file_name) as file:
|
33 |
+
tsv_file = csv.reader(file, delimiter="\t")
|
34 |
+
flag = 0
|
35 |
+
for line in tsv_file:
|
36 |
+
# if line[0] not in ["VD", "VS"]:
|
37 |
+
# if line[0] in ["NOTE", "category"]:
|
38 |
+
if "VD" not in line[0] and "VS" not in line[0]:
|
39 |
+
continue
|
40 |
+
data_dict = {}
|
41 |
+
try:
|
42 |
+
for k, v in col_idx.items():
|
43 |
+
data_dict[k] = line[v]
|
44 |
+
assert int(line[col_idx["gt_answer"]]) == 0 or int(line[col_idx["gt_answer"]]) == 1 or int(line[col_idx["gt_answer"]]) == 2
|
45 |
+
# assert int(line[col_idx["gpt4v_output_human_check"]]) == 0 or int(line[col_idx["gpt4v_output_human_check"]]) == 1 or int(line[col_idx["gpt4v_output_human_check"]]) == 2
|
46 |
+
# assert int(line[col_idx["llava_1_5_output_human_check"]]) == 0 or int(line[col_idx["llava_1_5_output_human_check"]]) == 1 or int(line[col_idx["llava_1_5_output_human_check"]]) == 2
|
47 |
+
except:
|
48 |
+
from IPython import embed;embed()
|
49 |
+
|
50 |
+
data_dict["filename"] = get_image_file_location(root_dir, data_dict)
|
51 |
+
if line[0] == "VD":
|
52 |
+
data_vd.append(data_dict)
|
53 |
+
else:
|
54 |
+
data_vs.append(data_dict)
|
55 |
+
|
56 |
+
result = data_vs + data_vd
|
57 |
+
print(len(result))
|
58 |
+
result1 = []
|
59 |
+
|
60 |
+
for re in result:
|
61 |
+
result1.append({"category": re['category'].strip(),
|
62 |
+
"subcategory": re['subcategory'].strip(),
|
63 |
+
"visual_input": re['visual_input'].strip(),
|
64 |
+
"set_id": re['set_id'].strip(),
|
65 |
+
"figure_id": re['figure_id'].strip(),
|
66 |
+
"sample_note": re['sample_note'].strip(),
|
67 |
+
"question_id": re['question_id'].strip(),
|
68 |
+
"question": re['question'].strip(),
|
69 |
+
"gt_answer": re['gt_answer'].strip(),
|
70 |
+
"gt_answer_details": re["gt_answer_details"].strip(),
|
71 |
+
"filename": re['filename'].strip() if re['filename'] else re['filename']})
|
72 |
+
|
73 |
+
print(len(result))
|
74 |
+
with open('./HallusionBench.json', 'w') as f:
|
75 |
+
json.dump(result, f)
|
utils.py
ADDED
@@ -0,0 +1,428 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import csv
|
2 |
+
import json
|
3 |
+
from tqdm import tqdm
|
4 |
+
import numpy as np
|
5 |
+
from prettytable import PrettyTable
|
6 |
+
import os
|
7 |
+
import time
|
8 |
+
import openai
|
9 |
+
import threading
|
10 |
+
|
11 |
+
try:
|
12 |
+
with open("apikey.txt", "r") as f:
|
13 |
+
api_key = f.read()
|
14 |
+
except:
|
15 |
+
api_key = ''
|
16 |
+
|
17 |
+
|
18 |
+
def get_image_file_location(root, row):
|
19 |
+
if int(row['visual_input']) == 0:
|
20 |
+
return None
|
21 |
+
img_file = row['set_id'] + "_" + row['figure_id'] + ".png"
|
22 |
+
return os.path.join(root, row['category'], row['subcategory'], img_file)
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
def evaluate_by_chatgpt(data, output_entry, correctness_entry, gpt_model="gpt-4", load_json=False, save_json_path="./hallusion_output.json"):
|
27 |
+
if load_json and os.path.exists(save_json_path):
|
28 |
+
with open(save_json_path, 'r') as f:
|
29 |
+
output = json.load(f)
|
30 |
+
else:
|
31 |
+
output = []
|
32 |
+
for sample in tqdm(data[len(output):]):
|
33 |
+
prompt = 'Imagine you are an intelligent teacher. Thoroughly read the question, reference answer and the prediction answer to ensure a clear understanding of the information provided. Assess the correctness of the predictions. '
|
34 |
+
prompt += 'If the prediction answer does not conflict with the reference answer, please generate “correct”. If the prediction answer conflict with the reference answer, please generate “incorrect”. If the prediction answer is unclear about the answer, please generate "unclear". \n\n Question:'
|
35 |
+
prompt += sample['question']
|
36 |
+
prompt += '\nReference answer: '
|
37 |
+
prompt += sample['gt_answer_details']
|
38 |
+
prompt += '\nPrediction answer:'
|
39 |
+
prompt += sample[output_entry]
|
40 |
+
prompt += '\nOutput:'
|
41 |
+
|
42 |
+
# https://github.com/openai/openai-python/issues/322#issuecomment-1767841683
|
43 |
+
while True:
|
44 |
+
try:
|
45 |
+
response = openai.ChatCompletion.create(
|
46 |
+
model=gpt_model,
|
47 |
+
messages=[{"role": "user", "content": prompt}],
|
48 |
+
api_key=api_key,
|
49 |
+
request_timeout=5)
|
50 |
+
break
|
51 |
+
except:
|
52 |
+
print("Timeout, retrying...")
