File size: 12,511 Bytes
dc05889 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 |
import json
from copy import deepcopy
import os
from pathlib import Path
import re
from PIL import Image
from multiprocessing import Pool
import ast
def merge_convs(conversations):
"""
Merge all successive 'human' conversations comprehensively.
Args:
conversations (list): List of conversation dictionaries
Returns:
list: Processed conversations with all successive human messages merged
Raises:
ValueError: If input is not a list or contains invalid conversation dictionaries
"""
# Validate input
if not isinstance(conversations, list):
raise ValueError("Input must be a list of conversation dictionaries")
# Validate each conversation dictionary structure
for conv in conversations:
if not isinstance(conv, dict):
raise ValueError("Each conversation must be a dictionary")
if 'from' not in conv or 'value' not in conv:
raise ValueError("Each conversation must have 'from' and 'value' keys")
processed_conversations = []
i = 0
while i < len(conversations):
current_conv = conversations[i]
# If current conversation is 'human', start merging
if current_conv['from'] == 'human':
# Collect all successive human conversations
merged_value = current_conv['value']
j = i + 1
while j < len(conversations) and conversations[j]['from'] == 'human':
merged_value += '\n\n' + conversations[j]['value']
j += 1
# Update current conversation with merged value
current_conv['value'] = merged_value
# Move index to last non-human conversation
i = j
else:
# For non-human conversations, just add to processed list
i += 1
processed_conversations.append(current_conv)
return processed_conversations
def transform_bbox(bbox, image_x, image_y):
# transform [y,x,height,width] to [x1,y1,x2,y2]
y, x, height, width = bbox
x1 = int(1000 * x / image_x)
y1 = int(1000 * y / image_y)
x2 = int(1000 * (x + width) / image_x)
y2 = int(1000 * (y + height) / image_y)
bbox_norm = [x1, y1, x2, y2]
return bbox_norm
grounding_step_prompt = "<|img|>Step {step_idx}. Given a GUI image, what are the relative (0-1000) pixel point coordinates for the element corresponding to the following instruction or description: {instruction}"
grounding_step_ans = "```\n{point_str}\n```"
act_step_prompt = "<|img|>Step {step_idx}. Instruction: {prev_instruction}"
act_step_ans = "The agent's action: {prev_action}"
user_start_prompt = "The agent is performing the ultimate task: {ultimate_task}."
user_history_instr_prompt = "History of the agent's steps:\n{history_list}."
def process_android_episodes(data, window_size=2):
"""
Process Android episodes and extract steps with click or long_press actions.
Args:
data (list): List of episode dictionaries
window_size (int, optional): Number of recent image-included conversations to include.
Defaults to 3 (current image + 2 previous image-included steps).
Returns:
dict: Dictionary with episode_id as key and list of filtered steps as value
"""
instructions = []
for episode in data:
for i, step in enumerate(episode):
res_touch_yx = ast.literal_eval(step["result_touch_yx"])
res_touch_yx = [round(res_touch_yx[0], 3), round(res_touch_yx[1], 3)]
res_lift_yx = ast.literal_eval(step["result_lift_yx"])
res_lift_yx = [round(res_lift_yx[0], 3), round(res_lift_yx[1], 3)]
is_tap = int(res_touch_yx[0]) != -1 and (res_touch_yx[0] == res_lift_yx[0] and res_touch_yx[1] == res_lift_yx[1])
step["is_tap"] = is_tap
if "coat_action_desc" not in step or step["coat_action_desc"] is None:
break
if not is_tap:
continue
if window_size > 0 and i == 0: # skip the first step if window_size > 0
continue
convs = [
{
"from": "human",
"value": user_start_prompt.format(
ultimate_task=step["instruction"]
),
},
]
cur_img_list = [Path("./") / Path(step["image_path"])]
if window_size > 0:
window_steps = episode[i-window_size:i] if i >= window_size else episode[:i]
if i > window_size: # has more history steps larger than window_size
convs.append(
{
"from": "human",
"value": user_history_instr_prompt.format(
history_list="\n".join(
[
f"\t{j+1}. " + prev_step["coat_action_desc"]
for j, prev_step in enumerate(episode[:i-window_size])
]
)
),
},
)
