code/make_report.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Build a single self-contained HTML report of the benchmark.
4 modules (Myopia / AMD / Glaucoma / DR). Each dataset card shows:
  - 采集背景 (acquisition background: FOV / device / source / resolution)
  - 类别分布 (class distribution by split: table + grouped bar chart)
  - 模型性能 (metrics table + grouped bar chart, 3 models)
  - 可展开的混淆矩阵 / ROC 图
All images are embedded as base64 -> works offline, one file.
"""
import os, json, base64, io, csv
from collections import defaultdict
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

PROJ = "/mnt/tidal-alsh-share2/dataset/qinshengqian/research/c3/GPT-Image"
RESULTS = f"{PROJ}/results"
DSROOT = f"{PROJ}/Dataset"
OUT = f"{RESULTS}/report.html"

MODELS = ["retfound", "resnet", "vit"]
MLABEL = {"retfound": "RetFound (ViT-L, CFP)", "resnet": "ResNet-50", "vit": "ViT-B/16"}
MCOLOR = {"retfound": "#4C72B0", "resnet": "#55A868", "vit": "#C44E52"}
SPLIT_COLOR = {"train": "#4C72B0", "val": "#DD8452", "test": "#55A868"}

CATS = [
    ("近视 · Myopic Maculopathy", "#2563eb", [("mmac", "MMAC 2023", "5-class grade 0–4")]),
    ("AMD · Age-related Macular Degeneration", "#16a34a", [("adam", "ADAM", "binary AMD / Non-AMD")]),
    ("青光眼 · Glaucoma", "#d97706", [("airogs", "AIROGS (EyePACS-AIROGS-light)", "binary RG / NRG"),
                                       ("papila", "PAPILA", "binary glaucoma / healthy")]),
    ("DR · Diabetic Retinopathy", "#dc2626", [("idrid", "IDRiD", "5-class grade 0–4"),
                                              ("aptos", "APTOS-2019", "5-class grade 0–4"),
                                              ("deepdrid", "DeepDRiD", "5-class grade 0–4")]),
]
SPLITS = {"mmac": "973/139/279", "adam": "280/40/80", "airogs": "5000/540/1000",
          "papila": "294/42/84", "idrid": "318/45/92", "aptos": "2930/366/366",
          "deepdrid": "1200/400/400"}
DSPATH = {"mmac": "Myopia/Classification_of_Myopic_Maculopathy", "adam": "AMD/adamdataset",
          "airogs": "Glaucoma/eyepacs-airogs-light", "papila": "Glaucoma/papila-retinal-fundus-images",
          "idrid": "DR/idrid-dataset", "aptos": "DR/aptos2019", "deepdrid": "DR/deepdrid"}

# 采集背景（已核实来源；不确定处据实标注）
BG = {
 "mmac": "彩色眼底照（非散瞳）｜FOV：45°（设备标称，论文正文未印）｜设备：Topcon TRC-NW400（单一设备）｜来源：上海健康医学中心 + 上海市第六人民医院（中国，均为中国人群）｜分辨率：未公开｜标注：META-PM 5 级，双医师分级（κ=0.91），单设备单人群为其局限。",
 "adam": "彩色眼底照｜FOV：未标注（仅说明取景中心为视盘 / 黄斑 / 两者中点）｜设备：Zeiss Visucam 500（2124×2056，824 张）+ Canon CR-2（1444×1444，376 张）｜来源：中山眼科中心（中国·广州）｜Training400：89 AMD / 311 非 AMD（AMD 被刻意过采样，非真实患病率）。",
 "airogs": "彩色眼底照，源自 EyePACS 远程筛查平台（美国约 500 个点、60071 人、多种族）｜设备：多相机混用（Optovue iCam100≈26%、Topcon NW200/400≈20%、Canon CR1/CR2/DGI、Centervue、Nidek、Crystalvue，约 21% 未知）｜FOV / 分辨率：因多设备未统一｜原为糖网筛查图后重标青光眼；全集 RG 仅约 3%（极不平衡），本「light」子集已平衡为 3270/3270。",
 "papila": "彩色眼底照，以视盘为中心｜FOV：30°｜设备：Topcon TRC-NW400（非散瞳）｜分辨率：2576×1934 JPEG｜来源：Reina Sofía 大学医院（西班牙·Murcia，2018–2020）｜244 人双眼共 488 张（healthy/glaucoma/suspect，本项目已剔除 suspect → 420）｜附临床数据与视盘/视杯分割。",
 "idrid": "彩色眼底照｜FOV：50°｜设备：Kowa VX-10α（散瞳，托吡卡胺 0.5%）｜分辨率：4288×2848 JPG｜来源：印度 Nanded（Maharashtra）眼科诊所，2009–2017｜全集 516 张（本项目有标签 455 张）｜DR 0–4（ICDR）+ 黄斑水肿风险分级。",
 "aptos": "彩色眼底照｜设备 / FOV / 分辨率：均未公开（多诊所、多相机、跨时间采集，异质性大）｜来源：Aravind 眼科医院（印度），乡村远程筛查｜训练集 3662 张，DR 0–4（ICDR）｜真实世界噪声明显（伪影 / 失焦 / 过曝欠曝 / 标签噪声）。",
 "deepdrid": "彩色眼底照（常规，非超广角）｜设备：Topcon 非散瞳（具体型号未公开）｜FOV≈45–60°、分辨率≈1956×1934（来自补充材料，中等可信）｜来源：上海市第六人民医院（中国）糖尿病筛查队列｜2000 张 / 500 人，每眼双视野（视盘中心 + 黄斑中心）｜DR 0–4 + 图像质量标注。",
}

