File size: 27,826 Bytes
e74a796 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 | #!/usr/bin/env python3
import json
from datetime import datetime
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
REPORT_DIR = ROOT / "reports"
OUTPUT = REPORT_DIR / "index.html"
def read_json(path, default=None):
try:
return json.loads(Path(path).read_text(encoding="utf-8"))
except Exception:
return default
def count_lines(path):
try:
with Path(path).open(encoding="utf-8") as f:
return sum(1 for _ in f)
except Exception:
return 0
def parse_log_status():
monitor = ROOT / "outputs/logs/training_monitor.log"
final_files = [
ROOT / "outputs/metrics/finetuned_struct_metrics.json",
ROOT / "outputs/metrics/finetuned_qa_metrics.json",
ROOT / "outputs/figures/metric_comparison.csv",
]
if monitor.exists():
log = monitor
else:
logs = sorted((ROOT / "outputs/logs").glob("*.log"), key=lambda p: p.stat().st_mtime, reverse=True)
if not logs:
return {"log_file": None, "tail": [], "stage": "not_started"}
log = logs[0]
tail = log.read_text(encoding="utf-8", errors="ignore").splitlines()[-120:]
joined = "\n".join(tail)
if all(p.exists() for p in final_files):
stage = "completed"
elif "scripts/train_qlora.py" in joined:
stage = "training"
elif "--run-name finetuned" in joined:
stage = "finetuned_eval"
elif "Traceback" in joined or "Error" in joined:
stage = "needs_attention"
else:
stage = "running"
return {"log_file": str(log.relative_to(ROOT)), "tail": tail[-20:], "stage": stage}
def metrics_from_prediction_file(path, limit=None):
required = [
"current_behavior",
"is_transition",
"elapsed_seconds_in_current_behavior",
"estimated_remaining_seconds",
"full_remaining_seconds",
"expected_end_time",
"next_possible_behavior",
"stage_index",
"total_stages",
"sequence_so_far",
]
rows = []
try:
with Path(path).open(encoding="utf-8") as f:
for line in f:
if line.strip():
rows.append(json.loads(line))
if limit and len(rows) >= limit:
break
except Exception:
return {}
if not rows:
return {}
parsed = [r for r in rows if isinstance(r.get("prediction"), dict)]
def acc(field):
pairs = [(r["target"].get(field), r["prediction"].get(field)) for r in parsed if field in r["prediction"]]
return sum(a == b for a, b in pairs) / len(pairs) if pairs else 0
def mae(field):
pairs = []
for r in parsed:
p = r.get("prediction", {})
t = r.get("target", {})
if isinstance(t.get(field), (int, float)) and isinstance(p.get(field), (int, float)):
pairs.append(abs(float(t[field]) - float(p[field])))
return sum(pairs) / len(pairs) if pairs else None
return {
"num_examples": len(rows),
"json_parse_rate": len(parsed) / len(rows),
"required_field_complete_rate": sum(all(f in r["prediction"] for f in required) for r in parsed) / len(rows),
"current_behavior_accuracy": acc("current_behavior"),
"next_possible_behavior_accuracy": acc("next_possible_behavior"),
"is_transition_accuracy": acc("is_transition"),
"stage_index_accuracy": acc("stage_index"),
"full_remaining_seconds_mae": mae("full_remaining_seconds"),
}
def load_metrics():
metric_dir = ROOT / "outputs/metrics"
metrics = {}
for path in sorted(metric_dir.glob("*.json")):
payload = read_json(path)
if not payload:
continue
key = f"{payload.get('run_name')}_{payload.get('task_type')}"
metrics[key] = payload
partial = metrics_from_prediction_file(ROOT / "outputs/predictions/base_struct_predictions.jsonl")
if partial:
metrics["base_struct_partial"] = {
"run_name": "base_partial",
