add full review and auto upload notebooks/colab_gas_valve_auto_label_train.py

notebooks/colab_gas_valve_auto_label_train.py

@@ -57,7 +57,7 @@
 
 get_ipython().system(
     "pip install -q ultralytics huggingface_hub opencv-python-headless "
-    "tqdm pyyaml matplotlib pandas"
+    "tqdm pyyaml matplotlib pandas ipywidgets"
 )
 
 import os
@@ -66,6 +66,7 @@ import time
 import shutil
 import random
 import subprocess
+import base64
 from pathlib import Path
 
 import cv2
@@ -132,6 +133,7 @@ DATASET_LOCAL = "/content/hf_11111"
 RAW_VALVE_PREFIX = "raw/valve"
 RAW_GAS_EXTRA_PREFIX = "raw/gas_extra"
 REVIEWED_VALVE_LABEL_PREFIX = "reviewed_labels/valve"
+CLEANING_PREFIX = "cleaning"
 PROCESS_SCRIPT_REPO_PATH = "scripts/process_valve_to_polygon.py"
 FORCE_REDOWNLOAD = True
 
@@ -145,6 +147,7 @@ allow_patterns = [
     f"{RAW_VALVE_PREFIX}/**",
     f"{RAW_GAS_EXTRA_PREFIX}/**",
     f"{REVIEWED_VALVE_LABEL_PREFIX}/**",
+    f"{CLEANING_PREFIX}/**",
     PROCESS_SCRIPT_REPO_PATH,
     "dataset.yaml",
 ]
@@ -162,6 +165,7 @@ snapshot_download(
 RAW_VALVE_DIR = base / RAW_VALVE_PREFIX
 RAW_GAS_EXTRA_DIR = base / RAW_GAS_EXTRA_PREFIX
 REVIEWED_VALVE_LABEL_DIR = base / REVIEWED_VALVE_LABEL_PREFIX
+CLEANING_DIR = base / CLEANING_PREFIX
 PROCESS_SCRIPT_LOCAL = base / PROCESS_SCRIPT_REPO_PATH
 
 print("下载结果统计")
@@ -171,6 +175,7 @@ for p in [
     RAW_VALVE_DIR,
     RAW_GAS_EXTRA_DIR,
     REVIEWED_VALVE_LABEL_DIR,
+    CLEANING_DIR,
 ]:
     n = len(list(p.glob("*"))) if p.exists() else 0
     print(f"  {p}: {n} entries")
@@ -575,7 +580,21 @@ def parse_exclude_keys(text: str) -> set[str]:
     return keys
 
 
-EXCLUDED_SAMPLE_KEYS = parse_exclude_keys(EXCLUDE_SAMPLE_KEYS)
+def read_exclude_keys_file(path: Path) -> set[str]:
+    if not path.exists():
+        return set()
+    return parse_exclude_keys(path.read_text(encoding="utf-8", errors="ignore"))
+
+
+PERSISTED_EXCLUDE_PATH = CLEANING_DIR / "excluded_sample_keys.txt"
+PERSISTED_EXCLUDED_SAMPLE_KEYS = read_exclude_keys_file(PERSISTED_EXCLUDE_PATH)
+MANUAL_EXCLUDED_SAMPLE_KEYS = parse_exclude_keys(EXCLUDE_SAMPLE_KEYS)
+EXCLUDED_SAMPLE_KEYS = PERSISTED_EXCLUDED_SAMPLE_KEYS | MANUAL_EXCLUDED_SAMPLE_KEYS
+
+if PERSISTED_EXCLUDED_SAMPLE_KEYS:
+    print(f"HuggingFace 已保存排除样本: {len(PERSISTED_EXCLUDED_SAMPLE_KEYS)}")
+if MANUAL_EXCLUDED_SAMPLE_KEYS:
+    print(f"本次手动填写排除样本: {len(MANUAL_EXCLUDED_SAMPLE_KEYS)}")
 print(f"Excluded sample keys: {len(EXCLUDED_SAMPLE_KEYS)}")
 for key in sorted(EXCLUDED_SAMPLE_KEYS):
     print(f"  {key}")
@@ -761,7 +780,10 @@ if DELETE_EXCLUDED_FROM_HF:
 # %%
 # @title Step 10 - 构建两类训练集
 
