JaceWei's picture
update
4469ee0
from dotenv import load_dotenv
# from utils.src.utils import get_json_from_response
from utils.src.model_utils import parse_pdf
import json
import random
import json_repair
from camel.models import ModelFactory
from camel.agents import ChatAgent
from tenacity import retry, stop_after_attempt
from docling_core.types.doc import ImageRefMode, PictureItem, TableItem
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions
from docling.document_converter import DocumentConverter, PdfFormatOption
from pathlib import Path
import PIL
from marker.models import create_model_dict
from utils.wei_utils import *
from utils.pptx_utils import *
from utils.critic_utils import *
import torch
from jinja2 import Template
import re
import argparse
import os
load_dotenv()
IMAGE_RESOLUTION_SCALE = 5.0
pipeline_options = PdfPipelineOptions()
pipeline_options.images_scale = IMAGE_RESOLUTION_SCALE
pipeline_options.generate_page_images = True
pipeline_options.generate_picture_images = True
doc_converter = DocumentConverter(
format_options={
InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)
}
)
def account_token(response):
input_token = response.info['usage']['prompt_tokens']
output_token = response.info['usage']['completion_tokens']
return input_token, output_token
def get_json_from_response(raw_response: str):
response = raw_response.strip()
l, r = response.rfind("```json"), response.rfind("```")
try:
if l == -1 or r == -1:
response = json_repair.loads(response)
else:
response = json_repair.loads(response[l + 7 : r].strip())
return response
except Exception as e:
raise RuntimeError("Failed to parse JSON from response", e)
@retry(stop=stop_after_attempt(5))
def parse_raw(args, actor_config, version=2):
raw_source = args.poster_path
markdown_clean_pattern = re.compile(r"<!--[\s\S]*?-->")
print(f'\nParsing raw content from {raw_source}...\n')
try:
raw_result = doc_converter.convert(raw_source)
print('✅ PDF converted to document format.')
raw_markdown = raw_result.document.export_to_markdown()
text_content = markdown_clean_pattern.sub("", raw_markdown)
print(f'Extracted {len(text_content)} characters from the document.')
if len(text_content) < 500:
print('\nParsing with docling failed, using marker instead\n')
parser_model = create_model_dict(device='cuda', dtype=torch.float16)
text_content, rendered = parse_pdf(raw_source, model_lst=parser_model, save_file=False)
except Exception as e:
print(f'❌ PDF parsing failed: {e}')
raise e
# Load prompt template safely
try:
if version == 1:
template = Template(open("utils/prompts/gen_poster_raw_content.txt").read())
elif version == 2:
print('Using v2 prompt template')
template = Template(open("utils/prompts/gen_poster_raw_content_v2.txt").read())
else:
raise ValueError("Invalid version number.")
except Exception as e:
print(f'❌ Failed to load prompt template: {e}')
raise e
# Initialize actor model
try:
if args.model_name_t.startswith('vllm_qwen'):
actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
url=actor_config['url'],
)
else:
actor_model = ModelFactory.create(
model_platform=actor_config['model_platform'],
model_type=actor_config['model_type'],
model_config_dict=actor_config['model_config'],
)
actor_sys_msg = 'You are the author of the paper, and you will create a poster for the paper.'
actor_agent = ChatAgent(
system_message=actor_sys_msg,
model=actor_model,
message_window_size=10,
token_limit=actor_config.get('token_limit', None)
)
except Exception as e:
print(f'❌ Failed to initialize actor model: {e}')
raise e
# === main conversation loop ===
max_retry = 5
content_json = {}
for attempt in range(max_retry):
try:
print(f'\n🌀 Generating poster content... (Attempt {attempt+1}/{max_retry})')
prompt = template.render(markdown_document=text_content)
actor_agent.reset()
response = actor_agent.step(prompt)
input_token, output_token = account_token(response)
content_json = get_json_from_response(response.msgs[0].content)
if len(content_json) > 0:
print('✅ Successfully parsed JSON content.')
break
else:
print('⚠️ Empty JSON response, retrying...')
if args.model_name_t.startswith('vllm_qwen'):
text_content = text_content[:80000]
except Exception as e:
print(f'❌ Error in actor_agent loop: {e}')
# short delay could be added if needed
else:
raise RuntimeError("Failed to get valid JSON content after multiple attempts.")
