eMCR / scripts /make_report.py
Lux1997's picture
Add scripts
cbb33d5 verified
Raw
History Blame Contribute Delete
12.4 kB
"""Render the mcr evaluation results into a self-contained HTML report.
Reads runs/full/<model>/<benchmark>/*.json, computes macro tables + multi-condition
stratified curves, and emits a single HTML file with inline SVG charts (no external
CDN, works offline).
Usage:
PYTHONNOUSERSITE=1 python scripts/make_report.py [runs_dir] [out_html]
"""
from __future__ import annotations
import glob
import json
import os
import sys
from collections import defaultdict
RUNS = sys.argv[1] if len(sys.argv) > 1 else "runs/full"
OUT = sys.argv[2] if len(sys.argv) > 2 else "report.html"
MODELS = ["emb-0.6B", "emb-4B", "emb-8B", "vl-2B", "vl-8B"]
BENCH = ["multiconir", "mcmr", "ssrb_subset", "merit"]
BENCH_LABEL = {
"multiconir": "MultiConIR<br><small>散文·纯文本</small>",
"mcmr": "MCMR<br><small>文→图文</small>",
"ssrb_subset": "SSRB<br><small>JSON 半结构化</small>",
"merit": "MERIT<br><small>交错多图</small>",
}
METRICS = ["ndcg@10", "recall@10", "mrr@10", "recall@100"]
def load():
d = defaultdict(list)
for p in glob.glob(os.path.join(RUNS, "*", "*", "*.json")):
if p.endswith("summary.json"):
continue
try:
r = json.load(open(p))
except Exception:
continue
if "dense" not in r:
continue
m = p.split(os.sep)[-3]
b = p.split(os.sep)[-2]
d[(m, b)].append(r)
return d
D = load()
def macro(m, b, metric):
reps = D.get((m, b), [])
vs = [r["dense"]["metrics"].get(metric) for r in reps]
vs = [x for x in vs if x is not None]
return sum(vs) / len(vs) if vs else None
def n_tasks(m, b):
return len(D.get((m, b), []))
def strat(m, b, dim, metric="ndcg@10"):
buck = defaultdict(list)
for r in D.get((m, b), []):
st = r["dense"].get("stratified", {}).get(dim, {})
for k, mm in st.items():
if mm.get(metric) is not None:
buck[k].append(mm[metric])
return {k: sum(v) / len(v) for k, v in buck.items()}
# ---------- HTML helpers ----------
def heat(v, lo, hi):
"""Green(high)->red(low) background for a value within [lo,hi]."""
if v is None:
return "background:#1b1f2a;color:#555"
t = 0.0 if hi == lo else (v - lo) / (hi - lo)
r = int(200 * (1 - t) + 40 * t)
g = int(60 * (1 - t) + 170 * t)
b = 70
return f"background:rgb({r},{g},{b});color:#fff"
def matrix_table(metric):
# per-column (benchmark) min/max for heat scaling
cols = {}
for b in BENCH:
vals = [macro(m, b, metric) for m in MODELS]
vals = [x for x in vals if x is not None]
cols[b] = (min(vals), max(vals)) if vals else (0, 1)
h = ['<table class="mat"><thead><tr><th>model</th>']
for b in BENCH:
h.append(f"<th>{BENCH_LABEL[b]}</th>")
h.append("</tr></thead><tbody>")
for m in MODELS:
h.append(f'<tr><td class="mname">{m}</td>')
for b in BENCH:
v = macro(m, b, metric)
style = heat(v, *cols[b])
txt = f"{v:.3f}" if v is not None else "—"
nt = n_tasks(m, b)
sub = f'<span class="nt">{nt}</span>' if v is not None else ""
h.append(f'<td style="{style}">{txt}{sub}</td>')
h.append("</tr>")
h.append("</tbody></table>")
return "".join(h)
PALETTE = ["#e15759", "#f28e2b", "#edc948", "#59a14f", "#4e79a7"]
def svg_line(series: dict, title: str, xlabel: str, width=560, height=300):
"""series: {name: {x(str-numeric): y}}; draws a line chart as inline SVG."""