+import pandas as pd
+
 COMBINED_DATASET = Path("/content/gas_valve_dataset")
+COMBINED_MANIFEST_PATH = COMBINED_DATASET / "training_manifest.csv"
 VAL_RATIO = 0.20              # 20% 做验证集，80% 做训练集。
 RANDOM_SEED = 42              # 固定随机种子，方便复现实验。
 INCLUDE_OLD_GAS_DATA = True   # 保留原来的气瓶数据一起训练。
@@ -833,6 +855,8 @@ for img_path in list_images(RAW_VALVE_DIR):
     })
 
 excluded_keys = globals().get("EXCLUDED_SAMPLE_KEYS", set())
+if "read_exclude_keys_file" in globals():
+    excluded_keys = set(excluded_keys) | read_exclude_keys_file(CLEANING_DIR / "excluded_sample_keys.txt")
 if excluded_keys:
     before_filter = len(samples)
     samples = [
@@ -863,6 +887,8 @@ for group in sorted({s["group"] for s in samples}):
 rng.shuffle(train_samples)
 rng.shuffle(val_samples)
 
+training_manifest_rows: list[dict] = []
+
 
 def materialize(split: str, split_samples: list[dict]) -> None:
     for idx, sample in enumerate(split_samples, 1):
@@ -872,10 +898,27 @@ def materialize(split: str, split_samples: list[dict]) -> None:
         dst_lbl = COMBINED_DATASET / f"labels/{split}/{stem}.txt"
         shutil.copy2(sample["image"], dst_img)
         write_lines(dst_lbl, sample["lines"])
+        training_manifest_rows.append({
+            "split": split,
+            "dataset_stem": stem,
+            "image_path": str(dst_img),
+            "label_path": str(dst_lbl),
+            "source_key": sample["source_key"],
+            "source": sample["source"],
+            "source_stem": sample["source_stem"],
+            "group": sample["group"],
+            "classes": ",".join(sorted({line.split()[0] for line in sample["lines"]})),
+            "objects": len(sample["lines"]),
+        })
 
 
 materialize("train", train_samples)
 materialize("val", val_samples)
+pd.DataFrame(training_manifest_rows).to_csv(
+    COMBINED_MANIFEST_PATH,
+    index=False,
+    encoding="utf-8-sig",
+)
 
 YAML_PATH = COMBINED_DATASET / "dataset.yaml"
 yaml_content = {
@@ -902,23 +945,29 @@ print(f"  val images: {len(list_images(COMBINED_DATASET / 'images/val'))}")
 print(f"  val labels: {len(val_label_files)}")
 print(f"  train class counts: {class_counts(train_label_files)}")
 print(f"  val class counts: {class_counts(val_label_files)}")
+print(f"  manifest: {COMBINED_MANIFEST_PATH}")
 print(YAML_PATH.read_text(encoding="utf-8"))
 
 
 # %% [markdown]
 # ---
-# ## 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
 
 
 # %%
-# @title Step 11 - 可视化抽��训练集
+# @title Step 11 - 全量审核训练集并删除坏样本
+# @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。
+
+from IPython.display import display
+import ipywidgets as widgets
 
-def draw_label_overlay(img_path: Path, label_path: Path) -> np.ndarray:
+
+def draw_label_overlay(img_path: Path, label_path: Path, max_side: int = 720) -> np.ndarray:
     img = read_image(img_path)
     if img is None:
         raise ValueError(f"Cannot read {img_path}")
     h, w = img.shape[:2]
-    scale = 720 / max(w, h)
+    scale = max_side / max(w, h)
     disp = cv2.resize(img, (int(w * scale), int(h * scale)), interpolation=cv2.INTER_AREA)
     dh, dw = disp.shape[:2]
     colors = {
@@ -942,23 +991,184 @@ def draw_label_overlay(img_path: Path, label_path: Path) -> np.ndarray:
     return cv2.cvtColor(disp, cv2.COLOR_BGR2RGB)
 