"task_type": "struct",
"input_file": "outputs/predictions/base_struct_predictions.jsonl",
"metrics": partial,
}
return metrics
def build_data():
summary = read_json(ROOT / "data/processed/summary.json", {})
progress = {
"base_struct_done": count_lines(ROOT / "outputs/predictions/base_struct_predictions.jsonl"),
"base_qa_done": count_lines(ROOT / "outputs/predictions/base_qa_predictions.jsonl"),
"finetuned_struct_done": count_lines(ROOT / "outputs/predictions/finetuned_struct_predictions.jsonl"),
"finetuned_qa_done": count_lines(ROOT / "outputs/predictions/finetuned_qa_predictions.jsonl"),
"val_total": summary.get("val_struct", {}).get("num_examples", 4030),
}
return {
"generated_at": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
"summary": summary,
"metrics": load_metrics(),
"progress": progress,
"log_status": parse_log_status(),
"files": {
"processed_summary": "data/processed/summary.json",
"base_predictions": "outputs/predictions/base_struct_predictions.jsonl",
"adapter_dir": "outputs/qwen35_9b_lora",
"figures_dir": "outputs/figures",
},
}
HTML_TEMPLATE = r"""<!doctype html>
<html lang="zh-CN">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>MWave Aircraft Lavatory Radar LLM Workflow</title>
<style>
:root{
--boeing:#0039A6;
--boeing-2:#0067B1;
--red:#C8102E;
--black:#080A0F;
--ink:#111827;
--steel:#5B6472;
--mist:#E7ECF3;
--panel:#F7F9FC;
--white:#FFFFFF;
--line:rgba(8,10,15,.14);
--shadow:0 24px 80px rgba(0,23,84,.16);
}
*{box-sizing:border-box}
html{scroll-behavior:smooth}
body{
margin:0;
color:var(--ink);
font-family:"Aptos Display","Bahnschrift","DIN Alternate","Noto Sans SC",sans-serif;
background:
radial-gradient(circle at 12% 8%, rgba(0,103,177,.22), transparent 30%),
radial-gradient(circle at 82% 18%, rgba(200,16,46,.16), transparent 26%),
linear-gradient(135deg,#f3f6fb 0%,#ffffff 42%,#e9eef6 100%);
}
.shell{max-width:1440px;margin:0 auto;padding:28px}
.hero{
position:relative;
min-height:520px;
border-radius:34px;
overflow:hidden;
color:white;
background:
linear-gradient(115deg,rgba(3,11,29,.96),rgba(0,57,166,.92) 48%,rgba(0,103,177,.72)),
repeating-linear-gradient(90deg,transparent 0 72px,rgba(255,255,255,.06) 72px 73px);
box-shadow:var(--shadow);
padding:42px;
}
.hero:after{
content:"";
position:absolute;
right:-160px;top:-110px;
width:620px;height:620px;
border:1px solid rgba(255,255,255,.18);
border-radius:50%;
box-shadow:inset 0 0 0 72px rgba(255,255,255,.04), inset 0 0 0 140px rgba(255,255,255,.035);
}
.kicker{display:inline-flex;gap:10px;align-items:center;letter-spacing:.16em;text-transform:uppercase;font-size:12px;color:#dbeafe}
.pulse{width:10px;height:10px;background:var(--red);border-radius:50%;box-shadow:0 0 0 10px rgba(200,16,46,.18)}
h1{font-size:clamp(42px,7vw,96px);line-height:.9;margin:28px 0 20px;letter-spacing:-.06em;max-width:980px}
.hero p{font-size:20px;line-height:1.65;max-width:920px;color:#e8eef8;margin:0}
.hero-grid{position:relative;z-index:1;display:grid;grid-template-columns:1.2fr .8fr;gap:24px;margin-top:38px}
.hero-card{background:rgba(255,255,255,.1);border:1px solid rgba(255,255,255,.2);border-radius:24px;padding:22px;backdrop-filter:blur(18px)}
.stat{font-size:38px;font-weight:800;letter-spacing:-.04em}
.stat-label{color:#cbd5e1;font-size:13px;margin-top:6px}
nav{position:sticky;top:0;z-index:10;margin:18px 0;padding:12px;border-radius:22px;background:rgba(255,255,255,.82);backdrop-filter:blur(18px);border:1px solid var(--line);display:flex;gap:8px;flex-wrap:wrap}