convs.append(
{
"from": "human",
"value": "The recent steps with the GUI images are as follows:\n",
}
)
for j, win_step_i in enumerate(window_steps):
if win_step_i["is_tap"]:
convs.append(
{
"from": "human",
"value": grounding_step_prompt.format(
instruction=win_step_i["coat_action_desc"], step_idx=i+1-(len(window_steps)-j)
),
}
)
convs.append(
{
"from": "gpt",
"value": grounding_step_ans.format(point_str=f"({win_step_i['coord_norm'][0]}, {win_step_i['coord_norm'][1]})"),
}
)
else:
convs.append(
{
"from": "human",
"value": act_step_prompt.format(
prev_instruction=win_step_i["coat_action_desc"], step_idx=i+1-(len(window_steps)-j)
),
}
)
if win_step_i["result_action_text"]:
convs.append(
{
"from": "human",
"value": act_step_ans.format(
prev_action=f"Type: {win_step_i['result_action_text']}"
),
}
)
else:
convs.append(
{
"from": "human",
"value": act_step_ans.format(
prev_action=win_step_i["coat_action_desc"]
),
}
)
win_img_list = [
str(Path("./") / Path(win_step["image_path"])) for win_step in window_steps
]
else:
convs.append(
{
"from": "human",
"value": user_history_instr_prompt.format(
history_list="\n".join(
[
f"\t{j+1}. " + prev_step["coat_action_desc"]
for j, prev_step in enumerate(episode[:i-window_size])
]
)
),
},
)
img_list = cur_img_list + win_img_list if window_size > 0 else cur_img_list
has_img_broken = False
for img_path in img_list:
try:
Image.open(str(img_path))
except Exception as e:
print(f"Error opening image {img_path}: {e}")
has_img_broken = True
break
if has_img_broken:
print(f"Image broken for episode {episode_id}, step {i+1}. Skipping...")
continue
# Current step details
convs.append(
{
"from": "human",
"value": grounding_step_prompt.format(instruction=step["coat_action_desc"], step_idx=i+1),
}
)
convs.append(
{
"from": "gpt",
"value": grounding_step_ans.format(point_str=f"({step['coord_norm'][0]}, {step['coord_norm'][1]})"),
}
)
convs = merge_convs(convs)
instructions.append(
{
"image": [str(img_path) for img_path in img_list],
"conversations": convs,
}
)
return instructions
# Example usage
if __name__ == "__main__":
# Sample data loading (replace with your actual file path)
data = []
episode_files = list(Path("./").rglob("*/*/*.json"))
for episode_file in episode_files:
with open(episode_file, "r", encoding="utf-8") as file:
episode_data = json.load(file)
data.append(episode_data)
img_parent_path = Path("./")
def preprocess_coord_norm(episode):
for step in episode:
if int(ast.literal_eval(step["result_touch_yx"])[0]) != -1:
if not Path(img_parent_path / step["image_path"]).exists():
continue
image_x, image_y = Image.open(img_parent_path / step["image_path"]).size
elem_bboxes = ast.literal_eval(step["ui_positions"]) # (y, x, height, width)
elem_bboxes = [transform_bbox(bbox, image_x, image_y) for bbox in elem_bboxes]
click_point_yx = ast.literal_eval(step["result_touch_yx"])
click_point = [1000*click_point_yx[1], 1000*click_point_yx[0]]
# check which element contains the click point
bbox = None
for elem_bbox in elem_bboxes:
if elem_bbox[0] <= click_point[0] <= elem_bbox[2] and elem_bbox[1] <= click_point[1] <= elem_bbox[3]:
if bbox is None:
bbox = elem_bbox
else:
# calculate area, take the smaller one
area = (bbox[2] - bbox[0]) * (bbox[3] - bbox[1])
new_area = (elem_bbox[2] - elem_bbox[0]) * (elem_bbox[3] - elem_bbox[1])
if new_area < area:
bbox = elem_bbox
if bbox is None:
coord_norm = [int(click_point[0]), int(click_point[1])]
else:
coord_norm = [(bbox[0] + bbox[2]) // 2, (bbox[1] + bbox[3]) // 2]
step["coord_norm"] = coord_norm
return episode
with Pool() as pool:
data = pool.map(preprocess_coord_norm, data)
# Process the episodes with default window_size=3
# window_size_list = [1, 2, 3]
window_size_list = [0, 1, 2, 3]
instructions = []
for window_size in window_size_list:
instructions.extend(process_android_episodes(data, window_size=window_size))
print(f"Number of context aware train instructions: {len(instructions)}")
with open(f"aitz_train_window_{'-'.join([str(e) for e in window_size_list])}_{len(instructions)//1000}k.json", "w", encoding="utf-8") as file:
json.dump(instructions, file, ensure_ascii=False, indent=4) |