 
-preview_imgs = list_images(COMBINED_DATASET / "images/train")
-preview_imgs = random.sample(preview_imgs, min(6, len(preview_imgs)))
+def overlay_data_url(img_path: Path, label_path: Path, max_side: int = 320) -> str:
+    rgb = draw_label_overlay(img_path, label_path, max_side=max_side)
+    bgr = cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR)
+    ok, buf = cv2.imencode(".jpg", bgr, [int(cv2.IMWRITE_JPEG_QUALITY), 82])
+    if not ok:
+        raise ValueError(f"Cannot encode overlay for {img_path}")
+    encoded = base64.b64encode(buf.tobytes()).decode("ascii")
+    return f"data:image/jpeg;base64,{encoded}"
+
+
+def write_persisted_exclusions(keys: set[str], selected_df: pd.DataFrame) -> tuple[Path, Path]:
+    CLEANING_DIR.mkdir(parents=True, exist_ok=True)
+    txt_path = CLEANING_DIR / "excluded_sample_keys.txt"
+    csv_path = CLEANING_DIR / "excluded_sample_keys_latest.csv"
+
+    txt_path.write_text(
+        "\n".join(sorted(keys)) + ("\n" if keys else ""),
+        encoding="utf-8",
+    )
+
+    selected_out = selected_df[
+        ["source_key", "source", "source_stem", "group", "split", "dataset_stem"]
+    ].copy()
+    selected_out["deleted_at"] = time.strftime("%Y-%m-%d %H:%M:%S")
+    selected_out.to_csv(csv_path, index=False, encoding="utf-8-sig")
+    return txt_path, csv_path
 
