add full review and auto upload notebooks/colab_gas_valve_auto_label_train.ipynb

notebooks/colab_gas_valve_auto_label_train.ipynb

@@ -80,7 +80,7 @@
         "\n",
         "get_ipython().system(\n",
         "    \"pip install -q ultralytics huggingface_hub opencv-python-headless \"\n",
-        "    \"tqdm pyyaml matplotlib pandas\"\n",
+        "    \"tqdm pyyaml matplotlib pandas ipywidgets\"\n",
         ")\n",
         "\n",
         "import os\n",
@@ -89,6 +89,7 @@
         "import shutil\n",
         "import random\n",
         "import subprocess\n",
+        "import base64\n",
         "from pathlib import Path\n",
         "\n",
         "import cv2\n",
@@ -175,6 +176,7 @@
         "RAW_VALVE_PREFIX = \"raw/valve\"\n",
         "RAW_GAS_EXTRA_PREFIX = \"raw/gas_extra\"\n",
         "REVIEWED_VALVE_LABEL_PREFIX = \"reviewed_labels/valve\"\n",
+        "CLEANING_PREFIX = \"cleaning\"\n",
         "PROCESS_SCRIPT_REPO_PATH = \"scripts/process_valve_to_polygon.py\"\n",
         "FORCE_REDOWNLOAD = True\n",
         "\n",
@@ -188,6 +190,7 @@
         "    f\"{RAW_VALVE_PREFIX}/**\",\n",
         "    f\"{RAW_GAS_EXTRA_PREFIX}/**\",\n",
         "    f\"{REVIEWED_VALVE_LABEL_PREFIX}/**\",\n",
+        "    f\"{CLEANING_PREFIX}/**\",\n",
         "    PROCESS_SCRIPT_REPO_PATH,\n",
         "    \"dataset.yaml\",\n",
         "]\n",
@@ -205,6 +208,7 @@
         "RAW_VALVE_DIR = base / RAW_VALVE_PREFIX\n",
         "RAW_GAS_EXTRA_DIR = base / RAW_GAS_EXTRA_PREFIX\n",
         "REVIEWED_VALVE_LABEL_DIR = base / REVIEWED_VALVE_LABEL_PREFIX\n",
+        "CLEANING_DIR = base / CLEANING_PREFIX\n",
         "PROCESS_SCRIPT_LOCAL = base / PROCESS_SCRIPT_REPO_PATH\n",
         "\n",
         "print(\"下载结果统计\")\n",
@@ -214,6 +218,7 @@
         "    RAW_VALVE_DIR,\n",
         "    RAW_GAS_EXTRA_DIR,\n",
         "    REVIEWED_VALVE_LABEL_DIR,\n",
+        "    CLEANING_DIR,\n",
         "]:\n",
         "    n = len(list(p.glob(\"*\"))) if p.exists() else 0\n",
         "    print(f\"  {p}: {n} entries\")\n",
@@ -678,7 +683,21 @@
         "    return keys\n",
         "\n",
         "\n",
-        "EXCLUDED_SAMPLE_KEYS = parse_exclude_keys(EXCLUDE_SAMPLE_KEYS)\n",
+        "def read_exclude_keys_file(path: Path) -> set[str]:\n",
+        "    if not path.exists():\n",
+        "        return set()\n",
+        "    return parse_exclude_keys(path.read_text(encoding=\"utf-8\", errors=\"ignore\"))\n",
+        "\n",
+        "\n",
+        "PERSISTED_EXCLUDE_PATH = CLEANING_DIR / \"excluded_sample_keys.txt\"\n",
+        "PERSISTED_EXCLUDED_SAMPLE_KEYS = read_exclude_keys_file(PERSISTED_EXCLUDE_PATH)\n",
+        "MANUAL_EXCLUDED_SAMPLE_KEYS = parse_exclude_keys(EXCLUDE_SAMPLE_KEYS)\n",
+        "EXCLUDED_SAMPLE_KEYS = PERSISTED_EXCLUDED_SAMPLE_KEYS | MANUAL_EXCLUDED_SAMPLE_KEYS\n",
+        "\n",
+        "if PERSISTED_EXCLUDED_SAMPLE_KEYS:\n",
+        "    print(f\"HuggingFace 已保存排除样本: {len(PERSISTED_EXCLUDED_SAMPLE_KEYS)}\")\n",
+        "if MANUAL_EXCLUDED_SAMPLE_KEYS:\n",
+        "    print(f\"本次手动填写排除样本: {len(MANUAL_EXCLUDED_SAMPLE_KEYS)}\")\n",
         "print(f\"Excluded sample keys: {len(EXCLUDED_SAMPLE_KEYS)}\")\n",
         "for key in sorted(EXCLUDED_SAMPLE_KEYS):\n",
         "    print(f\"  {key}\")\n",
@@ -874,7 +893,10 @@
       "source": [
         "# @title Step 10 - 构建两类训练集\n",
         "\n",
+        "import pandas as pd\n",
+        "\n",