xs = sorted({float(x) for s in series.values() for x in s}, key=float)
if not xs:
return f"<p>{title}: (no data)</p>"
ys = [y for s in series.values() for y in s.values()]
ymax = max(ys + [0.001])
ymax = (int(ymax / 0.1) + 1) * 0.1 # round up to 0.1
pad_l, pad_b, pad_t, pad_r = 48, 38, 36, 120
pw, ph = width - pad_l - pad_r, height - pad_b - pad_t
def px(x):
return pad_l + (xs.index(x) / max(1, len(xs) - 1)) * pw
def py(y):
return pad_t + ph - (y / ymax) * ph
out = [f'<svg viewBox="0 0 {width} {height}" class="chart">']
out.append(f'<text x="{width/2}" y="18" class="ctitle" text-anchor="middle">{title}</text>')
# y gridlines
steps = 5
for i in range(steps + 1):
yv = ymax * i / steps
yy = py(yv)
out.append(f'<line x1="{pad_l}" y1="{yy:.1f}" x2="{pad_l+pw}" y2="{yy:.1f}" class="grid"/>')
out.append(f'<text x="{pad_l-6}" y="{yy+3:.1f}" class="axlbl" text-anchor="end">{yv:.2f}</text>')
# x labels
for x in xs:
xx = px(x)
out.append(f'<text x="{xx:.1f}" y="{pad_t+ph+16:.1f}" class="axlbl" text-anchor="middle">{int(x)}</text>')
out.append(f'<text x="{pad_l+pw/2:.1f}" y="{height-4}" class="axlbl" text-anchor="middle">{xlabel}</text>')
# lines
for i, (name, s) in enumerate(series.items()):
col = PALETTE[i % len(PALETTE)]
pts = []
for x in xs:
key = str(int(x)) if str(int(x)) in s else (str(x) if str(x) in s else None)
if key is None:
continue
pts.append((px(x), py(s[key])))
if not pts:
continue
d = "M" + " L".join(f"{a:.1f},{b:.1f}" for a, b in pts)
out.append(f'<path d="{d}" fill="none" stroke="{col}" stroke-width="2.2"/>')
for a, b in pts:
out.append(f'<circle cx="{a:.1f}" cy="{b:.1f}" r="3" fill="{col}"/>')
ly = pad_t + 6 + i * 18
out.append(f'<line x1="{pad_l+pw+12}" y1="{ly}" x2="{pad_l+pw+30}" y2="{ly}" stroke="{col}" stroke-width="3"/>')
out.append(f'<text x="{pad_l+pw+34}" y="{ly+4}" class="leg">{name}</text>')
out.append("</svg>")
return "".join(out)
# ---------- build sections ----------
ladder = []
for b in BENCH:
best = max((macro(m, b, "ndcg@10") for m in MODELS if macro(m, b, "ndcg@10") is not None), default=None)
ladder.append((b, best))
mc_curves = {m: {k: round(v, 3) for k, v in strat(m, "multiconir", "n_cond").items()} for m in MODELS}
mc_curves = {m: s for m, s in mc_curves.items() if s}
merit_curves = {m: strat(m, "merit", "n_cond") for m in ["vl-2B", "vl-8B"]}
merit_curves = {m: s for m, s in merit_curves.items() if s}
ssrb_curves = {m: strat(m, "ssrb_subset", "nf") for m in MODELS}
ssrb_curves = {m: s for m, s in ssrb_curves.items() if s}
scaling_rows = []
for b in BENCH:
a, c = macro("vl-2B", b, "ndcg@10"), macro("vl-8B", b, "ndcg@10")
if a and c:
scaling_rows.append((b, a, c, (c - a) / a * 100))
html = f"""<!DOCTYPE html>
<html lang="zh"><head><meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>多条件检索统一评测 · 结果报告</title>
<style>
:root{{--bg:#0e1117;--card:#161b22;--ink:#e6edf3;--mut:#8b949e;--line:#30363d;--acc:#58a6ff;}}
*{{box-sizing:border-box}}
body{{margin:0;background:var(--bg);color:var(--ink);font:15px/1.6 -apple-system,'Segoe UI',Roboto,'PingFang SC','Microsoft YaHei',sans-serif}}
.wrap{{max-width:1040px;margin:0 auto;padding:32px 20px 80px}}
h1{{font-size:26px;margin:0 0 4px}} h2{{font-size:19px;margin:34px 0 12px;padding-bottom:6px;border-bottom:1px solid var(--line)}}
.sub{{color:var(--mut);margin:0 0 20px}}
.card{{background:var(--card);border:1px solid var(--line);border-radius:12px;padding:18px 20px;margin:14px 0}}
table{{border-collapse:collapse;width:100%;font-variant-numeric:tabular-nums}}
.mat th,.mat td{{padding:9px 10px;text-align:center;border:1px solid var(--line)}}
.mat th{{background:#1b2230;color:var(--ink);font-weight:600;font-size:13px}}
.mat td{{font-weight:600}} .mname{{text-align:left!important;background:#1b2230;color:var(--acc);font-family:ui-monospace,monospace}}