nav button,.tab-btn{
border:0;border-radius:999px;padding:12px 16px;background:#eef3fb;color:#1f2937;font-weight:800;cursor:pointer;transition:.18s ease;
}
nav button:hover,.tab-btn:hover,.tab-btn.active{background:var(--boeing);color:white;transform:translateY(-1px)}
section{margin-top:24px}
.panel{background:rgba(255,255,255,.9);border:1px solid var(--line);border-radius:28px;padding:28px;box-shadow:0 16px 44px rgba(17,24,39,.06)}
.section-title{display:flex;align-items:end;justify-content:space-between;gap:18px;margin-bottom:20px}
h2{font-size:34px;margin:0;letter-spacing:-.04em;color:var(--black)}
.muted{color:var(--steel);line-height:1.65}
.grid{display:grid;gap:18px}
.grid-4{grid-template-columns:repeat(4,minmax(0,1fr))}
.grid-3{grid-template-columns:repeat(3,minmax(0,1fr))}
.grid-2{grid-template-columns:repeat(2,minmax(0,1fr))}
.metric-card{border-radius:22px;padding:20px;background:linear-gradient(180deg,#fff,#f5f8fd);border:1px solid var(--line);position:relative;overflow:hidden}
.metric-card:before{content:"";position:absolute;left:0;top:0;width:5px;height:100%;background:var(--boeing)}
.metric-card.red:before{background:var(--red)}
.big{font-size:32px;font-weight:900;letter-spacing:-.04em}
.timeline{display:grid;grid-template-columns:repeat(6,1fr);gap:12px;margin-top:18px}
.step{position:relative;border-radius:22px;padding:18px;background:#f8fafc;border:1px solid var(--line);min-height:160px}
.step b{display:block;color:var(--boeing);font-size:18px;margin-bottom:8px}
.step .num{width:34px;height:34px;border-radius:50%;display:grid;place-items:center;background:var(--black);color:white;font-weight:900;margin-bottom:16px}
.step.active{background:linear-gradient(180deg,#fff,#eaf3ff);box-shadow:0 12px 36px rgba(0,57,166,.14)}
.chart-card{background:#fff;border:1px solid var(--line);border-radius:24px;padding:20px;min-height:360px}
canvas{width:100%;height:300px}
.tabs{display:flex;gap:8px;flex-wrap:wrap;margin-bottom:14px}
.tab{display:none}.tab.active{display:block}
table{width:100%;border-collapse:collapse;background:#fff;border-radius:18px;overflow:hidden}
th,td{text-align:left;padding:13px 14px;border-bottom:1px solid #e5e7eb}
th{background:#0b1220;color:#fff;font-size:13px}
tr:hover td{background:#f8fbff}
.badge{display:inline-flex;align-items:center;gap:8px;border-radius:999px;padding:8px 12px;background:#eef3fb;color:var(--boeing);font-weight:900;font-size:12px}
.badge.red{background:#fff0f2;color:var(--red)}
.architecture{display:grid;grid-template-columns:1fr 90px 1fr 90px 1fr;gap:16px;align-items:center}
.arch-box{padding:22px;border-radius:24px;border:1px solid var(--line);background:#fff;min-height:180px}
.arrow{height:2px;background:linear-gradient(90deg,var(--boeing),var(--red));position:relative}
.arrow:after{content:"";position:absolute;right:-2px;top:-5px;border-left:12px solid var(--red);border-top:6px solid transparent;border-bottom:6px solid transparent}
.risk{display:grid;grid-template-columns:repeat(2,1fr);gap:12px}
.risk div{border-radius:18px;padding:16px;background:#f8fafc;border:1px solid var(--line)}
.risk b{color:var(--red)}
.terminal{background:#05070c;color:#d1fae5;border-radius:22px;padding:18px;font-family:"Cascadia Mono","SFMono-Regular",monospace;overflow:auto;max-height:330px;font-size:13px}
.footer{padding:28px;color:#64748b;text-align:center}
