File size: 34,702 Bytes
7c08dc3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 |
from dotenv import load_dotenv
import os
import json
import copy
import yaml
import logging
import time
from jinja2 import Environment, StrictUndefined
from utils.src.utils import ppt_to_images, get_json_from_response
from camel.models import ModelFactory
from camel.agents import ChatAgent
from camel.messages import BaseMessage
from utils.pptx_utils import *
from utils.wei_utils import *
import pickle as pkl
import argparse
from concurrent.futures import ThreadPoolExecutor
from concurrent.futures import ProcessPoolExecutor, as_completed
import concurrent.futures
import sys
load_dotenv()
logging.basicConfig(
level=logging.DEBUG,
format='%(threadName)s: %(message)s',
stream=sys.stdout
)
logger = logging.getLogger(__name__)
IMAGE_SCALE_RATIO_MIN = 50
IMAGE_SCALE_RATIO_MAX = 40
TABLE_SCALE_RATIO_MIN = 100
TABLE_SCALE_RATIO_MAX = 80
def layout_process_section_wrapped(
sections,
new_outline,
init_template,
new_section_template,
init_actor_sys_msg,
new_section_actor_sys_msg,
actor_config,
documentation,
max_retry,
slide_width,
slide_height
):
logs = {}
parallel_results = {}
total_input_token, total_output_token = 0, 0
# Switch from ThreadPoolExecutor to ProcessPoolExecutor
with ThreadPoolExecutor() as executor:
futures = []
for section_index in range(len(sections)):
if section_index == 0:
sys_msg = init_actor_sys_msg
prompt_template = init_template
else:
sys_msg = new_section_actor_sys_msg
prompt_template = new_section_template
actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
future = executor.submit(
layout_process_section,
section_index,
sections,
new_outline,
prompt_template,
documentation,
sys_msg,
actor_model,
10,
max_retry,
slide_width,
slide_height
)
futures.append(future)
# Collect results as processes complete
for future in as_completed(futures):
section_index, section_logs, in_toks, out_toks = future.result()
# Store logs by section index
parallel_results[section_index] = section_logs
# Update token counters
total_input_token += in_toks
total_output_token += out_toks
# Merge results back into `logs`
for section_index, section_logs in parallel_results.items():
curr_section = sections[section_index]
logs[curr_section] = section_logs
return logs, total_input_token, total_output_token
def create_agent_fn(sys_msg, agent_model, window_size=10):
agent = ChatAgent(
system_message=sys_msg,
model=agent_model,
message_window_size=window_size,
)
return agent
def layout_h2_process_section(
section,
outline_no_sub_locations,
h2_actor_template,
create_h2_actor_agent, # If you need a fresh agent for each thread
):
"""
Run the logic for a single section.
Returns a tuple containing:
- section name (or id),
- updated subsection-location dict,
- input token count,
- output token count
"""
print(f'Generating h2 for section {section}...', flush=True)
# 1) Create the prompt
section_outline = {section: outline_no_sub_locations[section]}
section_jinja_args = {
'section_outline': json.dumps(section_outline, indent=4),
}
section_prompt = h2_actor_template.render(**section_jinja_args)
# 2) Prepare a fresh agent or reuse existing (thread-safe?) agent
# If your h2_actor_agent is not thread-safe, instantiate a new one here:
h2_actor_agent = create_h2_actor_agent()
h2_actor_agent.reset()
# 3) Get response
response = h2_actor_agent.step(section_prompt)
input_token, output_token = account_token(response)
# 4) Parse JSON
subsection_location = get_json_from_response(response.msgs[0].content)
# 5) Create a dict from the sub-locations
sec_bbox = outline_no_sub_locations[section]['location']
subsection_location_dict = {}
for k, v in subsection_location.items():
subsection_location_dict[k] = {
'left': v['location'][0],
'top': v['location'][1],
'width': v['location'][2],
'height': v['location'][3]
}
# 6) Validate and possibly revise
is_valid, revised = validate_and_adjust_subsections(sec_bbox, subsection_location_dict)
if not is_valid:
# Try once more
is_valid, revised = validate_and_adjust_subsections(sec_bbox, revised)
assert is_valid, "Failed to adjust subsections to fit section"
final_sub_loc = revised
else:
final_sub_loc = subsection_location
# Return all data needed by the main thread
return section, final_sub_loc, input_token, output_token
def layout_process_section(
section_index,
sections,
new_outline,
new_section_template,
documentation,
sys_msg,
agent_model,
window_size,
max_retry,
slide_width,
slide_height
):
"""
Runs the 'gen_layout' logic for a single section_index.
