PosterText-30K / code /postereval /prepare_ir.py
HGTasd's picture
Initial clean release
ec21fa4 verified
#!/usr/bin/env python3
"""Generate PosterEval IR from poster PNG/JPEG images.
The public flow intentionally uses two IR prompts:
- content IR for Order, Completeness, and Claim F1.
- figure IR for Local Text-Figure Alignment.
"""
import argparse
import json
import re
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path
from typing import Any, Dict, Iterable, List, Optional, Tuple
from openrouter_client import call_openrouter_json
DEFAULT_WORKERS = 8
PROMPT_DIR = Path(__file__).resolve().parent / "prompts"
PARSER_TO_PROMPT = {
"content": PROMPT_DIR / "content_ir_prompt.md",
"figure": PROMPT_DIR / "figure_grounding_ir_prompt.md",
}
def normalize_key(text: str) -> str:
return re.sub(r"[^a-z0-9]+", "", text.lower())
def extract_key(name: str, pattern: Optional[str]) -> str:
if pattern:
match = re.search(pattern, name)
if match:
return match.groupdict().get("key") or match.group(1)
if re.fullmatch(r"\d+", name):
return name
return normalize_key(name)
def should_ignore(name: str, patterns: Iterable[str]) -> bool:
return any(re.search(pattern, name) for pattern in patterns)
def choose_image(dir_path: Path, image_filename: Optional[str]) -> Optional[Path]:
if image_filename:
candidate = dir_path / image_filename
if candidate.exists():
return candidate
priority = [
"poster.png",
"paper.png",
"poster.jpg",
"paper.jpg",
"poster.jpeg",
"paper.jpeg",
]
for name in priority:
candidate = dir_path / name
if candidate.exists():
return candidate
images = []
for pattern in ("*.png", "*.jpg", "*.jpeg"):
images.extend(dir_path.glob(pattern))
return sorted(images, key=lambda p: p.name)[0] if images else None
def discover_images(
input_root: Path,
layout: str,
image_filename: Optional[str],
image_glob: str,
key_regex: Optional[str],
ignore_dir_regex: List[str],
) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
items: List[Dict[str, Any]] = []
missing: List[Dict[str, Any]] = []
if layout == "flat":
for image_path in sorted(input_root.glob(image_glob), key=lambda p: p.name):
if not image_path.is_file():
continue
key = extract_key(image_path.stem, key_regex)
items.append(
{
"key": key,
"source_name": image_path.stem,
"image_path": image_path,
"input_relpath": image_path.name,
}
)
return items, missing
for child in sorted(input_root.iterdir(), key=lambda p: p.name):
if not child.is_dir() or should_ignore(child.name, ignore_dir_regex):
continue
key = extract_key(child.name, key_regex)
image_path = choose_image(child, image_filename)
if image_path is None:
missing.append({"key": key, "source_name": child.name})
continue
items.append(
{
"key": key,
"source_name": child.name,
"image_path": image_path,
"input_relpath": str(image_path.relative_to(input_root)),
}
)
return items, missing
def infer_role(title: str, text_content: str) -> str:
text = f"{title} {text_content}".lower()
problem_kw = ["problem", "motivation", "challenge", "background"]
approach_kw = ["approach", "method", "framework", "model", "pipeline"]
evidence_kw = ["result", "evaluation", "experiment", "ablation", "analysis"]
if any(keyword in text for keyword in problem_kw):
return "Problem"
if any(keyword in text for keyword in evidence_kw):
return "Evidence"
if any(keyword in text for keyword in approach_kw):
return "Approach"
return "Approach"
def normalize_figure_ir(data: Dict[str, Any]) -> Dict[str, Any]:
if not isinstance(data, dict):
return {"sections": []}
sections = []
for section_index, section in enumerate(data.get("sections", []), start=1):
if not isinstance(section, dict):
continue
title = section.get("title", "") or ""
text_content = section.get("text_content", "") or ""
out_section = {
"id": section.get("id") or f"s{section_index}",
"title": title,
"bbox": section.get("bbox"),
"text_content": text_content,
"meta_role": section.get("meta_role") or infer_role(title, text_content),
"contains_figures": [],
}
figures = section.get("contains_figures") or section.get("figures") or []
for fig_index, figure in enumerate(figures, start=1):
if not isinstance(figure, dict):
continue
description = figure.get("description", "") or ""
out_section["contains_figures"].append(
{
"id": figure.get("id") or f"f{section_index}_{fig_index}",
"bbox": figure.get("bbox"),
"type": figure.get("type", "Chart"),
"caption": figure.get("caption") or description,
"semantic_summary": figure.get("semantic_summary") or description,
"visual_description": figure.get("visual_description") or description,
}
)
sections.append(out_section)
payload: Dict[str, Any] = {"sections": sections}
if "meta" in data:
payload["meta"] = data["meta"]
return payload
def parser_output_dir(output_root: Path, parser_name: str) -> Path:
if parser_name == "content":
return output_root / "content_ir"
if parser_name == "figure":
return output_root / "figure_ir"
raise ValueError(f"Unknown parser: {parser_name}")
def write_ir(
output_root: Path,
parser_name: str,
key: str,