@media(max-width:980px){
.hero-grid,.grid-4,.grid-3,.grid-2,.timeline,.architecture,.risk{grid-template-columns:1fr}
.arrow{height:42px;width:2px;margin:auto;background:linear-gradient(180deg,var(--boeing),var(--red))}
.arrow:after{right:-5px;top:auto;bottom:-2px;border-top:12px solid var(--red);border-left:6px solid transparent;border-right:6px solid transparent;border-bottom:0}
.shell{padding:14px}.hero{padding:26px}
}
</style>
</head>
<body>
<div class="shell">
<header class="hero">
<div class="kicker"><span class="pulse"></span> MWave Radar LLM · Aircraft Lavatory Intelligence</div>
<h1>毫米波雷达时序行为模型工作流报告</h1>
<p>从飞机厕所内毫米波雷达轨迹窗口和中间层表征出发，微调 Qwen3.5-9B，完成结构化行为预测、序列重建、剩余时间估计和 QA 状态问答，并支持部署到航空场景的边缘推理方案。</p>
<div class="hero-grid">
<div class="hero-card">
<div class="grid grid-4">
<div><div class="stat" id="heroTrain">-</div><div class="stat-label">训练结构样本</div></div>
<div><div class="stat" id="heroVal">-</div><div class="stat-label">验证结构样本</div></div>
<div><div class="stat" id="heroQA">-</div><div class="stat-label">QA 训练样本</div></div>
<div><div class="stat" id="heroProgress">-</div><div class="stat-label">整体评估进度</div></div>
</div>
</div>
<div class="hero-card">
<span class="badge red" id="runStage">RUNNING</span>
<p style="margin-top:14px;font-size:15px">报告生成时间：<b id="generatedAt"></b><br>模型：Qwen/Qwen3.5-9B + 4-bit QLoRA<br>输出：JSON schema + QA schema + charts</p>
</div>
</div>
</header>
<nav>
<button onclick="go('workflow')">Workflow</button>
<button onclick="go('data')">数据画像</button>
<button onclick="go('results')">评估结果</button>
<button onclick="go('analysis')">发现与风险</button>
<button onclick="go('deployment')">机载部署方案</button>
<button onclick="go('ops')">运行状态</button>
</nav>
<section id="workflow" class="panel">
<div class="section-title"><h2>端到端 Workflow</h2><span class="badge">Train · Evaluate · Deploy</span></div>
<div class="timeline">
<div class="step active"><div class="num">1</div><b>数据解析</b><span class="muted">读取 train/val JSONL，解析 chat 格式、轨迹窗口、背景知识和 assistant JSON。</span></div>
<div class="step active"><div class="num">2</div><b>标签规范</b><span class="muted">统一 `反复折返 -> 折返`，固定结构化输出字段，生成 QA 目标。</span></div>
<div class="step active"><div class="num">3</div><b>基线评估</b><span class="muted">Qwen3.5-9B base 先跑 val，保存微调前预测和指标。</span></div>
<div class="step active"><div class="num">4</div><b>QLoRA 微调</b><span class="muted">结构化预测与 QA 混合 SFT，4-bit NF4，LoRA 训练约 29M 参数。</span></div>
<div class="step active"><div class="num">5</div><b>微调后评估</b><span class="muted">同一套 val、同一套指标，输出 finetuned 预测、指标和图表。</span></div>
<div class="step active"><div class="num">6</div><b>机载部署</b><span class="muted">毫米波雷达边缘预处理 + 本地 LLM 推理 + 客舱系统状态输出。</span></div>
</div>
</section>
<section id="data" class="panel">
<div class="section-title"><h2>数据画像</h2><div class="tabs"><button class="tab-btn active" onclick="switchTab('dist','train',this)">Train</button><button class="tab-btn" onclick="switchTab('dist','val',this)">Val</button></div></div>
<div class="grid grid-2">
<div class="chart-card">
<h3>行为标签分布</h3>
<canvas id="labelChart"></canvas>
</div>
<div class="chart-card">
<h3>任务样本构成</h3>
<canvas id="taskChart"></canvas>
</div>
</div>
</section>
<section id="results" class="panel">
<div class="section-title"><h2>可视化结果</h2><span class="badge red">Base vs Finetuned</span></div>
<div class="grid grid-4" id="metricCards"></div>
<div class="grid grid-2" style="margin-top:18px">