BIN_COLS = [("accuracy", "Accuracy"), ("auroc", "AUROC"), ("auprc", "AUPRC"), ("f1_macro", "F1"),
            ("sensitivity", "Sensitivity"), ("specificity", "Specificity"), ("cohen_kappa", "Kappa"), ("mcc", "MCC")]
MUL_COLS = [("accuracy", "Accuracy"), ("balanced_accuracy", "Bal-Acc"), ("auroc_macro_ovr", "macro-AUROC"),
            ("quadratic_weighted_kappa", "QWK"), ("f1_macro", "F1-macro"), ("precision_macro", "Prec-macro"),
            ("recall_macro", "Rec-macro"), ("cohen_kappa", "Kappa")]
BIN_BAR = [("accuracy", "Acc"), ("auroc", "AUROC"), ("auprc", "AUPRC"), ("f1_macro", "F1"),
           ("sensitivity", "Sens"), ("specificity", "Spec")]
MUL_BAR = [("accuracy", "Acc"), ("auroc_macro_ovr", "AUROC"), ("quadratic_weighted_kappa", "QWK"),
           ("f1_macro", "F1"), ("balanced_accuracy", "Bal-Acc")]


def load(dsk, model):
    try:
        return json.load(open(os.path.join(RESULTS, dsk, model, "metrics.json")))
    except Exception:
        return None


def read_dist(dsk):
    """Return ordered [(label,class_name)], counts[split][label], total."""
    rows = list(csv.DictReader(open(os.path.join(DSROOT, DSPATH[dsk], "labels.csv"))))
    classes = sorted(set((r["label"], r.get("class_name", "")) for r in rows), key=lambda x: int(x[0]))
    cnt = defaultdict(lambda: defaultdict(int))
    for r in rows:
        cnt[r["split"]][r["label"]] += 1
    return classes, cnt, len(rows)


def b64_img(path):
    try:
        return "data:image/png;base64," + base64.b64encode(open(path, "rb").read()).decode()
    except Exception:
        return ""


def fig_b64(fig):
    buf = io.BytesIO()
    fig.savefig(buf, format="png", dpi=110, bbox_inches="tight")
    plt.close(fig)
    return "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()