-if preview_imgs:
-    fig, axes = plt.subplots(2, 3, figsize=(18, 10))
-    axes = axes.flatten()
-    for ax, img_path in zip(axes, preview_imgs):
-        lbl_path = COMBINED_DATASET / "labels/train" / f"{img_path.stem}.txt"
-        ax.imshow(draw_label_overlay(img_path, lbl_path))
-        ax.set_title(img_path.name, fontsize=9)
-        ax.axis("off")
-    for ax in axes[len(preview_imgs):]:
-        ax.axis("off")
-    plt.tight_layout()
-    plt.show()
-else:
-    print("没有找到可预览图片")
+
+def upload_persisted_exclusions(txt_path: Path, csv_path: Path) -> None:
+    api.upload_file(
+        repo_id=DATASET_REPO,
+        repo_type="dataset",
+        token=HF_TOKEN,
+        path_or_fileobj=str(txt_path),
+        path_in_repo=f"{CLEANING_PREFIX}/excluded_sample_keys.txt",
+        commit_message="update excluded training sample keys",
+    )
+    api.upload_file(
+        repo_id=DATASET_REPO,
+        repo_type="dataset",
+        token=HF_TOKEN,
+        path_or_fileobj=str(csv_path),
+        path_in_repo=f"{CLEANING_PREFIX}/excluded_sample_keys_latest.csv",
+        commit_message="update latest excluded training sample table",
+    )
+
+
+if not COMBINED_MANIFEST_PATH.exists():
+    raise FileNotFoundError(
+        f"Missing {COMBINED_MANIFEST_PATH}. Please run Step 10 before Step 11."
+    )
+
+training_review_df = pd.read_csv(COMBINED_MANIFEST_PATH)
+training_review_df = training_review_df.sort_values(
+    ["split", "source", "source_key"]
+).reset_index(drop=True)
+
+print(f"全量训练集样本数: {len(training_review_df)}")
+print(training_review_df.groupby(["split", "source"]).size().to_string())
+print("")
+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。")
+
+checkboxes: dict[int, widgets.Checkbox] = {}
+cards: list[widgets.Widget] = []
+
+for idx, row in training_review_df.iterrows():
+    img_path = Path(row["image_path"])
+    lbl_path = Path(row["label_path"])
+    if img_path.exists() and lbl_path.exists():
+        img_html = (
+            f"<img src='{overlay_data_url(img_path, lbl_path)}' "
+            "style='width:260px; height:auto; display:block;'>"
+        )
+    else:
+        img_html = "<div style='width:260px;height:160px;background:#eee;'>missing file</div>"
+
+    title = (
+        "<div style='font-size:12px;line-height:1.35;"
+        "font-family:Arial, sans-serif;word-break:break-all;'>"
+        f"<b>{row['split']}</b> | {row['source_key']}<br>"
+        f"{row['dataset_stem']} | class {row['classes']} | objects {row['objects']}"
+        "</div>"
+    )
+    checkbox = widgets.Checkbox(
+        value=False,
+        description="删除这个样本",
+        indent=False,
+        layout=widgets.Layout(width="260px"),
+    )
+    checkboxes[idx] = checkbox
+    cards.append(widgets.VBox(
+        [
+            widgets.HTML(title + img_html),
+            checkbox,
+        ],
+        layout=widgets.Layout(
+            width="280px",
+            border="1px solid #ddd",
+            padding="6px",
+            margin="4px",
+        ),
+    ))
+
+delete_button = widgets.Button(
+    description="删除勾选样本并保存到 HuggingFace",
+    button_style="danger",
+    layout=widgets.Layout(width="280px"),
+)
+summary_label = widgets.HTML(
+    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。"
+)
+delete_output = widgets.Output()
+
+
+def on_delete_clicked(_button):
+    selected_indices = [idx for idx, cb in checkboxes.items() if cb.value]
+    with delete_output:
+        delete_output.clear_output()
+        if not selected_indices:
+            print("没有勾选任何样本。")
+            return
+
+        selected_df = training_review_df.loc[selected_indices].copy()
+        selected_keys = set(selected_df["source_key"].astype(str))
+
+        existing_keys = set(globals().get("EXCLUDED_SAMPLE_KEYS", set()))
+        if "read_exclude_keys_file" in globals():
+            existing_keys |= read_exclude_keys_file(CLEANING_DIR / "excluded_sample_keys.txt")
+        merged_keys = existing_keys | selected_keys
+
+        deleted_files = 0
+        for _, selected_row in selected_df.iterrows():
+            for col in ["image_path", "label_path"]:
+                p = Path(selected_row[col])
+                if p.exists():
+                    p.unlink()
+                    deleted_files += 1
+
+        remaining_df = training_review_df.drop(index=selected_indices).reset_index(drop=True)
+        remaining_df.to_csv(COMBINED_MANIFEST_PATH, index=False, encoding="utf-8-sig")
+
+        txt_path, csv_path = write_persisted_exclusions(merged_keys, selected_df)
+        globals()["EXCLUDED_SAMPLE_KEYS"] = merged_keys
+
+        print(f"当前 Colab 训练集已删除样本: {len(selected_df)}")
+        print(f"删除 image/label 文件数: {deleted_files}")
+        print(f"累计排除 source_key 数: {len(merged_keys)}")
+        print(f"本地清单: {txt_path}")
+        train_labels_now = list((COMBINED_DATASET / "labels/train").glob("*.txt"))
+        val_labels_now = list((COMBINED_DATASET / "labels/val").glob("*.txt"))
+        print(f"剩余 train 图片: {len(list_images(COMBINED_DATASET / 'images/train'))}")
+        print(f"剩余 val 图片: {len(list_images(COMBINED_DATASET / 'images/val'))}")
+        print(f"剩余 train 类别计数: {class_counts(train_labels_now)}")
+        print(f"剩余 val 类别计数: {class_counts(val_labels_now)}")
+
+        try:
+            upload_persisted_exclusions(txt_path, csv_path)
+            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。")
+        except Exception as exc:
+            print("HuggingFace 上传排除清单失败，但当前 Colab 训练集已经删除。")
+            print(repr(exc))
+
+        for idx in selected_indices:
+            checkboxes[idx].disabled = True
+
+
+delete_button.on_click(on_delete_clicked)
+
+grid = widgets.GridBox(
+    cards,
+    layout=widgets.Layout(
+        grid_template_columns="repeat(3, 300px)",
+        grid_gap="8px",
+        align_items="flex-start",
+    ),
+)
+
+display(widgets.VBox([summary_label, delete_button, delete_output, grid]))
 