         "COMBINED_DATASET = Path(\"/content/gas_valve_dataset\")\n",
+        "COMBINED_MANIFEST_PATH = COMBINED_DATASET / \"training_manifest.csv\"\n",
         "VAL_RATIO = 0.20              # 20% 做验证集，80% 做训练集。\n",
         "RANDOM_SEED = 42              # 固定随机种子，方便复现实验。\n",
         "INCLUDE_OLD_GAS_DATA = True   # 保留原来的气瓶数据一起训练。\n",
@@ -946,6 +968,8 @@
         "    })\n",
         "\n",
         "excluded_keys = globals().get(\"EXCLUDED_SAMPLE_KEYS\", set())\n",
+        "if \"read_exclude_keys_file\" in globals():\n",
+        "    excluded_keys = set(excluded_keys) | read_exclude_keys_file(CLEANING_DIR / \"excluded_sample_keys.txt\")\n",
         "if excluded_keys:\n",
         "    before_filter = len(samples)\n",
         "    samples = [\n",
@@ -976,6 +1000,8 @@
         "rng.shuffle(train_samples)\n",
         "rng.shuffle(val_samples)\n",
         "\n",
+        "training_manifest_rows: list[dict] = []\n",
+        "\n",
         "\n",
         "def materialize(split: str, split_samples: list[dict]) -> None:\n",
         "    for idx, sample in enumerate(split_samples, 1):\n",
@@ -985,10 +1011,27 @@
         "        dst_lbl = COMBINED_DATASET / f\"labels/{split}/{stem}.txt\"\n",
         "        shutil.copy2(sample[\"image\"], dst_img)\n",
         "        write_lines(dst_lbl, sample[\"lines\"])\n",
+        "        training_manifest_rows.append({\n",
+        "            \"split\": split,\n",
+        "            \"dataset_stem\": stem,\n",
+        "            \"image_path\": str(dst_img),\n",
+        "            \"label_path\": str(dst_lbl),\n",
+        "            \"source_key\": sample[\"source_key\"],\n",
+        "            \"source\": sample[\"source\"],\n",
+        "            \"source_stem\": sample[\"source_stem\"],\n",
+        "            \"group\": sample[\"group\"],\n",
+        "            \"classes\": \",\".join(sorted({line.split()[0] for line in sample[\"lines\"]})),\n",
+        "            \"objects\": len(sample[\"lines\"]),\n",
+        "        })\n",
         "\n",
         "\n",
         "materialize(\"train\", train_samples)\n",
         "materialize(\"val\", val_samples)\n",
+        "pd.DataFrame(training_manifest_rows).to_csv(\n",
+        "    COMBINED_MANIFEST_PATH,\n",
+        "    index=False,\n",
+        "    encoding=\"utf-8-sig\",\n",
+        ")\n",
         "\n",
         "YAML_PATH = COMBINED_DATASET / \"dataset.yaml\"\n",
         "yaml_content = {\n",
@@ -1015,6 +1058,7 @@
         "print(f\"  val labels: {len(val_label_files)}\")\n",
         "print(f\"  train class counts: {class_counts(train_label_files)}\")\n",
         "print(f\"  val class counts: {class_counts(val_label_files)}\")\n",
+        "print(f\"  manifest: {COMBINED_MANIFEST_PATH}\")\n",
         "print(YAML_PATH.read_text(encoding=\"utf-8\"))"
       ]
     },
@@ -1023,7 +1067,7 @@
       "metadata": {},
       "source": [
         "---\n",
-        "## Step 11 - 可视化抽查训练集"
+        "## Step 11 - 全量审核训练集并删除坏样本"
       ]
     },
     {
@@ -1034,14 +1078,19 @@
       },
       "outputs": [],
       "source": [
-        "# @title Step 11 - 可视化抽查训练集\n",
+        "# @title Step 11 - 全量审核训练集并删除坏样本\n",
+        "# @markdown 这一步会显示 train/val 全部样本。勾选错误样本后点击红色按钮，会从当前 Colab 训练集删除，并把排除清单保存到 HuggingFace。\n",
+        "\n",
+        "from IPython.display import display\n",
+        "import ipywidgets as widgets\n",
         "\n",
-        "def draw_label_overlay(img_path: Path, label_path: Path) -> np.ndarray:\n",
