Yash-Butala
commited on
Commit
•
18c2115
1
Parent(s):
b2693bc
Upload eval.py
Browse files
eval.py
ADDED
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
from math import sqrt
|
3 |
+
import re
|
4 |
+
from nltk.translate.bleu_score import sentence_bleu
|
5 |
+
|
6 |
+
# gold label file
|
7 |
+
gold_fn = 'test.json'
|
8 |
+
|
9 |
+
pred_fn = 'llava-v1.5-13b.json'
|
10 |
+
gold = json.load(open(gold_fn))
|
11 |
+
pred = json.load(open(pred_fn))
|
12 |
+
|
13 |
+
sequence_match = 0
|
14 |
+
action_score = 0
|
15 |
+
total_click_penalty = 0
|
16 |
+
total_press_penalty = 0
|
17 |
+
total_write_penalty = 0
|
18 |
+
ideal_score = 0
|
19 |
+
max_click_penalty = 0
|
20 |
+
max_press_penalty = 0
|
21 |
+
max_write_penalty = 0
|
22 |
+
|
23 |
+
|
24 |
+
|
25 |
+
def get_bounds(box: dict(), cx, cy):
|
26 |
+
for i in box:
|
27 |
+
tl = box[i]["top_left"]
|
28 |
+
br = box[i]["bottom_right"]
|
29 |
+
if (tl[0]+br[0])/2 == cx and (tl[1]+br[1])/2 == cy:
|
30 |
+
return (tl,br)
|
31 |
+
|
32 |
+
assert False
|
33 |
+
|
34 |
+
|
35 |
+
def dynamic_dirichlet_l2_penalty(tl, br, px, py):
|
36 |
+
|
37 |
+
len_x = br[0] - tl[0]
|
38 |
+
len_y = br[1] - tl[1]
|
39 |
+
|
40 |
+
cx = ( br[0] - tl[0] ) / 2
|
41 |
+
cy = ( br[1] - tl[1] ) / 2
|
42 |
+
|
43 |
+
dx = abs(cx - px) - (len_x * 0.5)
|
44 |
+
dy = abs(cy - py) - (len_y * 0.5)
|
45 |
+
dist = sqrt((dx * (dx > 0)) ** 2 + (dy * (dy > 0)) ** 2)
|
46 |
+
|
47 |
+
mu = sqrt( len_x ** 2 + len_y ** 2)
|
48 |
+
|
49 |
+
score = mu / (dist+mu)
|
50 |
+
penalty = 1 - score
|
51 |
+
return penalty
|
52 |
+
|
53 |
+
for idx in gold:
|
54 |
+
|
55 |
+
gold_script = open(gold[idx]['task']).read().strip().split('\n')[2:]
|
56 |
+
llm_script = pred[idx].strip().split()
|
57 |
+
llm_script = [x for x in llm_script if x.strip().startswith('pyautogui')]
|
58 |
+
#find extreme case values
|
59 |
+
sample_weight = (len(gold_script)-0.9)
|
60 |
+
|
61 |
+
ideal_score += sample_weight
|
62 |
+
for gold_line in gold_script:
|
63 |
+
action_type = gold_line.split("pyautogui.")[1].split("(")[0]
|
64 |
+
if action_type == 'click' or action_type == 'rightClick' or action_type == 'moveTo' or action_type == 'dragTo':
|
65 |
+
max_click_penalty += sample_weight/len(gold_script)
|
66 |
+
if action_type == 'press' or action_type == 'hotkey':
|
67 |
+
max_press_penalty += sample_weight/len(gold_script)
|
68 |
+
if action_type == 'write':
|
69 |
+
max_write_penalty += sample_weight/len(gold_script)
|
70 |
+
|
71 |
+
seq_match_flag = 1
|
72 |
+
click_penalty = 0
|
73 |
+
press_penalty = 0
|
74 |
+
write_penalty = 0
|
75 |
+
|
76 |
+
# if length doesn't seq match is 0
|
77 |
+
# llm_script = llm_script[:len(gold_script)]
|
78 |
+
if len(llm_script) != len(gold_script):
|
79 |
+
seq_match_flag = 0
|
80 |
+
if seq_match_flag == 1:
|
81 |
+
for i in range(len(gold_script)):
|
82 |
+
gold_line = gold_script[i].strip()
|
83 |
+
gold_action = gold_line.split('pyautogui.')[1].split('(')[0]
|
84 |
+
pred_line = llm_script[i]
|
85 |
+
if pred_line.startswith('pyautogui.') == False:
|
86 |
+
seq_match_flag = 0
|
87 |
+
break
|
88 |
+
pred_action = pred_line.split('pyautogui.')[1].split('(')[0]
|
89 |
+
if pred_action != gold_action:
|
90 |
+
seq_match_flag = 0
|
91 |
+
break
|
92 |
+
|
93 |
+
# find penalties for correct and wrong sequences
|
94 |
+
box_path = gold[idx]['box']
|
95 |
+
