File size: 12,044 Bytes
cf8c8d4 | 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 | r"""Synthetic document pages with structure annotations -> page-level Markdown training records.
Input: parquet files with an `image` column and a `json` column holding the page annotation
(`page_information`, `element_bboxes` in reading order with types such as heading, paragraph, list, table
with cells, formula with LaTeX, caption, figure, header, footer, page_number, watermark).
The target is written in the parsing model's own output style (OvisOCR-style document Markdown):
- paragraphs separated by a blank line; headings as Markdown headings (`#` per level);
- tables as compact HTML `<table><thead><tr><td>..</td></tr></thead><tbody>..</tbody></table>`, with
rowspan / colspan, and the cells of every row in **visual left-to-right order** (the convention of
OmniDocBench-style ground truth, also for right-to-left scripts);
- display formulas as `$$ latex $$`; figures as `<img src="images/bbox_l_t_r_b.jpg" />` on a 0-1000 grid;
- running headers, footers and page numbers kept as plain text lines (as the model writes them);
watermarks dropped.
A target is only as good as its match to the model's style: check it with `--validate` first (the base
model's reading of clean pages should agree with the converted target).
python -m common.synth_convert --input "data/raw/synth_OCR/data/*/*.parquet" --out data/prepared/synth \
--per-subset 400 --seed 0
"""
import argparse
import glob
import io
import json
import random
import re
from collections import Counter
from pathlib import Path
from .vlm_utils import write_jsonl
SKIP = {"watermark"}
TEXT_TYPES = {"paragraph", "caption", "footnote", "quote", "stat", "header", "footer", "page_number", "title"}
NUM = re.compile(r"^\s*\d+[.)]\s*")
BULLET = re.compile(r"^\s*[-•·*]\s*")
REPEAT = re.compile(r"(.)\1{40,}|(.{2,12}?)\2{15,}", re.S) # degenerate runs: never train on them
def _cells_html(el):
"""HTML for a table element; cells of each row in visual left-to-right order."""
cells = el.get("cells") or []
if not cells:
return None
rows = {}
for c in cells:
rows.setdefault(c["row"], []).append(c)
head, body = [], []
for r in sorted(rows):
ordered = sorted(rows[r], key=lambda c: c["bbox"][0]) # left edge, left to right
tds = []
for c in ordered:
attrs = ""
if c.get("rowspan", 1) > 1:
attrs += f' rowspan="{c["rowspan"]}"'
if c.get("colspan", 1) > 1:
attrs += f' colspan="{c["colspan"]}"'
tds.append(f"<td{attrs}>{_inline(c.get('text', ''))}</td>")
tr = "<tr>" + "".join(tds) + "</tr>"
(head if all(c.get("header") for c in rows[r]) and not body else body).append(tr)
html = "<table>"
if head:
html += "<thead>" + "".join(head) + "</thead>"
if body:
html += "<tbody>" + "".join(body) + "</tbody>"
return html + "</table>"
def _inline(s):
return re.sub(r"\s*\n\s*", " ", s or "").strip()
def _list(el):
items = [x.strip() for x in (el.get("text") or "").split("\n") if x.strip()]
if el.get("ordered"): # (no backslashes inside f-strings: Python < 3.12 rejects them)
return "\n".join(str(i) + ". " + NUM.sub("", x) for i, x in enumerate(items, 1))
return "\n".join("- " + BULLET.sub("", x) for x in items)
def element_md(el, img_w, img_h):
t = el["type"]
if t in SKIP:
return None
if t == "table":
return _cells_html(el)
if t == "formula":
latex = (el.get("latex") or "").strip()
return f"$$\n{latex}\n$$" if latex else None
if t == "figure":
x, y, w, h = el["bbox"]
sc = lambda v, d: max(0, min(999, int(v / d * 1000))) # noqa: E731
return f'<img src="images/bbox_{sc(x, img_w)}_{sc(y, img_h)}_{sc(x + w, img_w)}_{sc(y + h, img_h)}.jpg" />'
if t == "heading":
return "#" * max(1, min(6, int(el.get("level") or 2))) + " " + _inline(el.get("text"))
if t == "list":
return _list(el)
text = (el.get("text") or "").strip()
return text or None
def page_markdown(ann):
pi = ann["page_information"]
w, h = pi.get("image_width") or pi["width"], pi.get("image_height") or pi["height"]
parts = [element_md(e, w, h) for e in ann["element_bboxes"]]
return "\n\n".join(p for p in parts if p)
GT_CATEGORY = {"paragraph": "text_block", "list": "text_block", "quote": "text_block", "stat": "text_block",
"footnote": "text_block", "heading": "title", "title": "title", "header": "header", "footer": "footer",
"page_number": "page_number", "table": "table", "formula": "equation_isolated", "figure": "figure"}
LANG_NAME = {"ar": "Arabic", "en": "English", "zh": "Chinese"}
IGNORED = {"header", "footer", "page_number"} # no reading-order index, as in the benchmark's own ground truth
def page_gt(ann, image_name):
"""The page in OmniDocBench / MDPBench ground-truth form, so the official scorer can grade readings of it."""