payload: Dict[str, Any],
) -> Path:
out_dir = parser_output_dir(output_root, parser_name) / key
out_dir.mkdir(parents=True, exist_ok=True)
out_path = out_dir / "poster_ir.json"
out_path.write_text(
json.dumps(payload, ensure_ascii=False, indent=2) + "\n",
encoding="utf-8",
)
return out_path
def run_one_parser(
item: Dict[str, Any],
output_root: Path,
parser_name: str,
model: str,
temperature: float,
include_paths: bool,
) -> Dict[str, Any]:
prompt_path = PARSER_TO_PROMPT[parser_name]
prompt = prompt_path.read_text(encoding="utf-8")
record = {
"key": item["key"],
"source_name": item["source_name"],
"input_relpath": item["input_relpath"],
"parser": parser_name,
"model": model,
"prompt_file": str(prompt_path.relative_to(PROMPT_DIR.parent)),
"error": "",
}
if include_paths:
record["image_path"] = str(item["image_path"])
try:
raw = call_openrouter_json(
prompt=prompt,
model=model,
image_path=item["image_path"],
max_tokens=16384,
temperature=temperature,
response_format_json=True,
)
if raw.get("parse_error"):
raise RuntimeError(raw.get("error", "model response JSON parse error"))
payload = normalize_figure_ir(raw) if parser_name == "figure" else raw
payload.setdefault("_postereval", {})
payload["_postereval"].update(
{
"source_name": item["source_name"],
"parser": parser_name,
"model": model,
"prompt_file": record["prompt_file"],
}
)
if include_paths:
payload["_postereval"]["image_path"] = str(item["image_path"])
else:
payload["_postereval"]["input_relpath"] = item["input_relpath"]
out_path = write_ir(output_root, parser_name, item["key"], payload)
record["output_relpath"] = str(out_path.relative_to(output_root))
except Exception as exc:
record["error"] = repr(exc)
return record
def expand_parsers(value: str) -> List[str]:
if value == "both":
return ["content", "figure"]
parsers = []
for part in value.split(","):
part = part.strip()
if part in {"content", "figure"}:
parsers.append(part)
elif part in {"content_ir"}:
parsers.append("content")
elif part in {"figure_ir"}:
parsers.append("figure")
else:
raise ValueError(f"Unknown parser: {part}")
return list(dict.fromkeys(parsers))
def run(args: argparse.Namespace) -> None:
input_root = Path(args.input_root).expanduser()
output_root = Path(args.output_root).expanduser()
output_root.mkdir(parents=True, exist_ok=True)
items, missing = discover_images(
input_root=input_root,
layout=args.layout,
image_filename=args.image_filename,
image_glob=args.image_glob,
key_regex=args.key_regex,
ignore_dir_regex=args.ignore_dir_regex,
)
parsers = expand_parsers(args.parser)
records = []
with ThreadPoolExecutor(max_workers=args.workers) as executor:
futures = []
for item in items:
for parser_name in parsers:
futures.append(
executor.submit(
run_one_parser,
item,
output_root,
parser_name,
args.model,
args.temperature,
args.include_paths,
)
)
for future in as_completed(futures):
records.append(future.result())
records.sort(key=lambda r: (r["parser"], r["key"], r["source_name"]))
summary = {
"input_root_exists": input_root.exists(),
"layout": args.layout,
"n_images": len(items),
"parsers": parsers,
"model": args.model,
"temperature": args.temperature,
"n_records": len(records),
"n_errors": sum(1 for record in records if record.get("error")),
"missing_images": missing,
"records": records,
}
if args.include_paths:
summary["input_root"] = str(input_root)
(output_root / "summary.json").write_text(
json.dumps(summary, ensure_ascii=False, indent=2) + "\n",
encoding="utf-8",
)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Generate PosterEval IR from poster images.")
parser.add_argument("--input-root", required=True, help="Poster image root.")
parser.add_argument("--output-root", required=True, help="IR output root.")
parser.add_argument(
"--parser",
default="both",
help="content / figure / both, or a comma-separated combination.",
)
parser.add_argument(
"--layout",
choices=["directories", "flat"],
default="directories",
help="Input layout: one image per subdirectory, or flat image files.",
)
parser.add_argument("--image-filename", help="Expected image filename inside each directory.")
parser.add_argument("--image-glob", default="*.png", help="Glob used when --layout flat.")
parser.add_argument("--key-regex", help="Regex with optional named group 'key'.")
parser.add_argument(
"--ignore-dir-regex",
action="append",
default=[r"^_", r"^\.", r"^__pycache__$"],
help="Directory regex to skip. Can be repeated.",
)
parser.add_argument("--model", default="qwen3-vl-235b", help="OpenRouter model alias or id.")
parser.add_argument(
"--temperature",
type=float,
default=0.02,
help="IR parser sampling temperature. Default matches the paper protocol.",
)
parser.add_argument("--workers", type=int, default=DEFAULT_WORKERS)
parser.add_argument(
"--include-paths",
action="store_true",
help="Include absolute local paths in outputs. Keep off for anonymous artifacts.",
)
return parser.parse_args()
def main() -> None:
run(parse_args())
if __name__ == "__main__":
main()