.nt{{display:block;font-size:10px;opacity:.7;font-weight:400}}
small{{color:var(--mut);font-weight:400}}
.kv{{padding:8px 10px;border:1px solid var(--line);text-align:center}}
.chart{{width:100%;height:auto;background:#0d1117;border-radius:8px;margin:6px 0}}
.ctitle{{fill:var(--ink);font-size:14px;font-weight:600}}
.grid{{stroke:#222b38;stroke-width:1}} .axlbl{{fill:var(--mut);font-size:11px}} .leg{{fill:var(--ink);font-size:11px;font-family:ui-monospace,monospace}}
.note{{color:var(--mut);font-size:13px}}
.find{{border-left:3px solid var(--acc);padding:4px 0 4px 14px;margin:12px 0}}
.grid2{{display:grid;grid-template-columns:1fr 1fr;gap:14px}} @media(max-width:760px){{.grid2{{grid-template-columns:1fr}}}}
.pill{{display:inline-block;background:#1b2230;border:1px solid var(--line);border-radius:999px;padding:2px 10px;font-size:12px;color:var(--mut);margin-right:6px}}
</style></head><body><div class="wrap">
<h1>多条件检索统一评测 · 结果报告</h1>
<p class="sub">框架 <code>mcr</code> · 4 benchmark × 5 模型(Qwen3-Embedding 0.6B/4B/8B · Qwen3-VL-Embedding 2B/8B)·
统一指标(自实现)· faiss 精确检索</p>
<div>
<span class="pill">MultiConIR 全量</span><span class="pill">MCMR 全量</span>
<span class="pill">MERIT 全量(修复多图)</span><span class="pill">SSRB 代表性子集 12 schema</span>
</div>
<h2>① 总览矩阵 · ndcg@10</h2>
<div class="card">{matrix_table('ndcg@10')}
<p class="note">单元格内小数字 = 该格聚合的 task 数(域/schema)。颜色按 <b>每个 benchmark 列</b>内相对高低着色(各 benchmark 量级差异极大,不可跨列比色)。</p></div>
<h2>② 难度阶梯(各 benchmark 最佳 ndcg@10)</h2>
<div class="card">
<table><tr>{"".join(f'<td class="kv"><div style="color:var(--mut);font-size:12px">{b.replace("_subset","").upper()}</div><div style="font-size:22px;font-weight:700;color:var(--acc)">{(v or 0):.3f}</div></td>' for b,v in ladder)}</tr></table>
<div class="find">纯文本多条件(MultiConIR ~0.74)已接近解决;一旦进入<b>跨模态(MCMR)→ 交错多图(MERIT)→ 半结构化大库(SSRB)</b>,性能跌一个数量级。</div>
</div>
<h2>③ 规模 × 模态:scaling 收益随模态复杂度放大</h2>
<div class="card">
<table class="mat"><thead><tr><th>benchmark</th><th>vl-2B</th><th>vl-8B</th><th>相对提升</th></tr></thead><tbody>
{"".join(f'<tr><td class="mname">{b.replace("_subset","")}</td><td>{a:.3f}</td><td>{c:.3f}</td><td style="color:{"#59a14f" if d>5 else "#8b949e"};font-weight:700">{d:+.0f}%</td></tr>' for b,a,c,d in scaling_rows)}
</tbody></table>
<div class="find">纯文本上堆参数几乎无用(MultiConIR ~0%);<b>跨模态、多图多条件上大模型才真正拉开差距</b>(MERIT +63%)。</div>
</div>
<h2>④ 组合性衰减:性能随条件数下降</h2>
<div class="grid2">
<div class="card">{svg_line(mc_curves, "MultiConIR · ndcg@10 vs 条件数", "条件数")}
<p class="note">全模型从 ~0.86(2 条件)单调跌到 ~0.47(10 条件),曲线高度重叠 → <b>与规模无关的普遍崩塌</b>。</p></div>
<div class="card">{svg_line(ssrb_curves, "SSRB · ndcg@10 vs 过滤条件数 nf", "过滤条件数 nf")}
<p class="note">过滤条件越多越差;<b>emb-4B 反超 emb-8B</b>(非单调 scaling)。</p></div>
</div>
<div class="card">{svg_line(merit_curves, "MERIT · ndcg@10 vs 条件数", "条件数", width=560, height=260)}
<p class="note">2 条件 0.14/0.22 → 3 条件骤降至 ~0.01。<b>注意</b>:3/4 条件样本极小(n=108 / n=5),该尾部统计上不稳。</p></div>
<h2>⑤ 其它指标</h2>
<div class="grid2">
<div class="card"><h3 style="margin:0 0 8px;font-size:15px">recall@10</h3>{matrix_table('recall@10')}</div>
<div class="card"><h3 style="margin:0 0 8px;font-size:15px">mrr@10</h3>{matrix_table('mrr@10')}</div>
</div>
<div class="card"><h3 style="margin:0 0 8px;font-size:15px">recall@100</h3>{matrix_table('recall@100')}</div>
<h2>⑥ 说明与 caveat</h2>
<div class="card note">
<ul>
<li><b>SSRB 为代表性子集</b>:6 域 × 各 2 schema = 12/99,非全量。</li>
<li><b>MERIT 多图已修复</b>:query 经 ST <code>message</code> 模态编码,全部参考图参与(早期"首图近似"结果已废弃)。</li>
<li><b>MERIT 高条件桶样本小</b>:2 条件 n≈9887 为主体;3/4 条件桶 n=108/5,尾部数值仅供参考。</li>
<li>统一 query 指令、统一 faiss 精确检索;未做模型专属调优。MAP 采用 trec_eval map_cut 约定。</li>
<li>指标自实现(不依赖 pytrec_eval),与手算/对拍一致。</li>
</ul>
</div>
</div></body></html>"""
with open(OUT, "w", encoding="utf-8") as f:
f.write(html)
print(f"wrote {OUT} ({len(html)} bytes)")