|
53 |
+
time.sleep(5) # Wait for 5 seconds before retrying
|
54 |
+
|
55 |
+
output_text = response['choices'][0]['message']['content']
|
56 |
+
|
57 |
+
|
58 |
+
if 'incorrect' in output_text.lower():
|
59 |
+
gpt_correctness = "0"
|
60 |
+
|
61 |
+
elif 'correct' in output_text.lower():
|
62 |
+
gpt_correctness = "1"
|
63 |
+
else:
|
64 |
+
gpt_correctness = "2"
|
65 |
+
|
66 |
+
sample[correctness_entry] = gpt_correctness
|
67 |
+
|
68 |
+
output.append(sample)
|
69 |
+
|
70 |
+
with open(save_json_path, 'w') as f:
|
71 |
+
json.dump(output, f)
|
72 |
+
|
73 |
+
return output
|
74 |
+
|
75 |
+
def check_same_by_chatgpt(data, output_entry, gpt_model="gpt-4", load_json=False, save_json_path="./hallusion_output.json"):
|
76 |
+
|
77 |
+
orig_response = {}
|
78 |
+
|
79 |
+
for r in data:
|
80 |
+
if str(r["figure_id"]) == "0":
|
81 |
+
key = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
82 |
+
orig_response[key] = r[output_entry]
|
83 |
+
|
84 |
+
for sample in tqdm(data):
|
85 |
+
if "same" not in sample.keys():
|
86 |
+
key = "_".join([sample["category"], sample["subcategory"], str(sample["set_id"]), str(sample["question_id"])])
|
87 |
+
response2 = orig_response[key]
|
88 |
+
|
89 |
+
prompt = 'Imagine you are an intelligent teacher. Thoroughly read the two responses to two different questions. Assess the consistency of the information provided within those two responses. '
|
90 |
+
prompt += 'You do not know the specific questions, but you can asssess the consistency among the two responses by checking for logical conflicts if both responses are correct. '
|
91 |
+
prompt += 'If response1 does not conflict with response2, please generate “same”. Otherwise, generate "different". \n\n response1:'
|
92 |
+
prompt += sample[output_entry]
|
93 |
+
prompt += '\nresponse2: '
|
94 |
+
prompt += response2
|
95 |
+
prompt += '\nOutput:'
|
96 |
+
|
97 |
+
# https://github.com/openai/openai-python/issues/322#issuecomment-1767841683
|
98 |
+
while True:
|
99 |
+
try:
|
100 |
+
response = openai.ChatCompletion.create(
|
101 |
+
model=gpt_model,
|
102 |
+
messages=[{"role": "user", "content": prompt}],
|
103 |
+
api_key=api_key,
|
104 |
+
request_timeout=5)
|
105 |
+
|
106 |
+
break
|
107 |
+
except:
|
108 |
+
print("Timeout, retrying...")