Returns a tuple:
(section_index, updated_log, input_tokens, output_tokens)
"""
curr_section = sections[section_index]
print(f'Generating h1 layout for section {curr_section}...')
# Build outline JSON just for current section
new_section_outline = {curr_section: new_outline[curr_section]}
if section_index == 0:
new_section_outline = {'meta': new_outline['meta'], curr_section: new_outline[curr_section]}
new_section_jinja_args = {
'json_outline': new_section_outline,
'function_docs': documentation,
'file_name': f'poster_{section_index}.pptx'
}
# Render prompt
new_section_prompt = new_section_template.render(**new_section_jinja_args)
existing_code = '' # Or fetch from a stable location that is not dependent on real-time results
# Call gen_layout
section_logs = gen_layout_parallel(
create_agent_fn(
sys_msg,
agent_model,
window_size
),
new_section_prompt,
max_retry,
existing_code=existing_code,
slide_width=slide_width,
slide_height=slide_height,
tmp_name=section_index
)
if section_logs[-1]['error'] is not None:
print(f'Failed to generate layout for section {curr_section}.')
return None
in_toks, out_toks = section_logs[-1]['cumulative_tokens']
return (section_index, section_logs, in_toks, out_toks)
def get_outline_location(outline, subsection=False):
outline_location = {}
for k, v in outline.items():
if k == 'meta':
continue
outline_location[k] = {
'location': v['location'],
}
if subsection:
if 'subsections' in v:
outline_location[k]['subsections'] = get_outline_location(v['subsections'])
return outline_location
def apply_outline_location(outline, location, subsection=False):
new_outline = {}
for k, v in outline.items():
if k == 'meta':
new_outline[k] = v
continue
new_outline[k] = copy.deepcopy(v)
new_outline[k]['location'] = location[k]['location']
if subsection:
if 'subsections' in v:
new_outline[k]['subsections'] = apply_outline_location(v['subsections'], location[k]['subsections'])
return new_outline
def fill_location(outline, section_name, location_dict):
new_outline = copy.deepcopy(outline)
if 'subsections' not in new_outline[section_name]:
return new_outline
for k, v in new_outline[section_name]['subsections'].items():
v['location'] = location_dict[k]['location']
return new_outline
def recover_name_and_location(outline_no_name, outline):
new_outline = copy.deepcopy(outline_no_name)
for k, v in outline_no_name.items():
if k == 'meta':
continue
new_outline[k]['name'] = outline[k]['name']
if type(new_outline[k]['location']) == list:
new_outline[k]['location'] = {
'left': v['location'][0],
'top': v['location'][1],
'width': v['location'][2],
'height': v['location'][3]
}
if 'subsections' in v:
for k_sub, v_sub in v['subsections'].items():
new_outline[k]['subsections'][k_sub]['name'] = outline[k]['subsections'][k_sub]['name']
if type(new_outline[k]['subsections'][k_sub]['location']) == list:
new_outline[k]['subsections'][k_sub]['location'] = {
'left': v_sub['location'][0],
'top': v_sub['location'][1],
'width': v_sub['location'][2],
'height': v_sub['location'][3]
}
return new_outline
def validate_and_adjust_subsections(section_bbox, subsection_bboxes):
"""
Validate that the given subsections collectively occupy the entire section.
If not, return an adjusted version that fixes the layout.
We assume all subsections are intended to be stacked vertically with no gaps,
spanning the full width of the section.
:param section_bbox: dict with keys ["left", "top", "width", "height"]
:param subsection_bboxes: dict of subsection_name -> bounding_box (each also
with keys ["left", "top", "width", "height"])
:return: (is_valid, revised_subsections)
where is_valid is True/False,
and revised_subsections is either the same as subsection_bboxes if valid,
or a new dict of adjusted bounding boxes if invalid.