<div class="chart-card"><h3>结构化指标</h3><canvas id="structMetricChart"></canvas></div>
<div class="chart-card"><h3>QA 指标</h3><canvas id="qaMetricChart"></canvas></div>
</div>
<div style="margin-top:18px">
<table id="metricTable"></table>
</div>
</section>
<section id="analysis" class="panel">
<div class="section-title"><h2>总结与发现</h2><span class="badge">Current Findings</span></div>
<div class="grid grid-3">
<div class="metric-card"><div class="big">1</div><b>base 模型不等于任务模型</b><p class="muted">未微调 Qwen3.5-9B 能理解中文指令，但结构化 schema 稳定性不足，尤其容易输出解释性文本或缺字段。</p></div>
<div class="metric-card red"><div class="big">2</div><b>QA 必须独立评估</b><p class="muted">QA 不是简单复述，需要从 `full_remaining_seconds`、`sequence_so_far` 和异常规则推导，占用、空出时间、区域使用和异常应单独打分。</p></div>
<div class="metric-card"><div class="big">3</div><b>短行为是关键风险</b><p class="muted">进入、门锁、坐下、起身、折返、犹豫等短时行为占比较低，但对流程阶段和剩余时间预测影响很大。</p></div>
</div>
<div class="risk" style="margin-top:18px">
<div><b>数据风险</b><p class="muted">训练集中坐用马桶占比最高，类别不平衡明显；若最终少数类 F1 低，需要重采样或 loss 权重。</p></div>
<div><b>工程风险</b><p class="muted">长 prompt 和 9B 推理导致全量评估耗时较长；部署时应缓存背景知识、压缩轨迹窗口并使用约束解码。</p></div>
<div><b>安全边界</b><p class="muted">系统只输出状态和行为，不输出身份识别，不存储原始可逆人体点云，降低隐私风险。</p></div>
<div><b>上线策略</b><p class="muted">先 shadow mode 与人工规则并行，确认误报/漏报边界，再进入客舱状态提示闭环。</p></div>
</div>
</section>
<section id="deployment" class="panel">
<div class="section-title"><h2>飞机厕所部署技术方案</h2><span class="badge">Edge-first · Privacy-preserving</span></div>
<div class="architecture">
<div class="arch-box"><h3>1. 传感与预处理</h3><p class="muted">毫米波雷达采集点云/轨迹；本地 MCU/边缘 SoC 做去噪、目标跟踪、窗口化、速度和区域特征提取。</p></div>
<div class="arrow"></div>
<div class="arch-box"><h3>2. 行为 LLM 推理</h3><p class="muted">Qwen3.5-9B LoRA/量化模型接收结构化窗口，输出严格 JSON：当前行为、阶段、剩余时间、序列。</p></div>
<div class="arrow"></div>
<div class="arch-box"><h3>3. 客舱系统集成</h3><p class="muted">输出占用、预计空出、异常、已使用区域。对接乘务终端、维护日志、客舱状态总线。</p></div>
</div>
<div class="grid grid-3" style="margin-top:18px">
<div class="metric-card"><b>边缘硬件</b><p class="muted">训练在地面 GPU；机载推理建议使用小型边缘 GPU/NPU 或将 LLM 部署在客舱边缘计算单元，厕所侧只传结构化特征。</p></div>
<div class="metric-card"><b>实时策略</b><p class="muted">每 0.5-1 秒更新窗口；稳定行为可降频推理，过渡态或异常候选升频推理。</p></div>
<div class="metric-card red"><b>失效保护</b><p class="muted">JSON 校验失败时回退规则模型；连续异常时只提示“需关注”，不直接做强制控制决策。</p></div>
</div>
</section>
<section id="ops" class="panel">
<div class="section-title"><h2>运行状态与文件</h2><span class="badge" id="opsBadge">Live</span></div>
<div class="grid grid-4">
<div class="metric-card"><div class="big" id="baseStructDone">0</div><b>base struct 已生成</b></div>
<div class="metric-card"><div class="big" id="baseQaDone">0</div><b>base QA 已生成</b></div>
<div class="metric-card"><div class="big" id="ftStructDone">0</div><b>finetuned struct 已生成</b></div>
<div class="metric-card"><div class="big" id="ftQaDone">0</div><b>finetuned QA 已生成</b></div>
</div>
<h3>最近日志</h3>
<pre class="terminal" id="logTail"></pre>
</section>
<div class="footer">MWave report · Boeing blue / red / black / gray visual system · generated locally</div>
</div>
<script id="report-data" type="application/json">__DATA__</script>
<script>
const DATA = JSON.parse(document.getElementById('report-data').textContent);
let distMode = 'train';
const colors = ['#0039A6','#0067B1','#C8102E','#111827','#5B6472','#8EA4C8','#B7C3D5','#7F1D1D','#334155','#64748B'];
function fmt(n){ if(n===null||n===undefined) return 'N/A'; if(typeof n==='number'){ return n>=1000 ? n.toLocaleString() : (Math.round(n*1000)/1000).toString(); } return n; }