|
109 |
+
time.sleep(5) # Wait for 5 seconds before retrying
|
110 |
+
|
111 |
+
|
112 |
+
output_text = response['choices'][0]['message']['content']
|
113 |
+
|
114 |
+
gpt_same = "0"
|
115 |
+
|
116 |
+
if 'same' in output_text.lower():
|
117 |
+
gpt_same = "1"
|
118 |
+
|
119 |
+
elif 'different' in output_text.lower():
|
120 |
+
gpt_same = "0"
|
121 |
+
|
122 |
+
|
123 |
+
sample["same"] = gpt_same
|
124 |
+
|
125 |
+
with open(save_json_path, 'w') as f:
|
126 |
+
json.dump(data, f)
|
127 |
+
|
128 |
+
return data
|
129 |
+
|
130 |
+
def get_eval_fig(data): # per figure
|
131 |
+
|
132 |
+
eval_fig_dict = dict()
|
133 |
+
|
134 |
+
for r in data:
|
135 |
+
if r["category"] == "VS" and str(r["figure_id"]) == "0": # no figure
|
136 |
+
continue
|
137 |
+
name = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["figure_id"])])
|
138 |
+
if name in eval_fig_dict:
|
139 |
+
c, t = eval_fig_dict[name]
|
140 |
+
eval_fig_dict[name] = (c + r["correct"], t+1)
|
141 |
+
else:
|
142 |
+
eval_fig_dict[name] = (r["correct"], 1)
|
143 |
+
|
144 |
+
eval_fig_stat = {}
|
145 |
+
eval_fig_stat["note"] = "all accuracy per image (consistency test)"
|
146 |
+
eval_fig_stat["total"] = len(eval_fig_dict.keys())
|
147 |
+
eval_fig_stat["correct"] = 0
|
148 |
+
eval_fig_stat["wrong"] = 0
|
149 |
+
eval_fig_stat["inconsistent"] = 0
|
150 |
+
eval_fig_stat["score"] = 0
|
151 |
+
|
152 |
+
|
153 |
+
for v in eval_fig_dict.values():
|
154 |
+
if v[0] == v[1]:
|
155 |
+
eval_fig_stat["correct"] += 1
|
156 |
+
elif v[0] == 0:
|
157 |
+
eval_fig_stat["wrong"] += 1
|
158 |
+
else:
|
159 |
+
eval_fig_stat["inconsistent"] += 1
|
160 |
+
eval_fig_stat["score"] += (v[0] / v[1])
|
161 |
+
|
162 |
+
eval_fig_stat["score"] = eval_fig_stat["score"] / eval_fig_stat["total"]
|
163 |
+
return eval_fig_stat
|
164 |
+
|
165 |
+
def get_eval_all(data, model_correctness_entry): # per question
|
166 |
+
|
167 |
+
eval_all_dict = dict()
|
168 |
+
eval_all_stat = {}
|
169 |
+
eval_all_stat["LH"] = 0
|
170 |
+
eval_all_stat["VI"] = 0
|
171 |
+
eval_all_stat["Mix"] = 0
|
172 |
+
|
173 |
+
for r in data:
|
174 |
+
name = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["figure_id"]), str(r["question_id"])])
|
175 |
+
assert name not in eval_all_dict
|
176 |
+
|
177 |
+
eval_all_dict[name] = r["correct"]
|
178 |
+
|
179 |
+
if str(r["category"]) == "VD": # VD
|
180 |
+
if str(r["figure_id"]) == "0":
|
181 |
+
if str(r[model_correctness_entry]) == "0" or str(r[model_correctness_entry]) == "2":
|
182 |
+
eval_all_stat["VI"] += 1
|
183 |
+
else:
|
184 |
+
if str(r[model_correctness_entry]) == "0":
|
185 |
+
eval_all_stat["Mix"] += 1
|
186 |
+
elif str(r[model_correctness_entry]) == "2":
|
187 |
+
eval_all_stat["VI"] += 1
|
188 |
+
else: # VS
|
189 |
+
if str(r["visual_input"]) == "0": # no visual
|
190 |
+
if str(r[model_correctness_entry]) == "0":
|
191 |
+
eval_all_stat["LH"] += 1
|
192 |
+
else: # original visual or modified visual (isual_input == 1 or 2)
|
193 |
+
if str(r[model_correctness_entry]) == "0":
|
194 |
+
eval_all_stat["Mix"] += 1
|
195 |
+
elif str(r[model_correctness_entry]) == "2":
|
196 |
+
eval_all_stat["VI"] += 1
|
197 |
+
|
198 |
+
eval_all_stat["note"] = "all accuracy per question"
|
199 |
+
eval_all_stat["total"] = len(eval_all_dict.keys())
|
200 |
+
eval_all_stat["correct"] = np.count_nonzero(list(eval_all_dict.values()))