"""
# Helper functions
def _right(bbox):
return bbox["left"] + bbox["width"]
def _bottom(bbox):
return bbox["top"] + bbox["height"]
section_left = section_bbox["left"]
section_top = section_bbox["top"]
section_right = section_left + section_bbox["width"]
section_bottom = section_top + section_bbox["height"]
# Convert dictionary to a list of (subsection_name, bbox) pairs
items = list(subsection_bboxes.items())
if not items:
# No subsections is definitely not valid if we want to fill the section
return False, None
# Sort subsections by their 'top' coordinate
items_sorted = sorted(items, key=lambda x: x[1]["top"])
# ---------------------------
# Step 1: Validate
# ---------------------------
# We'll check:
# 1. left/right boundaries match the section for each subsection
# 2. The first subsection's top == section_top
# 3. The last subsection's bottom == section_bottom
# 4. Each pair of consecutive subsections lines up exactly
# (previous bottom == current top) with no gap or overlap.
is_valid = True
# Check left/right for each
for name, bbox in items_sorted:
if bbox["left"] != section_left or _right(bbox) != section_right:
is_valid = False
break
# Check alignment for the first and last
if is_valid:
first_sub_name, first_sub_bbox = items_sorted[0]
if first_sub_bbox["top"] != section_top:
is_valid = False
if is_valid:
last_sub_name, last_sub_bbox = items_sorted[-1]
if _bottom(last_sub_bbox) != section_bottom:
is_valid = False
# Check consecutive alignment
if is_valid:
for i in range(len(items_sorted) - 1):
_, current_bbox = items_sorted[i]
_, next_bbox = items_sorted[i + 1]
if _bottom(current_bbox) != next_bbox["top"]:
is_valid = False
break
# If everything passed, we return
if is_valid:
return True, subsection_bboxes
# ---------------------------
# Step 2: Revise
# ---------------------------
# We will adjust all subsection bboxes so that they occupy
# the entire section exactly, preserving each original bbox's
# height *ratio* if possible.
# 2a. Compute total original height (in the order of sorted items)
original_heights = [bbox["height"] for _, bbox in items_sorted]
total_original_height = sum(original_heights)
# Avoid divide-by-zero if somehow there's a 0 height
if total_original_height <= 0:
# Fallback: split the section equally among subsections
# to avoid zero or negative heights
chunk_height = section_bbox["height"] / len(items_sorted)
scale_heights = [chunk_height] * len(items_sorted)
else:
# Scale each original height by the ratio of
# (section total height / sum of original heights)
scale = section_bbox["height"] / total_original_height
scale_heights = [h * scale for h in original_heights]
# 2b. Assign bounding boxes top->bottom, ensuring no gap
revised = {}
current_top = section_top
for i, (name, original_bbox) in enumerate(items_sorted):
revised_height = scale_heights[i]