+        "\n",
+        "def draw_label_overlay(img_path: Path, label_path: Path, max_side: int = 720) -> np.ndarray:\n",
         "    img = read_image(img_path)\n",
         "    if img is None:\n",
         "        raise ValueError(f\"Cannot read {img_path}\")\n",
         "    h, w = img.shape[:2]\n",
-        "    scale = 720 / max(w, h)\n",
+        "    scale = max_side / max(w, h)\n",
         "    disp = cv2.resize(img, (int(w * scale), int(h * scale)), interpolation=cv2.INTER_AREA)\n",
         "    dh, dw = disp.shape[:2]\n",
         "    colors = {\n",
@@ -1065,23 +1114,184 @@
         "    return cv2.cvtColor(disp, cv2.COLOR_BGR2RGB)\n",
         "\n",
         "\n",
-        "preview_imgs = list_images(COMBINED_DATASET / \"images/train\")\n",
-        "preview_imgs = random.sample(preview_imgs, min(6, len(preview_imgs)))\n",
+        "def overlay_data_url(img_path: Path, label_path: Path, max_side: int = 320) -> str:\n",
+        "    rgb = draw_label_overlay(img_path, label_path, max_side=max_side)\n",
+        "    bgr = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)\n",
+        "    ok, buf = cv2.imencode(\".jpg\", bgr, [int(cv2.IMWRITE_JPEG_QUALITY), 82])\n",
+        "    if not ok:\n",
+        "        raise ValueError(f\"Cannot encode overlay for {img_path}\")\n",
+        "    encoded = base64.b64encode(buf.tobytes()).decode(\"ascii\")\n",
+        "    return f\"data:image/jpeg;base64,{encoded}\"\n",
         "\n",
-        "if preview_imgs:\n",
-        "    fig, axes = plt.subplots(2, 3, figsize=(18, 10))\n",
-        "    axes = axes.flatten()\n",
-        "    for ax, img_path in zip(axes, preview_imgs):\n",
-        "        lbl_path = COMBINED_DATASET / \"labels/train\" / f\"{img_path.stem}.txt\"\n",
-        "        ax.imshow(draw_label_overlay(img_path, lbl_path))\n",
-        "        ax.set_title(img_path.name, fontsize=9)\n",
-        "        ax.axis(\"off\")\n",
-        "    for ax in axes[len(preview_imgs):]:\n",
-        "        ax.axis(\"off\")\n",
-        "    plt.tight_layout()\n",
-        "    plt.show()\n",
-        "else:\n",
-        "    print(\"没有找到可预览图片\")"
+        "\n",
+        "def write_persisted_exclusions(keys: set[str], selected_df: pd.DataFrame) -> tuple[Path, Path]:\n",
+        "    CLEANING_DIR.mkdir(parents=True, exist_ok=True)\n",
+        "    txt_path = CLEANING_DIR / \"excluded_sample_keys.txt\"\n",
+        "    csv_path = CLEANING_DIR / \"excluded_sample_keys_latest.csv\"\n",
+        "\n",
+        "    txt_path.write_text(\n",
+        "        \"\\n\".join(sorted(keys)) + (\"\\n\" if keys else \"\"),\n",
+        "        encoding=\"utf-8\",\n",
+        "    )\n",
+        "\n",
+        "    selected_out = selected_df[\n",
+        "        [\"source_key\", \"source\", \"source_stem\", \"group\", \"split\", \"dataset_stem\"]\n",
+        "    ].copy()\n",
+        "    selected_out[\"deleted_at\"] = time.strftime(\"%Y-%m-%d %H:%M:%S\")\n",
+        "    selected_out.to_csv(csv_path, index=False, encoding=\"utf-8-sig\")\n",
+        "    return txt_path, csv_path\n",
+        "\n",
+        "\n",
+        "def upload_persisted_exclusions(txt_path: Path, csv_path: Path) -> None:\n",
+        "    api.upload_file(\n",
+        "        repo_id=DATASET_REPO,\n",
+        "        repo_type=\"dataset\",\n",
+        "        token=HF_TOKEN,\n",
+        "        path_or_fileobj=str(txt_path),\n",