box_num = box_path.split("_")[-1].split(".json")[0]
|
96 |
+
box_path = "_".join(box_path.split("_")[:-1])+box_num+"_boxes.json"
|
97 |
+
box = json.load(open(box_path))
|
98 |
+
|
99 |
+
for i in range(len(gold_script)):
|
100 |
+
gold_line = gold_script[i].strip()
|
101 |
+
gold_action = gold_line.split('pyautogui.')[1].split('(')[0]
|
102 |
+
# just add the penalties
|
103 |
+
if seq_match_flag == 0:
|
104 |
+
if gold_action == 'click' or gold_action == 'rightClick' or gold_action == 'moveTo' or gold_action == 'dragTo':
|
105 |
+
click_penalty += 1/len(gold_script)
|
106 |
+
if gold_action == 'press' or gold_action == 'hotkey':
|
107 |
+
press_penalty += 1/len(gold_script)
|
108 |
+
if gold_action == 'write':
|
109 |
+
write_penalty += 1/len(gold_script)
|
110 |
+
continue
|
111 |
+
pred_line = llm_script[i]
|
112 |
+
pred_action = pred_line.split('pyautogui.')[1].split('(')[0]
|
113 |
+
|
114 |
+
# l2 penalty for click
|
115 |
+
|
116 |
+
if gold_action == 'click' or gold == 'rightClick':
|
117 |
+
# get original box bounds
|
118 |
+
gold_cx = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[0]
|
119 |
+
gold_cy = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[1].split(')')[0]
|
120 |
+
tl, br = get_bounds(box, float(gold_cx), float(gold_cy))
|
121 |
+
|
122 |
+
# get predicted point
|
123 |
+
pred_cx = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[0]
|
124 |
+
pred_cy = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[1].split(')')[0]
|
125 |
+
|
126 |
+
click_penalty += (1.0/len(gold_script)) * dynamic_dirichlet_l2_penalty(tl, br, float(pred_cx), float(pred_cy))
|
127 |
+
|
128 |
+
# penalty for press
|
129 |
+
if gold_action == 'press':
|
130 |
+
gold_key = gold_line.split("\"")[1]
|
131 |
+
pred_key = (re.split("\"|'", pred_line))[1]
|
132 |
+
if gold_key.strip() != pred_key.strip():
|
133 |
+
press_penalty += 1/len(gold_script)
|
134 |
+
|
135 |
+
# penalty for hotkey
|
136 |
+
if gold_action == 'hotkey':
|
137 |
+
gold_keys = gold_line.split("(")[1].split(")")[0].split(",")
|
138 |
+
pred_keys = pred_line.split("(")[1].split(")")[0].split(",")
|
139 |
+
|
140 |
+
gold_key_set = set([x[1:-1] for x in gold_keys if len(x)>2])
|
141 |
+
pred_key_set = set([x[1:-1] for x in pred_keys if len(x)>2])
|
142 |
+
if gold_key_set != pred_key_set:
|
143 |
+
press_penalty += 1/len(gold_script)
|
144 |
+
|
145 |
+
|
146 |
+
if gold_action == 'write':
|
147 |
+
reference = [gold_line.split("\"")[1]]
|
148 |
+
candidate = re.split("\"|'", pred_line)[1]
|
149 |
+
write_penalty += (1-sentence_bleu(reference, candidate, weights=(0.5, 0.5))) / len(gold_script)
|
150 |
+
|
151 |
+
sequence_match += (seq_match_flag) * sample_weight
|
152 |
+
action_score += (max(seq_match_flag - click_penalty - press_penalty - write_penalty, 0)) * sample_weight
|
153 |
+
if seq_match_flag:
|
154 |
+
total_click_penalty += click_penalty * sample_weight
|
155 |
+
total_press_penalty += press_penalty * sample_weight
|
156 |
+
total_write_penalty += write_penalty * sample_weight
|
157 |
+
|
158 |
+
|
159 |
+
print(ideal_score)
|
160 |
+
print(f"Sequence match: {sequence_match/ideal_score}")
|
161 |
+
print(f"Action match: {action_score/ideal_score}")
|
162 |
+
|
163 |
+
|
164 |
+
print(total_click_penalty/ideal_score)
|
165 |
+
print(total_press_penalty/ideal_score)
|
166 |
+
print(total_write_penalty/ideal_score)
|
167 |
+
|