def grouped_bar(labels, series, colors, ylabel, intlabels=False):
    """series: list of (name, [values]) aligned to labels."""
    x = np.arange(len(labels))
    w = 0.8 / max(1, len(series))
    fig, ax = plt.subplots(figsize=(7.2, 3.6))
    for i, (name, vals) in enumerate(series):
        bars = ax.bar(x + (i - (len(series) - 1) / 2) * w, vals, w, label=name, color=colors[name])
        for b, v in zip(bars, vals):
            if v:
                ax.text(b.get_x() + b.get_width() / 2, v, (f"{int(v)}" if intlabels else f"{v:.2f}"),
                        ha="center", va="bottom", fontsize=6.3)
    ax.set_xticks(x); ax.set_xticklabels(labels, fontsize=8.5)
    ax.set_ylabel(ylabel)
    ax.legend(fontsize=7.5, ncol=len(series), loc="lower center", bbox_to_anchor=(0.5, 1.0), frameon=False)
    ax.grid(axis="y", ls=":", alpha=0.4); ax.set_axisbelow(True)
    if not intlabels:
        ax.set_ylim(0, 1.08)
    for s in ("top", "right"):
        ax.spines[s].set_visible(False)
    return fig_b64(fig)


def perf_bar(metrics, bar_keys):
    keys = [k for k, _ in bar_keys]
    series = [(MLABEL[m], [(metrics.get(m) or {}).get(k) if isinstance((metrics.get(m) or {}).get(k), (int, float)) else 0
                           for k in keys]) for m in MODELS]
    colors = {MLABEL[m]: MCOLOR[m] for m in MODELS}
    return grouped_bar([l for _, l in bar_keys], series, colors, "score", intlabels=False)


def dist_bar(classes, cnt):
    labels = [f"{l}·{cn}" for l, cn in classes]
    series = [(sp, [cnt[sp].get(l, 0) for l, _ in classes]) for sp in ("train", "val", "test")]
    return grouped_bar(labels, series, SPLIT_COLOR, "images", intlabels=True)


def perf_table(metrics, cols):
    keys = [k for k, _ in cols]
    best = {}
    for k in keys:
        vals = [(metrics[m].get(k) if metrics.get(m) and isinstance(metrics[m].get(k), (int, float)) else None) for m in MODELS]
        vals = [v for v in vals if v is not None]
        best[k] = max(vals) if vals else None
    h = ['<table class="metrics"><thead><tr><th>Model</th>'] + [f"<th>{lab}</th>" for _, lab in cols] + ["</tr></thead><tbody>"]
    for m in MODELS:
        mm = metrics.get(m) or {}
        h.append(f'<tr><td class="model"><span class="dot" style="background:{MCOLOR[m]}"></span>{MLABEL[m]}</td>')
        for k, _ in cols:
            v = mm.get(k)
            if isinstance(v, (int, float)):
                cls = "best" if (best[k] is not None and abs(v - best[k]) < 1e-9) else ""
                h.append(f'<td class="{cls}">{v:.4f}</td>')
            else:
                h.append("<td>—</td>")
        h.append("</tr>")
    h.append("</tbody></table>")
    return "".join(h)


def dist_table(classes, cnt):
    h = ['<table class="metrics"><thead><tr><th>Split</th>'] + [f"<th>{l}·{cn}</th>" for l, cn in classes] + ["<th>合计</th></tr></thead><tbody>"]
    coltot = {l: 0 for l, _ in classes}
    for sp in ("train", "val", "test"):
        h.append(f'<tr><td class="model">{sp}</td>')
        s = 0
        for l, _ in classes:
            v = cnt[sp].get(l, 0); s += v; coltot[l] += v
            h.append(f"<td>{v}</td>")
        h.append(f"<td><b>{s}</b></td></tr>")
    h.append('<tr><td class="model">合计</td>')
    for l, _ in classes:
        h.append(f"<td><b>{coltot[l]}</b></td>")
    h.append(f'<td><b>{sum(coltot.values())}</b></td></tr></tbody></table>')
    return "".join(h)