+        "        path_in_repo=f\"{CLEANING_PREFIX}/excluded_sample_keys.txt\",\n",
+        "        commit_message=\"update excluded training sample keys\",\n",
+        "    )\n",
+        "    api.upload_file(\n",
+        "        repo_id=DATASET_REPO,\n",
+        "        repo_type=\"dataset\",\n",
+        "        token=HF_TOKEN,\n",
+        "        path_or_fileobj=str(csv_path),\n",
+        "        path_in_repo=f\"{CLEANING_PREFIX}/excluded_sample_keys_latest.csv\",\n",
+        "        commit_message=\"update latest excluded training sample table\",\n",
+        "    )\n",
+        "\n",
+        "\n",
+        "if not COMBINED_MANIFEST_PATH.exists():\n",
+        "    raise FileNotFoundError(\n",
+        "        f\"Missing {COMBINED_MANIFEST_PATH}. Please run Step 10 before Step 11.\"\n",
+        "    )\n",
+        "\n",
+        "training_review_df = pd.read_csv(COMBINED_MANIFEST_PATH)\n",
+        "training_review_df = training_review_df.sort_values(\n",
+        "    [\"split\", \"source\", \"source_key\"]\n",
+        ").reset_index(drop=True)\n",
+        "\n",
+        "print(f\"全量训练集样本数: {len(training_review_df)}\")\n",
+        "print(training_review_df.groupby([\"split\", \"source\"]).size().to_string())\n",
+        "print(\"\")\n",
+        "print(\"绿色是 class 0 气瓶，橙蓝色是 class 1 阀门。确认有问题的样本后，勾选图片下方复选框。\")\n",
+        "\n",
+        "checkboxes: dict[int, widgets.Checkbox] = {}\n",
+        "cards: list[widgets.Widget] = []\n",
+        "\n",
+        "for idx, row in training_review_df.iterrows():\n",
+        "    img_path = Path(row[\"image_path\"])\n",
+        "    lbl_path = Path(row[\"label_path\"])\n",
+        "    if img_path.exists() and lbl_path.exists():\n",
+        "        img_html = (\n",
+        "            f\"<img src='{overlay_data_url(img_path, lbl_path)}' \"\n",
+        "            \"style='width:260px; height:auto; display:block;'>\"\n",
+        "        )\n",
+        "    else:\n",
+        "        img_html = \"<div style='width:260px;height:160px;background:#eee;'>missing file</div>\"\n",
+        "\n",
+        "    title = (\n",
+        "        \"<div style='font-size:12px;line-height:1.35;\"\n",
+        "        \"font-family:Arial, sans-serif;word-break:break-all;'>\"\n",
+        "        f\"<b>{row['split']}</b> | {row['source_key']}<br>\"\n",
+        "        f\"{row['dataset_stem']} | class {row['classes']} | objects {row['objects']}\"\n",
+        "        \"</div>\"\n",
+        "    )\n",
+        "    checkbox = widgets.Checkbox(\n",
+        "        value=False,\n",
+        "        description=\"删除这个样本\",\n",
+        "        indent=False,\n",
+        "        layout=widgets.Layout(width=\"260px\"),\n",
+        "    )\n",
+        "    checkboxes[idx] = checkbox\n",
+        "    cards.append(widgets.VBox(\n",
+        "        [\n",
+        "            widgets.HTML(title + img_html),\n",
+        "            checkbox,\n",
+        "        ],\n",
+        "        layout=widgets.Layout(\n",
+        "            width=\"280px\",\n",
+        "            border=\"1px solid #ddd\",\n",
+        "            padding=\"6px\",\n",
+        "            margin=\"4px\",\n",
+        "        ),\n",
+        "    ))\n",
+        "\n",
+        "delete_button = widgets.Button(\n",
+        "    description=\"删除勾选样本并保存到 HuggingFace\",\n",
+        "    button_style=\"danger\",\n",