function pct(n){ return n===null||n===undefined ? 'N/A' : Math.round(n*1000)/10 + '%'; }
function go(id){ document.getElementById(id).scrollIntoView({behavior:'smooth'}); }
function switchTab(group, mode, btn){ distMode=mode; document.querySelectorAll('.tab-btn').forEach(b=>b.classList.remove('active')); btn.classList.add('active'); drawAll(); }
function drawBar(canvas, labels, values, opts={}){
const ctx = canvas.getContext('2d'); const dpr = window.devicePixelRatio || 1;
const rect = canvas.getBoundingClientRect(); canvas.width = rect.width*dpr; canvas.height = rect.height*dpr; ctx.scale(dpr,dpr);
const w=rect.width,h=rect.height,pad=42; ctx.clearRect(0,0,w,h);
const max=Math.max(...values,1); const bw=(w-pad*2)/values.length*0.68;
ctx.strokeStyle='#E5E7EB'; ctx.lineWidth=1;
for(let i=0;i<5;i++){ const y=pad+(h-pad*2)*i/4; ctx.beginPath();ctx.moveTo(pad,y);ctx.lineTo(w-pad,y);ctx.stroke(); }
values.forEach((v,i)=>{ const x=pad+(w-pad*2)*i/values.length+(w-pad*2)/values.length*.16; const bh=(h-pad*2)*v/max; const y=h-pad-bh;
const grad=ctx.createLinearGradient(0,y,0,h-pad); grad.addColorStop(0, colors[i%colors.length]); grad.addColorStop(1,'rgba(0,57,166,.18)');
ctx.fillStyle=grad; roundRect(ctx,x,y,bw,bh,8); ctx.fill();
ctx.save(); ctx.translate(x+bw/2,h-pad+8); ctx.rotate(-Math.PI/4); ctx.fillStyle='#334155'; ctx.font='11px Bahnschrift, sans-serif'; ctx.fillText(labels[i],0,0); ctx.restore();
});
ctx.fillStyle='#111827'; ctx.font='12px Bahnschrift, sans-serif'; ctx.fillText(opts.yLabel||'',8,18);
}
function drawDonut(canvas, labels, values){
const ctx=canvas.getContext('2d'); const dpr=window.devicePixelRatio||1; const rect=canvas.getBoundingClientRect(); canvas.width=rect.width*dpr; canvas.height=rect.height*dpr; ctx.scale(dpr,dpr);
const w=rect.width,h=rect.height,cx=w*.34,cy=h*.48,r=Math.min(w,h)*.27; ctx.clearRect(0,0,w,h); const total=values.reduce((a,b)=>a+b,0)||1; let start=-Math.PI/2;
values.forEach((v,i)=>{ const end=start+Math.PI*2*v/total; ctx.beginPath(); ctx.moveTo(cx,cy); ctx.arc(cx,cy,r,start,end); ctx.closePath(); ctx.fillStyle=colors[i%colors.length]; ctx.fill(); start=end; });
ctx.beginPath(); ctx.arc(cx,cy,r*.58,0,Math.PI*2); ctx.fillStyle='#fff'; ctx.fill(); ctx.fillStyle='#111827'; ctx.font='900 24px Bahnschrift'; ctx.textAlign='center'; ctx.fillText(total.toLocaleString(),cx,cy+8);
ctx.textAlign='left'; labels.forEach((l,i)=>{ const y=50+i*28; ctx.fillStyle=colors[i%colors.length]; ctx.fillRect(w*.62,y-10,12,12); ctx.fillStyle='#334155'; ctx.font='13px Bahnschrift'; ctx.fillText(`${l}: ${values[i].toLocaleString()}`,w*.62+20,y); });
}
function roundRect(ctx,x,y,w,h,r){ ctx.beginPath(); ctx.moveTo(x+r,y); ctx.arcTo(x+w,y,x+w,y+h,r); ctx.arcTo(x+w,y+h,x,y+h,r); ctx.arcTo(x,y+h,x,y,r); ctx.arcTo(x,y,x+w,y,r); ctx.closePath(); }
function initHero(){
const s=DATA.summary||{}, p=DATA.progress||{}; const val=p.val_total||1;
document.getElementById('heroTrain').textContent=fmt(s.train_struct?.num_examples);
document.getElementById('heroVal').textContent=fmt(s.val_struct?.num_examples);
document.getElementById('heroQA').textContent=fmt(s.train_qa?.num_examples);
const done=(p.base_struct_done||0)+(p.base_qa_done||0)+(p.finetuned_struct_done||0)+(p.finetuned_qa_done||0);
document.getElementById('heroProgress').textContent=pct(done/(val*4));
document.getElementById('generatedAt').textContent=DATA.generated_at;