|
201 |
+
eval_all_stat["wrong"] = eval_all_stat["total"] - eval_all_stat["correct"]
|
202 |
+
|
203 |
+
return eval_all_stat
|
204 |
+
|
205 |
+
def get_eval_pair_all(data, model_correctness_entry): # per question pair
|
206 |
+
|
207 |
+
orig_correctness = dict()
|
208 |
+
counter = 0
|
209 |
+
lh_counter = 0
|
210 |
+
vi_counter = 0
|
211 |
+
both_counter = 0
|
212 |
+
|
213 |
+
for r in data:
|
214 |
+
if str(r["figure_id"]) == "0":
|
215 |
+
key = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
216 |
+
orig_correctness[key] = r[model_correctness_entry]
|
217 |
+
|
218 |
+
get_eval_pair_dict = dict()
|
219 |
+
get_analysis_pair_dict = dict()
|
220 |
+
|
221 |
+
for r in data:
|
222 |
+
name = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
223 |
+
if name in get_eval_pair_dict:
|
224 |
+
c, t = get_eval_pair_dict[name]
|
225 |
+
get_eval_pair_dict[name] = (c + r["correct"], t+1)
|
226 |
+
else:
|
227 |
+
get_eval_pair_dict[name] = (r["correct"], 1)
|
228 |
+
counter += 1
|
229 |
+
|
230 |
+
# (LH, VI)
|
231 |
+
analysis = (0, 0)
|
232 |
+
if str(r["figure_id"]) == "0": # when it's original question
|
233 |
+
if str(r["category"]) == "VD": # VD
|
234 |
+
if str(r[model_correctness_entry]) == "0" or str(r[model_correctness_entry]) == "2":
|
235 |
+
analysis = (0, 1) # VI -- get original image wrong, bad vision
|
236 |
+
else: # VS
|
237 |
+
if str(r[model_correctness_entry]) == "0":
|
238 |
+
analysis = (1, 0) # LH -- wrong answer without visual, making things up
|
239 |
+
else: # when it's not original question
|
240 |
+
key = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
241 |
+
orig_c = orig_correctness[key]
|
242 |
+
if str(r["category"]) == "VD": # VD
|
243 |
+
if str(orig_c) == "1" and str(r[model_correctness_entry]) == "0":
|
244 |
+
if str(r["same"]) == "1":
|
245 |
+
analysis = (1, 1) # Mixed -- orig correct but modified wrong, with the same answer as the original question, could be bad vision or language hallucination
|
246 |
+
else:
|
247 |
+
analysis = (0, 1) # VI -- orig correct but modified wrong, but answer differently, only due to bad vision
|
248 |
+
elif str(orig_c) == "1" and str(r[model_correctness_entry]) == "2":
|
249 |
+
analysis = (0, 1) # VI -- orig correct but modified uncertain, bad vision
|
250 |
+
elif str(r[model_correctness_entry]) == "0" or str(r[model_correctness_entry]) == "2":
|
251 |
+
# when orig_c == 0 or 2 and current is wrong
|
252 |
+
analysis = (0, 1) # VI -- when original is wrong and current is wrong, bad vision
|
253 |
+
else: # VS
|
254 |
+
key = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
255 |
+
orig_c = orig_correctness[key]
|
256 |
+
if str(orig_c) == "0": # No visual wrong
|
257 |
+
if str(r[model_correctness_entry]) == "0" and str(r["same"]) == "1":
|
258 |
+
analysis = (1, 0) # LH -- same answer with and without visual, LH overtake visual
|
259 |
+
elif str(r[model_correctness_entry]) == "0":
|
260 |
+
analysis = (1, 1) # LH -- different answer with and without visual but both wrong, both language and visual are bad
|
261 |
+
elif str(r[model_correctness_entry]) == "2":
|
262 |
+
analysis = (1, 1) # Mixed -- no visual wrong, but with visual uncertain, could be either
|
263 |
+
elif str(orig_c) == "2":# No visual uncertain
|
264 |
+
if str(r[model_correctness_entry]) == "0" or str(r[model_correctness_entry]) == "2":