# Trim sections if too many
if len(content_json.get('sections', [])) > 9:
selected_sections = (
content_json['sections'][:2]
+ random.sample(content_json['sections'][2:-2], 5)
+ content_json['sections'][-2:]
)
content_json['sections'] = selected_sections
has_title = False
for section in content_json['sections']:
if type(section) != dict or not 'title' in section or not 'content' in section:
print(f"Ouch! The response is invalid, the LLM is not following the format :(")
print('Trying again...')
raise
# if 'title' in section['title'].lower():
# has_title = True
# if not has_title:
# print('Ouch! The response is invalid, the LLM is not following the format :(')
# raise
os.makedirs('contents', exist_ok=True)
json.dump(content_json, open(f'contents/<{args.model_name_t}_{args.model_name_v}>_{args.poster_name}_raw_content.json', 'w'), indent=4)
return input_token, output_token, raw_result
def gen_image_and_table(args, conv_res):
input_token, output_token = 0, 0
raw_source = args.poster_path
output_dir = Path(f'<{args.model_name_t}_{args.model_name_v}>_images_and_tables/{args.poster_name}')
output_dir.mkdir(parents=True, exist_ok=True)
doc_filename = args.poster_name
# Save page images
for page_no, page in conv_res.document.pages.items():
page_no = page.page_no
page_image_filename = output_dir / f"{doc_filename}-{page_no}.png"
with page_image_filename.open("wb") as fp:
page.image.pil_image.save(fp, format="PNG")
# Save images of figures and tables
table_counter = 0
picture_counter = 0
for element, _level in conv_res.document.iterate_items():
if isinstance(element, TableItem):
table_counter += 1
element_image_filename = (
output_dir / f"{doc_filename}-table-{table_counter}.png"
)
with element_image_filename.open("wb") as fp:
element.get_image(conv_res.document).save(fp, "PNG")
if isinstance(element, PictureItem):
picture_counter += 1
element_image_filename = (
output_dir / f"{doc_filename}-picture-{picture_counter}.png"
)
with element_image_filename.open("wb") as fp:
element.get_image(conv_res.document).save(fp, "PNG")
# Save markdown with embedded pictures
md_filename = output_dir / f"{doc_filename}-with-images.md"
conv_res.document.save_as_markdown(md_filename, image_mode=ImageRefMode.EMBEDDED)
# Save markdown with externally referenced pictures
md_filename = output_dir / f"{doc_filename}-with-image-refs.md"
conv_res.document.save_as_markdown(md_filename, image_mode=ImageRefMode.REFERENCED)
# Save HTML with externally referenced pictures
html_filename = output_dir / f"{doc_filename}-with-image-refs.html"
conv_res.document.save_as_html(html_filename, image_mode=ImageRefMode.REFERENCED)
tables = {}
table_index = 1
for table in conv_res.document.tables:
caption = table.caption_text(conv_res.document)
if len(caption) > 0:
table_img_path = f'<{args.model_name_t}_{args.model_name_v}>_images_and_tables/{args.poster_name}/{args.poster_name}-table-{table_index}.png'
table_img = PIL.Image.open(table_img_path)
tables[str(table_index)] = {
'caption': caption,
'table_path': table_img_path,
'width': table_img.width,
'height': table_img.height,
'figure_size': table_img.width * table_img.height,
'figure_aspect': table_img.width / table_img.height,
}
table_index += 1
images = {}
image_index = 1
for image in conv_res.document.pictures:
caption = image.caption_text(conv_res.document)
if len(caption) > 0:
image_img_path = f'<{args.model_name_t}_{args.model_name_v}>_images_and_tables/{args.poster_name}/{args.poster_name}-picture-{image_index}.png'
image_img = PIL.Image.open(image_img_path)
images[str(image_index)] = {
'caption': caption,
'image_path': image_img_path,
'width': image_img.width,
'height': image_img.height,
'figure_size': image_img.width * image_img.height,
'figure_aspect': image_img.width / image_img.height,
}
image_index += 1
json.dump(images, open(f'<{args.model_name_t}_{args.model_name_v}>_images_and_tables/{args.poster_name}_images.json', 'w'), indent=4)
json.dump(tables, open(f'<{args.model_name_t}_{args.model_name_v}>_images_and_tables/{args.poster_name}_tables.json', 'w'), indent=4)
return input_token, output_token, images, tables
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)
args = parser.parse_args()
agent_config = get_agent_config(args.model_name)
if args.poster_name is None:
args.poster_name = args.poster_path.split('/')[-1].replace('.pdf', '').replace(' ', '_')
# Parse raw content
input_token, output_token = parse_raw(args, agent_config)
# Generate images and tables
_, _ = gen_image_and_table(args)
print(f'Token consumption: {input_token} -> {output_token}')