document.getElementById('runStage').textContent=(DATA.log_status?.stage||'unknown').toUpperCase();
document.getElementById('baseStructDone').textContent=fmt(p.base_struct_done);
document.getElementById('baseQaDone').textContent=fmt(p.base_qa_done);
document.getElementById('ftStructDone').textContent=fmt(p.finetuned_struct_done);
document.getElementById('ftQaDone').textContent=fmt(p.finetuned_qa_done);
document.getElementById('logTail').textContent=(DATA.log_status?.tail||[]).join('\n');
}
function drawData(){
const counts = DATA.summary?.[distMode+'_struct']?.label_counts || {};
const entries = Object.entries(counts).sort((a,b)=>b[1]-a[1]).slice(0,12);
drawBar(document.getElementById('labelChart'), entries.map(x=>x[0]), entries.map(x=>x[1]), {yLabel:'Top labels'});
const s=DATA.summary||{}; const vals = distMode==='train' ? [s.train_struct?.num_examples||0, s.train_qa?.num_examples||0] : [s.val_struct?.num_examples||0, s.val_qa?.num_examples||0];
drawDonut(document.getElementById('taskChart'), ['结构化预测','QA 问答'], vals);
}
function metricRows(task){
const out=[]; Object.values(DATA.metrics||{}).forEach(p=>{ if(p.task_type===task){ Object.entries(p.metrics||{}).forEach(([k,v])=>{ if(typeof v==='number') out.push({run:p.run_name, metric:k, value:v}); }); }}); return out;
}
function drawMetrics(){
const struct = metricRows('struct').filter(x=>['json_parse_rate','required_field_complete_rate','current_behavior_accuracy','next_possible_behavior_accuracy','is_transition_accuracy','stage_index_accuracy'].includes(x.metric));
const qa = metricRows('qa').filter(x=>['json_parse_rate','required_field_complete_rate','occupied_accuracy','is_abnormal_accuracy','used_areas_micro_f1'].includes(x.metric));
drawBar(document.getElementById('structMetricChart'), struct.map(x=>x.run+'\\n'+x.metric.replaceAll('_',' ')), struct.map(x=>x.value), {yLabel:'score'});
drawBar(document.getElementById('qaMetricChart'), qa.map(x=>x.run+'\\n'+x.metric.replaceAll('_',' ')), qa.map(x=>x.value), {yLabel:'score'});
const keyMetrics=[
['Struct behavior acc', DATA.metrics?.finetuned_struct?.metrics?.current_behavior_accuracy, ''],
['Struct schema complete', DATA.metrics?.finetuned_struct?.metrics?.required_field_complete_rate, ''],
['QA abnormal F1', DATA.metrics?.finetuned_qa?.metrics?.is_abnormal_f1, 'red'],
['QA used areas F1', DATA.metrics?.finetuned_qa?.metrics?.used_areas_micro_f1, '']
];
document.getElementById('metricCards').innerHTML=keyMetrics.map(([name,val,klass])=>`<div class="metric-card ${klass}"><div class="big">${typeof val==='number'?pct(val):'N/A'}</div><b>${name}</b></div>`).join('');
const all=[]; Object.values(DATA.metrics||{}).forEach(p=>{ Object.entries(p.metrics||{}).forEach(([k,v])=>all.push([p.run_name,p.task_type,k,v])); });
document.getElementById('metricTable').innerHTML='<tr><th>Run</th><th>Task</th><th>Metric</th><th>Value</th></tr>'+all.map(r=>`<tr><td>${r[0]}</td><td>${r[1]}</td><td>${r[2]}</td><td>${fmt(r[3])}</td></tr>`).join('');
}
function drawAll(){ initHero(); drawData(); drawMetrics(); }
window.addEventListener('resize', drawAll);
drawAll();
</script>
</body>
</html>
"""
def main():
REPORT_DIR.mkdir(parents=True, exist_ok=True)
data = build_data()
html = HTML_TEMPLATE.replace("__DATA__", json.dumps(data, ensure_ascii=False))
OUTPUT.write_text(html, encoding="utf-8")
print(f"wrote {OUTPUT}")
if __name__ == "__main__":
main()
|