File size: 6,124 Bytes
5e4bcb2 |
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 |
import json
from math import sqrt
import re
from nltk.translate.bleu_score import sentence_bleu
# gold label file
gold_fn = 'test.json'
pred_fn = 'llava-v1.5-13b.json'
gold = json.load(open(gold_fn))
pred = json.load(open(pred_fn))
sequence_match = 0
action_score = 0
total_click_penalty = 0
total_press_penalty = 0
total_write_penalty = 0
ideal_score = 0
max_click_penalty = 0
max_press_penalty = 0
max_write_penalty = 0
def get_bounds(box: dict(), cx, cy):
for i in box:
tl = box[i]["top_left"]
br = box[i]["bottom_right"]
if (tl[0]+br[0])/2 == cx and (tl[1]+br[1])/2 == cy:
return (tl,br)
assert False
def dynamic_dirichlet_l2_penalty(tl, br, px, py):
len_x = br[0] - tl[0]
len_y = br[1] - tl[1]
cx = ( br[0] - tl[0] ) / 2
cy = ( br[1] - tl[1] ) / 2
dx = abs(cx - px) - (len_x * 0.5)
dy = abs(cy - py) - (len_y * 0.5)
dist = sqrt((dx * (dx > 0)) ** 2 + (dy * (dy > 0)) ** 2)
mu = sqrt( len_x ** 2 + len_y ** 2)
score = mu / (dist+mu)
penalty = 1 - score
return penalty
for idx in gold:
gold_script = open(gold[idx]['task']).read().strip().split('\n')[2:]
llm_script = pred[idx].strip().split()
llm_script = [x for x in llm_script if x.strip().startswith('pyautogui')]
#find extreme case values
sample_weight = (len(gold_script)-0.9)
ideal_score += sample_weight
for gold_line in gold_script:
action_type = gold_line.split("pyautogui.")[1].split("(")[0]
if action_type == 'click' or action_type == 'rightClick' or action_type == 'moveTo' or action_type == 'dragTo':
max_click_penalty += sample_weight/len(gold_script)
if action_type == 'press' or action_type == 'hotkey':
max_press_penalty += sample_weight/len(gold_script)
if action_type == 'write':
max_write_penalty += sample_weight/len(gold_script)
seq_match_flag = 1
click_penalty = 0
press_penalty = 0
write_penalty = 0
# if length doesn't seq match is 0
# llm_script = llm_script[:len(gold_script)]
if len(llm_script) != len(gold_script):
seq_match_flag = 0
if seq_match_flag == 1:
for i in range(len(gold_script)):
gold_line = gold_script[i].strip()
gold_action = gold_line.split('pyautogui.')[1].split('(')[0]
pred_line = llm_script[i]
if pred_line.startswith('pyautogui.') == False:
seq_match_flag = 0
break
pred_action = pred_line.split('pyautogui.')[1].split('(')[0]
if pred_action != gold_action:
seq_match_flag = 0
break
# find penalties for correct and wrong sequences
box_path = gold[idx]['box']
box_num = box_path.split("_")[-1].split(".json")[0]
box_path = "_".join(box_path.split("_")[:-1])+box_num+"_boxes.json"
box = json.load(open(box_path))
for i in range(len(gold_script)):
gold_line = gold_script[i].strip()
gold_action = gold_line.split('pyautogui.')[1].split('(')[0]
# just add the penalties
if seq_match_flag == 0:
if gold_action == 'click' or gold_action == 'rightClick' or gold_action == 'moveTo' or gold_action == 'dragTo':
click_penalty += 1/len(gold_script)
if gold_action == 'press' or gold_action == 'hotkey':
press_penalty += 1/len(gold_script)
if gold_action == 'write':
write_penalty += 1/len(gold_script)
continue
pred_line = llm_script[i]
pred_action = pred_line.split('pyautogui.')[1].split('(')[0]
# l2 penalty for click
if gold_action == 'click' or gold == 'rightClick':
# get original box bounds
gold_cx = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[0]
gold_cy = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[1].split(')')[0]
tl, br = get_bounds(box, float(gold_cx), float(gold_cy))
# get predicted point
pred_cx = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[0]
pred_cy = gold_line.split("pyautogui.")[1].split('(')[1].split(',')[1].split(')')[0]
click_penalty += (1.0/len(gold_script)) * dynamic_dirichlet_l2_penalty(tl, br, float(pred_cx), float(pred_cy))
# penalty for press
if gold_action == 'press':
gold_key = gold_line.split("\"")[1]
pred_key = (re.split("\"|'", pred_line))[1]
if gold_key.strip() != pred_key.strip():
press_penalty += 1/len(gold_script)
# penalty for hotkey
if gold_action == 'hotkey':
gold_keys = gold_line.split("(")[1].split(")")[0].split(",")
pred_keys = pred_line.split("(")[1].split(")")[0].split(",")
gold_key_set = set([x[1:-1] for x in gold_keys if len(x)>2])
pred_key_set = set([x[1:-1] for x in pred_keys if len(x)>2])
if gold_key_set != pred_key_set:
press_penalty += 1/len(gold_script)
if gold_action == 'write':
reference = [gold_line.split("\"")[1]]
candidate = re.split("\"|'", pred_line)[1]
write_penalty += (1-sentence_bleu(reference, candidate, weights=(0.5, 0.5))) / len(gold_script)
sequence_match += (seq_match_flag) * sample_weight
action_score += (max(seq_match_flag - click_penalty - press_penalty - write_penalty, 0)) * sample_weight
if seq_match_flag:
total_click_penalty += click_penalty * sample_weight
total_press_penalty += press_penalty * sample_weight
total_write_penalty += write_penalty * sample_weight
print(ideal_score)
print(f"Sequence match: {sequence_match/ideal_score}")
print(f"Action match: {action_score/ideal_score}")
print(total_click_penalty/ideal_score)
print(total_press_penalty/ideal_score)
print(total_write_penalty/ideal_score)
|