+        "    layout=widgets.Layout(width=\"280px\"),\n",
+        ")\n",
+        "summary_label = widgets.HTML(\n",
+        "    value=\"<b>操作提示：</b>删除后不用重新跑 Step 10，直接运行 Step 12 训练即可。\"\n",
+        ")\n",
+        "delete_output = widgets.Output()\n",
+        "\n",
+        "\n",
+        "def on_delete_clicked(_button):\n",
+        "    selected_indices = [idx for idx, cb in checkboxes.items() if cb.value]\n",
+        "    with delete_output:\n",
+        "        delete_output.clear_output()\n",
+        "        if not selected_indices:\n",
+        "            print(\"没有勾选任何样本。\")\n",
+        "            return\n",
+        "\n",
+        "        selected_df = training_review_df.loc[selected_indices].copy()\n",
+        "        selected_keys = set(selected_df[\"source_key\"].astype(str))\n",
+        "\n",
+        "        existing_keys = set(globals().get(\"EXCLUDED_SAMPLE_KEYS\", set()))\n",
+        "        if \"read_exclude_keys_file\" in globals():\n",
+        "            existing_keys |= read_exclude_keys_file(CLEANING_DIR / \"excluded_sample_keys.txt\")\n",
+        "        merged_keys = existing_keys | selected_keys\n",
+        "\n",
+        "        deleted_files = 0\n",
+        "        for _, selected_row in selected_df.iterrows():\n",
+        "            for col in [\"image_path\", \"label_path\"]:\n",
+        "                p = Path(selected_row[col])\n",
+        "                if p.exists():\n",
+        "                    p.unlink()\n",
+        "                    deleted_files += 1\n",
+        "\n",
+        "        remaining_df = training_review_df.drop(index=selected_indices).reset_index(drop=True)\n",
+        "        remaining_df.to_csv(COMBINED_MANIFEST_PATH, index=False, encoding=\"utf-8-sig\")\n",
+        "\n",
+        "        txt_path, csv_path = write_persisted_exclusions(merged_keys, selected_df)\n",
+        "        globals()[\"EXCLUDED_SAMPLE_KEYS\"] = merged_keys\n",
+        "\n",
+        "        print(f\"当前 Colab 训练集已删除样本: {len(selected_df)}\")\n",
+        "        print(f\"删除 image/label 文件数: {deleted_files}\")\n",
+        "        print(f\"累计排除 source_key 数: {len(merged_keys)}\")\n",
+        "        print(f\"本地清单: {txt_path}\")\n",
+        "        train_labels_now = list((COMBINED_DATASET / \"labels/train\").glob(\"*.txt\"))\n",
+        "        val_labels_now = list((COMBINED_DATASET / \"labels/val\").glob(\"*.txt\"))\n",
+        "        print(f\"剩余 train 图片: {len(list_images(COMBINED_DATASET / 'images/train'))}\")\n",
+        "        print(f\"剩余 val 图片: {len(list_images(COMBINED_DATASET / 'images/val'))}\")\n",
+        "        print(f\"剩余 train 类别计数: {class_counts(train_labels_now)}\")\n",
+        "        print(f\"剩余 val 类别计数: {class_counts(val_labels_now)}\")\n",
+        "\n",
+        "        try:\n",
+        "            upload_persisted_exclusions(txt_path, csv_path)\n",
+        "            print(\"HuggingFace 排除清单已保存。以后重新跑 Step 3-10 会自动排除这些样本。\")\n",
+        "        except Exception as exc:\n",
+        "            print(\"HuggingFace 上传排除清单失败，但当前 Colab 训练集已经删除。\")\n",
+        "            print(repr(exc))\n",
+        "\n",
+        "        for idx in selected_indices:\n",
+        "            checkboxes[idx].disabled = True\n",
+        "\n",
+        "\n",
+        "delete_button.on_click(on_delete_clicked)\n",
+        "\n",
+        "grid = widgets.GridBox(\n",
+        "    cards,\n",