# If there's floating error, we can clamp in the last iteration
# so that the bottom exactly matches section_bottom.
# But for simplicity, we'll keep it straightforward unless needed.
revised[name] = {
"left": section_left,
"top": current_top,
"width": section_bbox["width"],
"height": revised_height
}
# Update current_top for next subsection
current_top += revised_height
# Due to potential float rounding, we can enforce the last subsection
# to exactly end at section_bottom:
last_name = items_sorted[-1][0]
# Recompute the actual bottom after the above assignment
new_bottom = revised[last_name]["top"] + revised[last_name]["height"]
diff = new_bottom - section_bottom
if abs(diff) > 1e-9:
# Adjust the last subsection's height
revised[last_name]["height"] -= diff
# Return the revised dictionary
return False, revised
def filter_image_table(args, filter_config):
images = json.load(open(f'images_and_tables/{args.poster_name}_images.json', 'r'))
tables = json.load(open(f'images_and_tables/{args.poster_name}_tables.json', 'r'))
doc_json = json.load(open(f'contents/{args.model_name}_{args.poster_name}_raw_content.json', 'r'))
agent_filter = 'image_table_filter_agent'
with open(f"prompt_templates/{agent_filter}.yaml", "r") as f:
config_filter = yaml.safe_load(f)
image_information = {}
for k, v in images.items():
image_information[k] = copy.deepcopy(v)
image_information[k]['min_width'] = v['width'] // IMAGE_SCALE_RATIO_MIN
image_information[k]['min_height'] = v['height'] // IMAGE_SCALE_RATIO_MIN
image_information[k]['max_width'] = v['width'] // IMAGE_SCALE_RATIO_MAX
image_information[k]['max_height'] = v['height'] // IMAGE_SCALE_RATIO_MAX
table_information = {}
for k, v in tables.items():
table_information[k] = copy.deepcopy(v)
table_information[k]['min_width'] = v['width'] // TABLE_SCALE_RATIO_MIN
table_information[k]['min_height'] = v['height'] // TABLE_SCALE_RATIO_MIN
table_information[k]['max_width'] = v['width'] // TABLE_SCALE_RATIO_MAX
table_information[k]['max_height'] = v['height'] // TABLE_SCALE_RATIO_MAX
filter_actor_sys_msg = config_filter['system_prompt']
filter_model = ModelFactory.create(
model_platform=filter_config['model_platform'],
model_type=filter_config['model_type'],
model_config_dict=filter_config['model_config'],
)
filter_actor_agent = ChatAgent(
system_message=filter_actor_sys_msg,
model=filter_model,
message_window_size=10, # [Optional] the length for chat memory
)
filter_jinja_args = {
'json_content': doc_json,
'table_information': table_information,
'image_information': image_information,
}
jinja_env = Environment(undefined=StrictUndefined)
filter_prompt = jinja_env.from_string(config_filter["template"])
response = filter_actor_agent.step(filter_prompt.render(**filter_jinja_args))
input_token, output_token = account_token(response)
response_json = get_json_from_response(response.msgs[0].content)
table_information = response_json['table_information']
image_information = response_json['image_information']
json.dump(images, open(f'images_and_tables/{args.poster_name}_images_filtered.json', 'w'), indent=4)
json.dump(tables, open(f'images_and_tables/{args.poster_name}_tables_filtered.json', 'w'), indent=4)
return input_token, output_token
def gen_outline_layout(args, actor_config, critic_config):
poster_log_path = f'log/{args.model_name}_{args.poster_name}_poster_{args.index}'
if not os.path.exists(poster_log_path):
os.mkdir(poster_log_path)
total_input_token, total_output_token = 0, 0
consumption_log = {
'outline': [],
'h1_actor': [],
'h2_actor': [],
'h1_critic': [],
'gen_layout': []
}
jinja_env = Environment(undefined=StrictUndefined)
outline_file_path = f'outlines/{args.model_name}_{args.poster_name}_outline_{args.index}.json'
agent_name = 'poster_planner_new'
agent_init_name = 'layout_agent_init_parallel'
agent_new_section_name = 'layout_agent_new_section_parallel'
h1_critic_name = 'critic_layout_hierarchy_1'
h2_actor_name = 'actor_layout_hierarchy_2'
doc_json = json.load(open(f'contents/{args.model_name}_{args.poster_name}_raw_content.json', 'r'))
filtered_table_information = json.load(open(f'images_and_tables/{args.poster_name}_tables_filtered.json', 'r'))
filtered_image_information = json.load(open(f'images_and_tables/{args.poster_name}_images_filtered.json', 'r'))