|
265 |
+
analysis = (0, 1) # VI -- no visual uncertain, with visual still wrong or uncertain, visual capability is bad
|
266 |
+
else: # No visual correct
|
267 |
+
if str(r[model_correctness_entry]) == "2":
|
268 |
+
analysis = (0, 1) # VI -- no visual correct, with visual uncertain, visual capability is bad
|
269 |
+
elif str(r[model_correctness_entry]) == "0": # current is wrong
|
270 |
+
if str(r["visual_input"]) == "1": # common sense visual question
|
271 |
+
analysis = (0, 1) # VI -- no visual correct, with visual wrong on common sense question, visual capability is bad
|
272 |
+
elif str(r["visual_input"]) == "2": # counter-common sense visual question
|
273 |
+
if str(r["same"]) == "1":
|
274 |
+
analysis = (1, 0) # LH -- with visual correct, but modified question wrong with the same answer, not considering visual so the error is attributed to Language
|
275 |
+
else:
|
276 |
+
analysis = (0, 1) # VI -- with visual correct, but modified question wrong with different answers, visual capability is bad
|
277 |
+
else:
|
278 |
+
assert False, "Data error"
|
279 |
+
|
280 |
+
if analysis[0] > 0 and analysis[1] > 0:
|
281 |
+
both_counter += 1
|
282 |
+
elif analysis[0] > 0:
|
283 |
+
lh_counter += 1
|
284 |
+
elif analysis[1] > 0:
|
285 |
+
vi_counter += 1
|
286 |
+
|
287 |
+
|
288 |
+
if name in get_analysis_pair_dict:
|
289 |
+
lh, vi = get_analysis_pair_dict[name]
|
290 |
+
get_analysis_pair_dict[name] = (lh + analysis[0], vi + analysis[1])
|
291 |
+
else:
|
292 |
+
get_analysis_pair_dict[name] = analysis
|
293 |
+
|
294 |
+
eval_all_pair_stat = {}
|
295 |
+
eval_all_pair_stat["note"] = "all accuracy per question pair"
|
296 |
+
eval_all_pair_stat["total"] = len(get_eval_pair_dict.keys())
|
297 |
+
eval_all_pair_stat["total_q"] = counter
|
298 |
+
eval_all_pair_stat["correct"] = 0
|
299 |
+
eval_all_pair_stat["wrong"] = 0
|
300 |
+
eval_all_pair_stat["LH"] = 0
|
301 |
+
eval_all_pair_stat["VI"] = 0
|
302 |
+
eval_all_pair_stat["Mix"] = 0
|
303 |
+
|
304 |
+
eval_all_pair_stat["LH_cg"] = lh_counter
|
305 |
+
eval_all_pair_stat["VI_cg"] = vi_counter
|
306 |
+
eval_all_pair_stat["Mix_cg"] = both_counter
|
307 |
+
|
308 |
+
# for v in get_eval_pair_dict.values():
|
309 |
+
# if v[0] == v[1]:
|
310 |
+
# eval_all_pair_stat["correct"] += 1
|
311 |
+
# else:
|
312 |
+
# eval_all_pair_stat["wrong"] += 1
|
313 |
+
|
314 |
+
# for v in get_analysis_pair_dict.values():
|
315 |
+
# if v[0] > 0 and v[1] > 0:
|
316 |
+
# eval_all_pair_stat["Mix"] += 1
|
317 |
+
# elif v[0] > 0:
|
318 |
+
# eval_all_pair_stat["LH"] += 1
|
319 |
+
# elif v[1] > 0:
|
320 |
+
# eval_all_pair_stat["VI"] += 1
|
321 |
+
|
322 |
+
for k in get_eval_pair_dict.keys():
|
323 |
+
v = get_eval_pair_dict[k]
|
324 |
+
a = get_analysis_pair_dict[k]
|
325 |
+
if v[0] == v[1]:
|
326 |
+
eval_all_pair_stat["correct"] += 1
|
327 |
+
else:
|
328 |
+
eval_all_pair_stat["wrong"] += 1
|
329 |
+
if a[0] > 0 and a[1] > 0:
|
330 |
+
eval_all_pair_stat["Mix"] += 1
|
331 |
+
elif a[0] > 0:
|
332 |
+
eval_all_pair_stat["LH"] += 1
|
333 |
+
elif a[1] > 0:
|
334 |
+
eval_all_pair_stat["VI"] += 1
|
335 |
+
|
336 |
+
assert (eval_all_pair_stat["wrong"] == (eval_all_pair_stat["Mix"] + eval_all_pair_stat["LH"] + eval_all_pair_stat["VI"]))
|
337 |
+
|
338 |
+
return eval_all_pair_stat
|
339 |
+
|
340 |
+
def get_eval_pair_easy(data):
|
341 |
+
|