def perclass_table(metrics, classes):
    """Per-class Recall / F1 / AUROC for each model + shared support. Best F1 per class highlighted."""
    def fmt(v):
        return f"{v:.3f}" if isinstance(v, (int, float)) else "—"
    h = ['<table class="metrics pc"><thead><tr><th rowspan="2">Class</th><th rowspan="2">Support</th>']
    for m in MODELS:
        h.append(f'<th colspan="3" class="grp">{MLABEL[m]}</th>')
    h.append("</tr><tr>")
    for m in MODELS:
        h.append('<th class="grp">Recall</th><th>F1</th><th>AUROC</th>')
    h.append("</tr></thead><tbody>")
    for l, cn in classes:
        sup = None
        for m in MODELS:
            pc = (metrics.get(m) or {}).get("per_class", {})
            if str(l) in pc:
                sup = int(pc[str(l)].get("support", 0)); break
        f1s = [(metrics.get(m) or {}).get("per_class", {}).get(str(l), {}).get("f1-score") for m in MODELS]
        bf = max([v for v in f1s if isinstance(v, (int, float))], default=None)
        h.append(f'<tr><td class="model">{l}·{cn}</td><td>{sup if sup is not None else "—"}</td>')
        for m in MODELS:
            mm = metrics.get(m) or {}
            pc = mm.get("per_class", {}).get(str(l), {})
            au = (mm.get("auroc_per_class") or {}).get(str(l))
            f1 = pc.get("f1-score")
            f1cls = "best" if (isinstance(f1, (int, float)) and bf is not None and abs(f1 - bf) < 1e-9) else ""
            h.append(f'<td class="grp">{fmt(pc.get("recall"))}</td><td class="{f1cls}">{fmt(f1)}</td><td>{fmt(au)}</td>')
        h.append("</tr>")
    h.append("</tbody></table>")
    return "".join(h)


def perclass_f1_bar(metrics, classes):
    labels = [f"{l}·{cn}" for l, cn in classes]
    series = []
    for m in MODELS:
        pc = (metrics.get(m) or {}).get("per_class", {})
        vals = [pc.get(str(l), {}).get("f1-score") if isinstance(pc.get(str(l), {}).get("f1-score"), (int, float)) else 0
                for l, _ in classes]
        series.append((MLABEL[m], vals))
    return grouped_bar(labels, series, {MLABEL[m]: MCOLOR[m] for m in MODELS}, "per-class F1", intlabels=False)


def gallery_html(dsk):
    rows = []
    for kind, title in [("confusion_matrix", "混淆矩阵"), ("roc", "ROC")]:
        cells = []
        for m in MODELS:
            img = b64_img(os.path.join(RESULTS, dsk, m, f"{kind}.png"))
            if img:
                cells.append(f'<figure><img src="{img}"><figcaption>{MLABEL[m]}</figcaption></figure>')
        rows.append(f'<div class="grow"><h4>{title}</h4><div class="imgrow">{"".join(cells)}</div></div>')
    return f'<details><summary>详细图：混淆矩阵 / ROC 曲线</summary><div class="gallery">{"".join(rows)}</div></details>'