+        "    layout=widgets.Layout(\n",
+        "        grid_template_columns=\"repeat(3, 300px)\",\n",
+        "        grid_gap=\"8px\",\n",
+        "        align_items=\"flex-start\",\n",
+        "    ),\n",
+        ")\n",
+        "\n",
+        "display(widgets.VBox([summary_label, delete_button, delete_output, grid]))"
       ]
     },
     {
@@ -1111,6 +1321,7 @@
         "BATCH = -1                      # 让 Ultralytics 自动选择显存能承受的 batch。\n",
         "PATIENCE = 35                   # 验证集 35 轮不提升就早停。\n",
         "RUN_NAME = f\"gas_valve_seg_reviewed_{time.strftime('%Y%m%d_%H%M')}\"\n",
+        "AUTO_UPLOAD_TRAINING_RESULTS = True\n",
         "\n",
         "print(\"训练参数\")\n",
         "print(f\"  model: {TRAIN_MODEL}\")\n",
@@ -1120,6 +1331,65 @@
         "print(f\"  batch: {BATCH}\")\n",
         "print(f\"  run: {RUN_NAME}\")\n",
         "\n",
+        "\n",
+        "TRAINING_UPLOAD_ITEMS = [\n",
+        "    \"weights/best.pt\",\n",
+        "    \"weights/last.pt\",\n",
+        "    \"results.png\",\n",
+        "    \"results.csv\",\n",
+        "    \"args.yaml\",\n",
+        "    \"confusion_matrix_normalized.png\",\n",
+        "    \"confusion_matrix.png\",\n",
+        "    \"BoxP_curve.png\",\n",
+        "    \"BoxR_curve.png\",\n",
+        "    \"BoxF1_curve.png\",\n",
+        "    \"BoxPR_curve.png\",\n",
+        "    \"MaskP_curve.png\",\n",
+        "    \"MaskR_curve.png\",\n",
+        "    \"MaskF1_curve.png\",\n",
+        "    \"MaskPR_curve.png\",\n",
+        "]\n",
+        "\n",
+        "\n",
+        "def upload_training_outputs(run_dir: Path, run_name: str) -> None:\n",
+        "    operations = []\n",
+        "    for rel in TRAINING_UPLOAD_ITEMS:\n",
+        "        local_file = run_dir / rel\n",
+        "        if not local_file.exists():\n",
+        "            continue\n",
+        "        operations.append(CommitOperationAdd(\n",
+        "            path_in_repo=f\"runs/{run_name}/{rel}\",\n",
+        "            path_or_fileobj=str(local_file),\n",
+        "        ))\n",
+        "        print(f\"准备上传: {rel}\")\n",
+        "\n",
+        "    extra_files = [\n",
+        "        (YAML_PATH, \"dataset.yaml\"),\n",
+        "        (COMBINED_MANIFEST_PATH, \"training_manifest.csv\"),\n",
+        "        (CLEANING_DIR / \"excluded_sample_keys.txt\", \"excluded_sample_keys.txt\"),\n",
+        "    ]\n",
+        "    for local_file, rel in extra_files:\n",
+        "        if local_file.exists():\n",
+        "            operations.append(CommitOperationAdd(\n",
+        "                path_in_repo=f\"runs/{run_name}/{rel}\",\n",
+        "                path_or_fileobj=str(local_file),\n",
+        "            ))\n",
+        "            print(f\"准备上传: {rel}\")\n",
+        "\n",
+        "    if not operations:\n",
+        "        print(\"没有找到可上传的训练结果文件。\")\n",
+        "        return\n",
+        "\n",
+        "    api.create_commit(\n",
+        "        repo_id=DATASET_REPO,\n",
+        "        repo_type=\"dataset\",\n",
+        "        token=HF_TOKEN,\n",
+        "        operations=operations,\n",
+        "        commit_message=f\"Training_results_{run_name}\",\n",
+        "    )\n",
+        "    print(f\"训练结果已上传: https://huggingface.co/datasets/{DATASET_REPO}/tree/main/runs/{run_name}\")\n",
+        "\n",
+        "\n",
         "model = YOLO(TRAIN_MODEL)\n",