342 |
+
get_eval_pair_dict = dict()
|
343 |
+
counter = 0
|
344 |
+
|
345 |
+
for r in data:
|
346 |
+
if str(r["visual_input"]) == "2":
|
347 |
+
# if str(r["figure_id"]) != "0":
|
348 |
+
continue
|
349 |
+
name = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
350 |
+
if name in get_eval_pair_dict:
|
351 |
+
c, t = get_eval_pair_dict[name]
|
352 |
+
get_eval_pair_dict[name] = (c + r["correct"], t+1)
|
353 |
+
else:
|
354 |
+
get_eval_pair_dict[name] = (r["correct"], 1)
|
355 |
+
counter += 1
|
356 |
+
|
357 |
+
eval_all_pair_stat = {}
|
358 |
+
eval_all_pair_stat["note"] = "all accuracy per question pair"
|
359 |
+
eval_all_pair_stat["total"] = len(get_eval_pair_dict.values())
|
360 |
+
eval_all_pair_stat["total_q"] = counter
|
361 |
+
eval_all_pair_stat["correct"] = 0
|
362 |
+
eval_all_pair_stat["wrong"] = 0
|
363 |
+
|
364 |
+
for v in get_eval_pair_dict.values():
|
365 |
+
if v[0] == v[1]:
|
366 |
+
eval_all_pair_stat["correct"] += 1
|
367 |
+
else:
|
368 |
+
eval_all_pair_stat["wrong"] += 1
|
369 |
+
|
370 |
+
return eval_all_pair_stat
|
371 |
+
|
372 |
+
def get_eval_pair_hard(data):
|
373 |
+
|
374 |
+
get_eval_pair_dict = dict()
|
375 |
+
counter = 0
|
376 |
+
|
377 |
+
for r in data:
|
378 |
+
if str(r["visual_input"]) != "2":
|
379 |
+
# if str(r["figure_id"]) == "0":
|
380 |
+
continue
|
381 |
+
name = "_".join([r["category"], r["subcategory"], str(r["set_id"]), str(r["question_id"])])
|
382 |
+
if name in get_eval_pair_dict:
|
383 |
+
c, t = get_eval_pair_dict[name]
|
384 |
+
get_eval_pair_dict[name] = (c + r["correct"], t+1)
|
385 |
+
else:
|
386 |
+
get_eval_pair_dict[name] = (r["correct"], 1)
|
387 |
+
counter += 1
|
388 |
+
|
389 |
+
eval_all_pair_stat = {}
|
390 |
+
eval_all_pair_stat["note"] = "all accuracy per question pair"
|
391 |
+
eval_all_pair_stat["total"] = len(get_eval_pair_dict.values())
|
392 |
+
eval_all_pair_stat["total_q"] = counter
|
393 |
+
eval_all_pair_stat["correct"] = 0
|
394 |
+
eval_all_pair_stat["wrong"] = 0
|
395 |
+
|
396 |
+
for v in get_eval_pair_dict.values():
|
397 |
+
if v[0] == v[1]:
|
398 |
+
eval_all_pair_stat["correct"] += 1
|
399 |
+
else:
|
400 |
+
eval_all_pair_stat["wrong"] += 1
|
401 |
+
|
402 |
+
return eval_all_pair_stat
|
403 |
+
|
404 |
+
def assign_correctness(data_arr, correctness_entry):
|
405 |
+
for r in data_arr:
|
406 |
+
assert int(r[correctness_entry]) == 0 or int(r[correctness_entry]) == 1 or int(r[correctness_entry]) == 2
|
407 |
+
if r["category"] == "VS" and int(r["figure_id"]) == 0: # if there is no visual supplement and the model does not know, count it as correct
|
408 |
+
r["correct"] = 1 if int(r[correctness_entry]) == 1 or int(r[correctness_entry]) == 2 else 0
|
409 |
+
else:
|
410 |
+
r["correct"] = 1 if int(r[correctness_entry]) == 1 else 0
|
411 |
+
return data_arr
|
412 |
+
|
413 |
+
|
414 |
+
def yes_ratio_stats(data):
|
415 |
+
|
416 |
+
|
417 |
+
yes_gt = [int(i["gt_answer"]) for i in data]
|
418 |
+
yes_pred = [int(int(i["correct"]) == int(i["gt_answer"])) for i in data]
|
419 |
+
|
420 |
+
fp_sample = [i for i in data if int(i["correct"]) == 0]
|
421 |
+
fp = [int(i["gt_answer"]) for i in fp_sample]
|
422 |
+
|
423 |
+
stats = {}
|
424 |
+
stats["diff"] = sum(yes_pred)/len(yes_pred) - sum(yes_gt)/len(yes_gt)
|
425 |
+
stats["fp"] = (len(fp) - sum(fp))/len(fp)
|
426 |
+
|
427 |
+
return stats
|
428 |
+
|