with open(f"prompt_templates/{agent_name}.yaml", "r") as f:
planner_config = yaml.safe_load(f)
with open(f"prompt_templates/{agent_init_name}.yaml", "r") as f:
config_init = yaml.safe_load(f)
with open(f"prompt_templates/{agent_new_section_name}.yaml", "r") as f:
config_new_section = yaml.safe_load(f)
with open(f"prompt_templates/{h1_critic_name}.yaml", "r") as f:
config_h1_critic = yaml.safe_load(f)
with open(f"prompt_templates/{h2_actor_name}.yaml", "r") as f:
config_h2_actor = yaml.safe_load(f)
planner_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
planner_agent = ChatAgent(
system_message=planner_config['system_prompt'],
model=planner_model,
message_window_size=10,
)
outline_template = jinja_env.from_string(planner_config["template"])
planner_jinja_args = {
'json_content': doc_json,
'table_information': filtered_table_information,
'image_information': filtered_image_information,
}
actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
init_actor_sys_msg = config_init['system_prompt']
def create_init_actor_agent():
actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
init_actor_agent = ChatAgent(
system_message=init_actor_sys_msg,
model=actor_model,
message_window_size=10,
)
return init_actor_agent
new_section_actor_sys_msg = config_new_section['system_prompt']
def create_new_section_actor_agent():
actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
new_section_actor_agent = ChatAgent(
system_message=new_section_actor_sys_msg,
model=actor_model,
message_window_size=10,
)
return new_section_actor_agent
h1_critic_model = ModelFactory.create(
model_platform=critic_config['model_platform'],
model_type=critic_config['model_type'],
model_config_dict=critic_config['model_config'],
)
h1_critic_sys_msg = config_h1_critic['system_prompt']
h1_critic_agent = ChatAgent(
system_message=h1_critic_sys_msg,
model=h1_critic_model,
message_window_size=None,
)
h1_pos_example = Image.open('h1_example/h1_pos.jpg')
h1_neg_example = Image.open('h1_example/h1_neg.jpg')
h2_actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
h2_actor_sys_msg = config_h2_actor['system_prompt']
def create_h2_actor_agent():
h2_actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
h2_actor_agent = ChatAgent(
system_message=h2_actor_sys_msg,
model=h2_actor_model,
message_window_size=10,
)
return h2_actor_agent
init_template = jinja_env.from_string(config_init["template"])
new_section_template = jinja_env.from_string(config_new_section["template"])
h1_critic_template = jinja_env.from_string(config_h1_critic["template"])
attempt = 0
while True:
print(f'Generating outline attempt {attempt}...', flush=True)
planner_prompt = outline_template.render(**planner_jinja_args)
planner_agent.reset()
response = planner_agent.step(planner_prompt)
outline = get_json_from_response(response.msgs[0].content)
input_token, output_token = account_token(response)
sections = list(outline.keys())
sections = [x for x in sections if x != 'meta']
slide_width = outline['meta']['width']
slide_height = outline['meta']['height']
name_to_hierarchy = get_hierarchy(outline)
consumption_log['outline'].append((input_token, output_token))
total_input_token += input_token
total_output_token += output_token
init_outline = {'meta': outline['meta'], sections[0]: outline[sections[0]]}
new_outline = outline
init_jinja_args = {
'json_outline': init_outline,
'function_docs': documentation
}
init_prompt = init_template.render(**init_jinja_args)
# hierarchy 1 only
outline_location = get_outline_location(outline, subsection=False)
logs, layout_cumulative_input_token, layout_cumulative_output_token = layout_process_section_wrapped(
sections,
new_outline,
init_template,
new_section_template,
init_actor_sys_msg,
new_section_actor_sys_msg,
actor_config,
documentation,
args.max_retry,
slide_width,
slide_height
)
concatenated_code = utils_functions
for section_index in range(len(sections)):
section = sections[section_index]
concatenated_code += '\n' + logs[section][-1]['code']
presentation_object_name = logs[section][-1]['output'].replace('\n', '')
concatenated_code += '\n' + f'save_presentation({presentation_object_name}, file_name="poster_{section_index + 1}.pptx")'
concatenated_code += f'''
name_to_hierarchy = {name_to_hierarchy}