         "t_start = time.time()\n",
         "train_results = model.train(\n",
@@ -1153,7 +1423,15 @@
         "elapsed = time.time() - t_start\n",
         "print(f\"训练完成，用时 {elapsed / 60:.1f} 分钟\")\n",
         "print(f\"Run dir: {RUN_DIR}\")\n",
-        "print(f\"Best weights: {RUN_DIR / 'weights/best.pt'}\")"
+        "print(f\"Best weights: {RUN_DIR / 'weights/best.pt'}\")\n",
+        "\n",
+        "if AUTO_UPLOAD_TRAINING_RESULTS:\n",
+        "    print(\"开始自动上传训练结果到 HuggingFace...\")\n",
+        "    try:\n",
+        "        upload_training_outputs(RUN_DIR, RUN_NAME)\n",
+        "    except Exception as exc:\n",
+        "        print(\"自动上传失败。训练结果仍保存在当前 Colab 的 RUN_DIR，可以运行 Step 14 手动重传。\")\n",
+        "        print(repr(exc))"
       ]
     },
     {
@@ -1214,59 +1492,12 @@
       "outputs": [],
       "source": [
         "# @title Step 14 - 上传训练结果\n",
+        "# @markdown Step 12 训练完成后已经会自动上传。本步骤只用于自动上传失败后的手动重传。\n",
         "\n",
-        "UPLOAD_TRAINING_RESULTS = True\n",
-        "UPLOAD_REPO = DATASET_REPO\n",
-        "UPLOAD_REPO_TYPE = \"dataset\"\n",
-        "\n",
-        "if UPLOAD_TRAINING_RESULTS:\n",
-        "    upload_items = [\n",
-        "        \"weights/best.pt\",\n",
-        "        \"weights/last.pt\",\n",
-        "        \"results.png\",\n",
-        "        \"results.csv\",\n",
-        "        \"args.yaml\",\n",
-        "        \"confusion_matrix_normalized.png\",\n",
-        "        \"confusion_matrix.png\",\n",
-        "        \"BoxP_curve.png\",\n",
-        "        \"BoxR_curve.png\",\n",
-        "        \"BoxF1_curve.png\",\n",
-        "        \"BoxPR_curve.png\",\n",
-        "        \"MaskP_curve.png\",\n",
-        "        \"MaskR_curve.png\",\n",
-        "        \"MaskF1_curve.png\",\n",
-        "        \"MaskPR_curve.png\",\n",
-        "    ]\n",
-        "\n",
-        "    operations = []\n",
-        "    for rel in upload_items:\n",
-        "        local_file = RUN_DIR / rel\n",
-        "        if not local_file.exists():\n",
-        "            continue\n",
-        "        operations.append(CommitOperationAdd(\n",
-        "            path_in_repo=f\"runs/{RUN_NAME}/{rel}\",\n",
-        "            path_or_fileobj=local_file.read_bytes(),\n",
-        "        ))\n",
-        "        print(f\"Prepared {rel}\")\n",
+        "if \"RUN_DIR\" not in globals() or \"RUN_NAME\" not in globals():\n",
+        "    raise RuntimeError(\"还没有训练结果。请先运行 Step 12。\")\n",
         "\n",
-        "    operations.append(CommitOperationAdd(\n",
-        "        path_in_repo=f\"runs/{RUN_NAME}/dataset.yaml\",\n",
-        "        path_or_fileobj=YAML_PATH.read_bytes(),\n",
-        "    ))\n",
-        "\n",
-        "    if operations:\n",
-        "        api.create_commit(\n",
-        "            repo_id=UPLOAD_REPO,\n",
-        "            repo_type=UPLOAD_REPO_TYPE,\n",
-        "            token=HF_TOKEN,\n",
-        "            operations=operations,\n",
-        "            commit_message=f\"Training_results_{RUN_NAME}\",\n",
-        "        )\n",
-        "        print(f\"Uploaded results to https://huggingface.co/datasets/{UPLOAD_REPO}/tree/main/runs/{RUN_NAME}\")\n",
-        "    else:\n",
-        "        print(\"No result files found to upload\")\n",
-        "else:\n",
-        "    print(\"Skipping result upload\")"
+        "upload_training_outputs(RUN_DIR, RUN_NAME)"
       ]
     },
     {