def downsample_block(dsk):
    """Per-dataset data-scarcity: for each fraction, a card with table + bar chart of all 4 metrics."""
    if dsk not in ("adam", "airogs", "papila"):
        return ""
    DS = os.path.join(RESULTS, "downsample")
    frac_show = [100, 50, 25, 10, 5]
    dnames = {"adam": "ADAM", "airogs": "AIROGS", "papila": "PAPILA"}
    train_counts = {"adam": {100:280,50:140,25:70,10:28,5:14},
                    "airogs": {100:5000,50:2500,25:1250,10:500,5:250},
                    "papila": {100:294,50:146,25:73,10:29,5:15}}
    metrics_keys = [("accuracy", "Acc"), ("auroc", "AUROC"), ("auprc", "AUPRC"), ("f1_macro", "F1"), ("sensitivity", "Sens"), ("specificity", "Spec")]
    # load per-fraction metrics
    frac_data = {}
    for f in frac_show:
        frac_data[f] = {}
        for m in MODELS:
            p = os.path.join(DS, dsk, f"{f:03d}", m, "metrics.json")
            if os.path.isfile(p):
                frac_data[f][m] = json.load(open(p))
    cards = []
    for f in frac_show:
        labels = [kn for _, kn in metrics_keys]
        series = []
        for m in MODELS:
            mm = frac_data[f].get(m)
            vals = []
            for k, _ in metrics_keys:
                v = mm.get(k) if mm else None
                vals.append(v if isinstance(v, (int, float)) else 0)
            series.append((MLABEL[m], vals))
        bar = grouped_bar(labels, series, {MLABEL[m]: MCOLOR[m] for m in MODELS}, "score", intlabels=False)
        # table
        th = ['<table class="metrics"><thead><tr><th>Model</th>']
        for _, kn in metrics_keys:
            th.append(f'<th>{kn}</th>')
        th.append("</tr></thead><tbody>")
        best_map = {}
        for k, _ in metrics_keys:
            bv = None
            for m in MODELS:
                mm = frac_data[f].get(m)
                if mm:
                    v = mm.get(k)
                    if isinstance(v, (int, float)):
                        if bv is None or v > bv:
                            bv = v
            best_map[k] = bv
        for m in MODELS:
            mm = frac_data[f].get(m)
            th.append(f'<tr><td class="model"><span class="dot" style="background:{MCOLOR[m]}"></span>{m}</td>')
            for k, _ in metrics_keys:
                v = mm.get(k) if mm else None
                cls = "best" if (isinstance(v, (int, float)) and best_map[k] is not None and abs(v - best_map[k]) < 1e-9) else ""
                vs = f"{v:.4f}" if isinstance(v, (int, float)) else "—"
                th.append(f'<td class="{cls}">{vs}</td>')
            th.append("</tr>")
        th.append("</tbody></table>")
        cards.append(f"""
        <div style="border:1px solid #eef0f3;border-radius:8px;margin:12px 20px;overflow:hidden">
          <div class="card-h" style="background:#fafbfc"><h4 style="font-size:14px;margin:0">{f}% · {train_counts[dsk][f]} 训练样本</h4><span class="meta">val/test 保持完整</span></div>
          <div class="content" style="padding:10px 20px">
            <div class="tbl">{"".join(th)}</div>
            <div class="chart"><img src="{bar}"></div>
          </div>
        </div>""")
    return f"""
        <h4 class="sec">数据稀缺性分析 · Data-scarcity experiment</h4>
        <div class="bg" style="font-size:12px;color:#6b21a8">训练数据按类别分层抽样至 100/50/25/10/5%，保持 val/test 完整。<b>PAPILA 随数据量下降最快</b>，最适合作合成数据增广实验。</div>
        {''.join(cards)}"""


def main():
    parts = []
    for cat_name, color, dsets in CATS:
        cards = []
        for dsk, title, desc in dsets:
            metrics = {m: load(dsk, m) for m in MODELS}
            if not any(metrics.values()):
                continue
            task = next(v for v in metrics.values() if v)["task"]
            cols = BIN_COLS if task == "binary" else MUL_COLS
            bar = BIN_BAR if task == "binary" else MUL_BAR
            ntest = next(v for v in metrics.values() if v).get("n_test")
            classes, cnt, total = read_dist(dsk)
            perclass_block = "" if task == "binary" else f"""
        <h4 class="sec">每类指标 · Per-class metrics</h4>
        <div class="content">
          <div class="tbl">{perclass_table(metrics, classes)}</div>
          <div class="chart"><img src="{perclass_f1_bar(metrics, classes)}"></div>
        </div>"""
            cards.append(f"""
      <div class="card">
        <div class="card-h"><h3>{title}</h3>
          <span class="meta">{desc} · 划分 train/val/test = {SPLITS.get(dsk,'?')} · 总计 {total} 张</span></div>
        <div class="bg"><b>📷 采集背景：</b>{BG.get(dsk,'')}</div>
        <h4 class="sec">类别分布 · Class distribution（按 split）</h4>
        <div class="content">
          <div class="tbl">{dist_table(classes, cnt)}</div>
          <div class="chart"><img src="{dist_bar(classes, cnt)}"></div>
        </div>
        <h4 class="sec">模型性能 · Performance</h4>
        <div class="content">
          <div class="tbl">{perf_table(metrics, cols)}</div>
          <div class="chart"><img src="{perf_bar(metrics, bar)}"></div>
        </div>{perclass_block}
        {downsample_block(dsk)}
        {gallery_html(dsk)}
      </div>""")
        parts.append(f'\n    <section class="module" style="--accent:{color}">\n      <h2>{cat_name}</h2>\n      {"".join(cards)}\n    </section>')