identifier = "parallel"
poster_path = "poster_{section_index + 1}.pptx"
get_visual_cues(name_to_hierarchy, identifier, poster_path)
'''
output, error = run_code_with_utils(concatenated_code, utils_functions)
if error is not None:
print(error, flush=True)
attempt += 1
continue
consumption_log['h1_actor'].append((layout_cumulative_input_token, layout_cumulative_output_token))
total_input_token += layout_cumulative_input_token
total_output_token += layout_cumulative_output_token
h1_path = f'tmp/poster_<parallel>_hierarchy_1.pptx'
h2_path = f'tmp/poster_<parallel>_hierarchy_2.pptx'
h1_filled_path = f'tmp/poster_<parallel>_hierarchy_1_filled.pptx'
h2_filled_path = f'tmp/poster_<parallel>_hierarchy_2_filled.pptx'
ppt_to_images(h1_path, 'tmp/layout_h1')
ppt_to_images(h2_path, 'tmp/layout_h2')
ppt_to_images(h1_filled_path, 'tmp/layout_h1_filled')
ppt_to_images(h2_filled_path, 'tmp/layout_h2_filled')
h1_img = Image.open('tmp/layout_h1/slide_0001.jpg')
h2_img = Image.open('tmp/layout_h2/slide_0001.jpg')
h1_filled_img = Image.open('tmp/layout_h1_filled/slide_0001.jpg')
h2_filled_img = Image.open('tmp/layout_h2_filled/slide_0001.jpg')
h1_critic_msg = BaseMessage.make_user_message(
role_name='User',
content=h1_critic_template.render(),
image_list=[h1_neg_example, h1_pos_example, h1_filled_img]
)
outline_bbox_dict = {}
for k, v in outline_location.items():
outline_bbox_dict[k] = v['location']
bbox_check_result = check_bounding_boxes(
outline_bbox_dict,
new_outline['meta']['width'],
new_outline['meta']['height']
)
if len(bbox_check_result) != 0:
print(bbox_check_result, flush=True)
attempt += 1
continue
h1_critic_agent.reset()
response = h1_critic_agent.step(h1_critic_msg)
input_token, output_token = account_token(response)
consumption_log['h1_critic'].append((input_token, output_token))
total_input_token += input_token
total_output_token += output_token
if response.msgs[0].content == 'T':
print('Blank area detected.', flush=True)
attempt += 1
continue
print('Sucessfully generated outline.', flush=True)
break
outline_bbox_dict = {}
for k, v in outline_location.items():
outline_bbox_dict[k] = v['location']
# Generate subsection locations
outline_no_sub_locations = copy.deepcopy(new_outline)
if 'meta' in outline_no_sub_locations:
outline_no_sub_locations.pop('meta')
for k, v in outline_no_sub_locations.items():
if 'subsections' in v:
subsections = v['subsections']
for k_sub, v_sub in subsections.items():
del v_sub['location']
del v_sub['name']
h2_actor_template = jinja_env.from_string(config_h2_actor["template"])
h2_cumulative_input_token = 0
h2_cumulative_output_token = 0
updated_sections = []
with ThreadPoolExecutor() as executor:
# Kick off all tasks
future_to_section = {
executor.submit(
layout_h2_process_section,
section,
outline_no_sub_locations,
h2_actor_template,
create_h2_actor_agent # pass the factory function
): section
for section in sections
}
# Gather results as they complete
for future in concurrent.futures.as_completed(future_to_section):
section = future_to_section[future]
sec, final_sub_loc, in_toks, out_toks = future.result()
# Accumulate token usage
h2_cumulative_input_token += in_toks
h2_cumulative_output_token += out_toks
# Stash the final sub-loc for merging
updated_sections.append((sec, final_sub_loc))
# Now merge each updated subsection location back into outline_no_sub_locations
for (section, final_sub_loc) in updated_sections:
outline_no_sub_locations = fill_location(
outline_no_sub_locations,
section,
final_sub_loc
)
consumption_log['h2_actor'].append((h2_cumulative_input_token, h2_cumulative_output_token))
total_input_token += h2_cumulative_input_token
total_output_token += h2_cumulative_output_token
outline_no_sub_locations['meta'] = outline['meta']
outline_no_sub_locations_with_name = recover_name_and_location(outline_no_sub_locations, new_outline)
new_outline = outline_no_sub_locations_with_name
### Outline finalized, actually generate layout
logs = {}
gen_layout_cumulative_input_token = 0
gen_layout_cumulative_output_token = 0
init_outline = {'meta': outline['meta'], sections[0]: outline[sections[0]]}
new_outline = outline
init_jinja_args = {
'json_outline': init_outline,
'function_docs': documentation
}
outline_location = get_outline_location(outline, subsection=False)