els, dets, order = ann["element_bboxes"], [], 0
for i, e in enumerate(els):
t = e["type"]
if t == "caption":
tgt = e.get("target")
cat = "table_caption" if isinstance(tgt, int) and 0 <= tgt < len(els) and els[tgt]["type"] == "table" \
else "figure_caption"
else:
cat = GT_CATEGORY.get(t)
if cat is None:
continue
x, y, w, h = e["bbox"]
d = {"category_type": cat, "poly": [x, y, x + w, y, x + w, y + h, x, y + h],
"order": None if cat in IGNORED else order, "anno_id": i,
"attribute": {"text_language": LANG_NAME.get((e.get("lang") or ["und"])[0], "Other")},
"text": "", "html": "", "latex": ""}
if cat == "table":
d["html"] = _cells_html(e) or ""
elif cat == "equation_isolated":
d["latex"] = (e.get("latex") or "").strip()
elif cat != "figure":
d["text"] = (e.get("text") or "").strip() if t != "heading" else _inline(e.get("text"))
if cat not in IGNORED:
order += 1
dets.append(d)
lang = page_language(ann)
return {"layout_dets": dets, "extra": {"relation": []},
"page_info": {"image_path": image_name, "page_attribute": {"language": LANG_NAME.get(lang, "Other")}}}
def page_language(ann):
c = Counter(l for e in ann["element_bboxes"] for l in e.get("lang", []))
return c.most_common(1)[0][0] if c else "und"
def line_height(ann):
"""Median text-line height in image pixels: below ~16 px Arabic dots and diacritics are barely resolved,
and training on text the model cannot see teaches it to guess."""
hs = sorted(l["bbox"][3] for l in ann.get("line_bboxes", []) if l.get("bbox") and l["bbox"][3] > 0)
return hs[len(hs) // 2] if hs else 0.0
def page_id(subset, parquet_file, ann, fallback):
"""Unique per page: subset, shard number (train-00002-of-00008 -> 00002) and the page id in the shard."""
stem = Path(parquet_file).stem
m = re.match(r".*?-(\d+)-of-\d+$", stem)
shard = m.group(1) if m else stem
return f"synth_{subset}_{shard}_{ann['page_information'].get('id', fallback)}"
def effects(ann):
pi = ann["page_information"]
return "photo" if pi.get("photo") else "scan" if pi.get("scan") else "clean"
def main(argv=None):
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--input", required=True, help="glob of parquet files (subset = parent folder name)")
ap.add_argument("--out", required=True, help="folder for images/ and synth.jsonl")
ap.add_argument("--per-subset", type=int, default=0, help="pages per subset (0 = all)")
ap.add_argument("--only-clean", action="store_true", help="no photo / scan effects (for --validate sets)")
ap.add_argument("--languages", nargs="*", help="keep pages whose main language is one of these")
ap.add_argument("--min-line-px", type=float, default=0.0,
help="drop pages whose median text-line height is below this many image pixels (e.g. 16)")
ap.add_argument("--skip-subsets", nargs="*", default=[],
help="subset folders to leave out (e.g. receipts, whose key-value blocks are labelled as tables)")
ap.add_argument("--seed", type=int, default=0)
ap.add_argument("--bench", help="also write a benchmark subset of this name (gt.json + items.jsonl) for "
"`python -m common.mdpbench score`; page ids get a language prefix (ar_...)")
a = ap.parse_args(argv)
import pyarrow.parquet as pq
out = Path(a.out)
(out / "images").mkdir(parents=True, exist_ok=True)
by_subset = {}
for f in sorted(glob.glob(a.input)):
if Path(f).parent.name not in a.skip_subsets:
by_subset.setdefault(Path(f).parent.name, []).append(f)
rng = random.Random(a.seed)
records, dropped, gts, seen = [], Counter(), [], set()
for sub, files in sorted(by_subset.items()):
kept = 0
files = files[:]
rng.shuffle(files)
for f in files:
pf = pq.ParquetFile(f)
for batch in pf.iter_batches(batch_size=32):
for row in batch.to_pylist():
if a.per_subset and kept >= a.per_subset:
break
ann = json.loads(row["json"])
if a.only_clean and effects(ann) != "clean":
continue
lang = page_language(ann)
if a.languages and lang not in a.languages:
continue
lh = line_height(ann)
if lh < a.min_line_px:
dropped["text too small"] += 1
continue
text = page_markdown(ann)
if not text.strip():
dropped["empty"] += 1
continue
if REPEAT.search(text):
dropped["repetition"] += 1
continue
pid = page_id(sub, f, ann, kept)
if pid in seen:
raise SystemExit(f"duplicate page id {pid}: ids must be unique across shards")
seen.add(pid)
img = out / "images" / f"{pid}.png"
if not img.exists():
img.write_bytes(row["image"]["bytes"])
records.append({"id": pid, "image": str(img.resolve()), "text": text, "source": f"synth_OCR/{sub}",
"license": "see dataset card", "language": lang, "level": "page", "split": "train",
"effects": effects(ann), "template": ann["page_information"].get("template"),
"line_px": round(lh, 1)})
if a.bench:
gts.append(page_gt(ann, f"{lang}_{pid}.png"))
kept += 1
if a.per_subset and kept >= a.per_subset:
break
if a.per_subset and kept >= a.per_subset:
break
print(f"{sub}: {kept} pages", flush=True)
write_jsonl(out / "synth.jsonl", records)
if a.bench:
from .mdpbench import BENCH_DIR
d = BENCH_DIR / "subsets" / a.bench
d.mkdir(parents=True, exist_ok=True)
(d / "gt.json").write_text(json.dumps(gts, ensure_ascii=False), encoding="utf-8")
write_jsonl(d / "items.jsonl", [{"id": f"{r['language']}_{r['id']}", "image": r["image"], "text": "",
"level": "page", "language": r["language"]} for r in records])
print(f"benchmark subset {a.bench}: {len(gts)} pages -> {d}")
langs = Counter(r["language"] for r in records)
print(f"{len(records)} records -> {out / 'synth.jsonl'} | languages {dict(langs)} | dropped {dict(dropped)}")
if __name__ == "__main__":
main()
|