ftu-VietFinTab / scripts /stats.py
dinhanhx's picture
Claude Opus 5.5
Add conversion scripts (YOLO, COCO, HTML) and tests
eb58dca
Raw History Blame Contribute Delete
4.14 kB
# /// script
# requires-python = ">=3.10"
# dependencies = ["matplotlib", "numpy"]
# ///
"""Dataset-level charts: tables per company, company x period heatmap, structure distributions.
Usage: uv run scripts/stats.py [--root .] [--out figures]
"""
import argparse
import json
import xml.etree.ElementTree as ET
from collections import Counter
from pathlib import Path
import matplotlib.pyplot as plt
import numpy as np
BLUE = "#2a78d6"
INK, MUTED = "#0b0b0b", "#52514e"
QORDER = {"Q1": 0, "Q2": 1, "Q3": 2, "Q4": 3, "CN": 4}
plt.rcParams.update({
"axes.spines.top": False, "axes.spines.right": False,
"axes.edgecolor": MUTED, "axes.labelcolor": MUTED, "xtick.color": MUTED,
"ytick.color": MUTED, "text.color": INK, "axes.titlelocation": "left",
"axes.titleweight": "bold", "figure.dpi": 150,
})
def load(root):
return [json.loads(l) for l in (root / "metadata.jsonl").read_text().splitlines()]
def per_company(meta, out):
counts = Counter(m["company"] for m in meta).most_common()
names, vals = zip(*counts)
fig, ax = plt.subplots(figsize=(7, 4.5))
bars = ax.barh(names[::-1], vals[::-1], color=BLUE, height=0.6)
ax.bar_label(bars, padding=3, color=INK)
ax.set(xlabel="Table images", title=f"Tables per company (n={len(meta):,})")
ax.set_xlim(0, max(vals) * 1.1)
fig.tight_layout()
fig.savefig(out / "tables_per_company.png")
def company_period(meta, out):
periods = sorted({(m["year"], m["quarter"]) for m in meta}, key=lambda p: (p[0], QORDER[p[1]]))
companies = sorted({m["company"] for m in meta})
grid = np.zeros((len(companies), len(periods)), dtype=int)
for m in meta:
grid[companies.index(m["company"]), periods.index((m["year"], m["quarter"]))] += 1
fig, ax = plt.subplots(figsize=(10, 5.5))
im = ax.imshow(grid, cmap="Blues", aspect="auto")
ax.set_xticks(range(len(periods)), [f"{q}\n{y}" for y, q in periods])
ax.set_yticks(range(len(companies)), companies)
for (i, j), v in np.ndenumerate(grid):
ax.text(j, i, v or "", ha="center", va="center", fontsize=7,
color="white" if v > grid.max() * 0.55 else INK)
ax.set_title("Tables per company and reporting period")
for s in ax.spines.values():
s.set_visible(False)
fig.colorbar(im, ax=ax, label="Table images", shrink=0.8)
fig.tight_layout()
fig.savefig(out / "tables_per_period.png")
def parse(path):
r = ET.parse(path).getroot()
names = Counter(o.findtext("name") for o in r.iter("object"))
return (int(r.findtext("size/width")), int(r.findtext("size/height")),
names["table row"], names["table column"], names["table spanning cell"])
def structure(meta, root, out):
w, h, rows, cols, spans = np.array([parse(root / m["structure_file"]) for m in meta]).T
panels = [("Rows per table", rows), ("Columns per table", cols),
("Spanning cells per table", spans), ("Image width (px)", w),
("Image height (px)", h), ("Aspect ratio (w/h)", w / h)]
fig, axes = plt.subplots(2, 3, figsize=(11, 6))
for ax, (title, x) in zip(axes.flat, panels):
hi = np.percentile(x, 99) # clip the long tail so the bulk stays readable
ax.hist(x[x <= hi], bins=30, color=BLUE, edgecolor="white", linewidth=0.5)
ax.set(title=title, ylabel="Tables")
ax.annotate(f"median {np.median(x):.4g}\np99 {hi:.4g}", (0.97, 0.95),
xycoords="axes fraction", ha="right", va="top", color=MUTED, fontsize=8)
fig.suptitle("Structure annotation statistics (x-axis clipped at p99)", x=0.01, ha="left",
fontweight="bold")
fig.tight_layout()
fig.savefig(out / "structure_stats.png")
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--root", type=Path, default=Path(__file__).resolve().parent.parent)
ap.add_argument("--out", type=Path, default=Path("figures"))
a = ap.parse_args()
a.out.mkdir(exist_ok=True)
meta = load(a.root)
per_company(meta, a.out)
company_period(meta, a.out)
structure(meta, a.root, a.out)
print(f"wrote 3 figures to {a.out}/")