logs = {}
# We'll store all updated logs here, keyed by section_index.
parallel_results = {}
with concurrent.futures.ThreadPoolExecutor() as executor:
futures = []
for section_index in range(len(sections)):
if section_index == 0:
create_agent_fn = create_init_actor_agent
prompt_template = init_template
else:
create_agent_fn = create_new_section_actor_agent
prompt_template = new_section_template
future = executor.submit(
layout_process_section,
section_index,
sections,
new_outline,
prompt_template,
documentation,
create_agent_fn,
args.max_retry,
name_to_hierarchy,
slide_width,
slide_height
)
futures.append(future)
# Collect the results as they come in
for future in concurrent.futures.as_completed(futures):
try:
section_index, section_logs, in_toks, out_toks = future.result()
# Store these logs in a dictionary keyed by the section index
parallel_results[section_index] = section_logs
# Update token counters
gen_layout_cumulative_input_token += in_toks
gen_layout_cumulative_output_token += out_toks
except Exception as exc:
print(f"[ERROR] A section failed: {exc}", flush=True)
# Possibly re-raise if you want to stop everything on error
# raise
# After all tasks complete, merge the results back into `logs`
for section_index, section_logs in parallel_results.items():
curr_section = sections[section_index]
logs[curr_section] = section_logs
concatenated_code = utils_functions
for section_index in range(len(sections)):
section = sections[section_index]
concatenated_code += '\n' + logs[section][-1]['code']
concatenated_code += '\n' + f'save_presentation(presentation, file_name="poster_{section_index + 1}.pptx")'
concatenated_code += f'''
name_to_hierarchy = {name_to_hierarchy}
identifier = "parallel"
poster_path = "poster_{section_index + 1}.pptx"
get_visual_cues(name_to_hierarchy, identifier, poster_path)
'''
output, error = run_code(concatenated_code)
if error is not None:
print(f'Failed to generate layout for section {curr_section}.')
consumption_log['h1_actor'].append((layout_cumulative_input_token, layout_cumulative_output_token))
total_input_token += gen_layout_cumulative_input_token
total_output_token += gen_layout_cumulative_output_token
h1_path = f'tmp/poster_<parallel>_hierarchy_1.pptx'
h2_path = f'tmp/poster_<parallel>_hierarchy_2.pptx'
h1_filled_path = f'tmp/poster_<parallel>_hierarchy_1_filled.pptx'
h2_filled_path = f'tmp/poster_<parallel>_hierarchy_2_filled.pptx'
ppt_to_images(h1_path, 'tmp/layout_h1')
ppt_to_images(h2_path, 'tmp/layout_h2')
ppt_to_images(h1_filled_path, 'tmp/layout_h1_filled')
ppt_to_images(h2_filled_path, 'tmp/layout_h2_filled')
h1_img = Image.open('tmp/layout_h1/slide_0001.jpg')
h2_img = Image.open('tmp/layout_h2/slide_0001.jpg')
h1_filled_img = Image.open('tmp/layout_h1_filled/slide_0001.jpg')
h2_filled_img = Image.open('tmp/layout_h2_filled/slide_0001.jpg')
ckpt = {
'logs': logs,
'outline': new_outline,
'name_to_hierarchy': name_to_hierarchy,
'consumption_log': consumption_log,
'total_input_token': total_input_token,
'total_output_token': total_output_token,
}
with open(f'checkpoints/{args.model_name}_{args.poster_name}_ckpt_{args.index}.pkl', 'wb') as f:
pkl.dump(ckpt, f)
json.dump(
new_outline,
open(outline_file_path, "w"),
ensure_ascii=False,
indent=4,
)
return total_input_token, total_output_token
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--poster_name', type=str, default=None)
parser.add_argument('--model_name', type=str, default='4o')
parser.add_argument('--poster_path', type=str, required=True)
parser.add_argument('--index', type=int, default=0)
parser.add_argument('--max_retry', type=int, default=3)
args = parser.parse_args()
actor_config = get_agent_config(args.model_name)
critic_config = get_agent_config(args.model_name)
if args.poster_name is None:
args.poster_name = args.poster_path.split('/')[-1].replace('.pdf', '').replace(' ', '_')
input_token, output_token = filter_image_table(args, actor_config)
print(f'Token consumption: {input_token} -> {output_token}', flush=True)
input_token, output_token = gen_outline_layout(args, actor_config, critic_config)
print(f'Token consumption: {input_token} -> {output_token}', flush=True) |