    html = f"""<!doctype html><html lang="zh"><head><meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>眼底图像分类 Benchmark</title>
<style>
  :root{{font-family:-apple-system,"Segoe UI",Roboto,"PingFang SC","Microsoft YaHei",sans-serif}}
  body{{margin:0;background:#f5f6f8;color:#1f2430;line-height:1.5}}
  header{{background:linear-gradient(135deg,#1e293b,#334155);color:#fff;padding:32px 40px}}
  header h1{{margin:0 0 6px;font-size:26px}} header p{{margin:2px 0;opacity:.85;font-size:14px}}
  .wrap{{max-width:1180px;margin:0 auto;padding:24px 20px 60px}}
  .module{{margin:34px 0}}
  .module h2{{border-left:7px solid var(--accent);padding:6px 0 6px 14px;font-size:21px;margin:0 0 16px;color:var(--accent)}}
  .card{{background:#fff;border-radius:12px;box-shadow:0 1px 4px rgba(0,0,0,.08);margin:0 0 20px;overflow:hidden}}
  .card-h{{padding:14px 20px;border-bottom:1px solid #eef0f3;display:flex;align-items:baseline;gap:14px;flex-wrap:wrap}}
  .card-h h3{{margin:0;font-size:17px}} .card-h .meta{{color:#788196;font-size:12.5px}}
  .bg{{margin:14px 20px 4px;padding:10px 14px;background:#f8fafc;border:1px solid #eef0f3;border-radius:8px;font-size:12.5px;color:#475569;line-height:1.8}}
  .bg b{{color:#334155}}
  h4.sec{{margin:16px 20px 2px;font-size:13.5px;color:#334155;border-left:3px solid var(--accent);padding-left:9px}}
  .content{{display:flex;gap:18px;padding:8px 20px 14px;flex-wrap:wrap;align-items:flex-start}}
  .tbl{{flex:1 1 440px;overflow-x:auto}} .chart{{flex:1 1 480px;text-align:center}} .chart img{{max-width:100%;height:auto}}
  table.metrics{{border-collapse:collapse;width:100%;font-size:13px}}
  table.metrics th,table.metrics td{{padding:6px 8px;text-align:center;border-bottom:1px solid #eef0f3}}
  table.metrics thead th{{background:#f8f9fb;color:#5b6472;font-weight:600;font-size:11.5px}}
  table.metrics td.model{{text-align:left;white-space:nowrap;font-weight:600}}
  .dot{{display:inline-block;width:9px;height:9px;border-radius:50%;margin-right:7px;vertical-align:middle}}
  table.metrics td.best{{background:#fff7e6;color:#b45309;font-weight:700}}
  table.metrics .grp{{border-left:2px solid #e2e6ea}}
  details{{border-top:1px solid #eef0f3;padding:10px 20px 16px}} summary{{cursor:pointer;color:#475569;font-size:13px;font-weight:600}}
  .grow h4{{margin:14px 0 8px;font-size:13px;color:#5b6472}}
  .imgrow{{display:flex;gap:12px;flex-wrap:wrap}}
  .imgrow figure{{margin:0;flex:1 1 300px;max-width:360px;text-align:center}}
  .imgrow img{{width:100%;border:1px solid #eef0f3;border-radius:6px}} .imgrow figcaption{{font-size:11.5px;color:#788196;margin-top:4px}}
  footer{{text-align:center;color:#9aa3b2;font-size:12px;padding:20px}}
</style></head><body>
<header>
  <h1>眼底图像分类 Benchmark · RetFound vs ResNet vs ViT</h1>
  <p>7 个数据集 · 4 个疾病方向 · 三模型（RetFound ViT-L / ResNet-50 / ViT-B/16，均预训练后全参数微调）</p>
  <p>每个数据集含：采集背景（FOV/设备/来源/分辨率）· 类别分布（按 split）· 模型性能（指标表 + 柱状图）· 混淆矩阵/ROC</p>
  <p>评估协议：输入 224 · 官方划分优先（否则 7:1:2 分层）· val 选最优→测 test · 指标由统一脚本计算</p>
</header>
<div class="wrap">
  {''.join(parts)}
  <footer>采集背景均据权威来源整理，不确定处已标注「设备标称/未公开/补充材料」· 单文件离线可用</footer>
</div></body></html>"""
    open(OUT, "w").write(html)
    print(f"wrote {OUT}  ({len(html)/1024:.0f} KB)")


if __name__ == "__main__":
    main()