 
 # %% [markdown]
@@ -978,6 +1188,7 @@ IMG_SIZE = 640                  # YOLO 分割常用图片尺寸。
 BATCH = -1                      # 让 Ultralytics 自动选择显存能承受的 batch。
 PATIENCE = 35                   # 验证集 35 轮不提升就早停。
 RUN_NAME = f"gas_valve_seg_reviewed_{time.strftime('%Y%m%d_%H%M')}"
+AUTO_UPLOAD_TRAINING_RESULTS = True
 
 print("训练参数")
 print(f"  model: {TRAIN_MODEL}")
@@ -987,6 +1198,65 @@ print(f"  imgsz: {IMG_SIZE}")
 print(f"  batch: {BATCH}")
 print(f"  run: {RUN_NAME}")
 
+
+TRAINING_UPLOAD_ITEMS = [
+    "weights/best.pt",
+    "weights/last.pt",
+    "results.png",
+    "results.csv",
+    "args.yaml",
+    "confusion_matrix_normalized.png",
+    "confusion_matrix.png",
+    "BoxP_curve.png",
+    "BoxR_curve.png",
+    "BoxF1_curve.png",
+    "BoxPR_curve.png",
+    "MaskP_curve.png",
+    "MaskR_curve.png",
+    "MaskF1_curve.png",
+    "MaskPR_curve.png",
+]
+
+
+def upload_training_outputs(run_dir: Path, run_name: str) -> None:
+    operations = []
+    for rel in TRAINING_UPLOAD_ITEMS:
+        local_file = run_dir / rel
+        if not local_file.exists():
+            continue
+        operations.append(CommitOperationAdd(
+            path_in_repo=f"runs/{run_name}/{rel}",
+            path_or_fileobj=str(local_file),
+        ))
+        print(f"准备上传: {rel}")
+
+    extra_files = [
+        (YAML_PATH, "dataset.yaml"),
+        (COMBINED_MANIFEST_PATH, "training_manifest.csv"),
+        (CLEANING_DIR / "excluded_sample_keys.txt", "excluded_sample_keys.txt"),
+    ]
+    for local_file, rel in extra_files:
+        if local_file.exists():
+            operations.append(CommitOperationAdd(
+                path_in_repo=f"runs/{run_name}/{rel}",
+                path_or_fileobj=str(local_file),
+            ))
+            print(f"准备上传: {rel}")
+
+    if not operations:
+        print("没有找到可上传的训练结果文件。")
+        return
+
+    api.create_commit(
+        repo_id=DATASET_REPO,
+        repo_type="dataset",
+        token=HF_TOKEN,
+        operations=operations,
+        commit_message=f"Training_results_{run_name}",
+    )
+    print(f"训练结果已上传: https://huggingface.co/datasets/{DATASET_REPO}/tree/main/runs/{run_name}")
+
+
 model = YOLO(TRAIN_MODEL)
 t_start = time.time()
 train_results = model.train(
@@ -1022,6 +1292,14 @@ print(f"训练完成，用时 {elapsed / 60:.1f} 分钟")
 print(f"Run dir: {RUN_DIR}")
 print(f"Best weights: {RUN_DIR / 'weights/best.pt'}")
 
+if AUTO_UPLOAD_TRAINING_RESULTS:
+    print("开始自动上传训练结果到 HuggingFace...")
+    try:
+        upload_training_outputs(RUN_DIR, RUN_NAME)
+    except Exception as exc:
+        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。")
+        print(repr(exc))
+
 
 # %% [markdown]
 # ---
@@ -1061,59 +1339,12 @@ for name in ["results.png", "confusion_matrix_normalized.png", "confusion_matrix
 
 # %%
 # @title Step 14 - 上传训练结果
+# @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。
 
-UPLOAD_TRAINING_RESULTS = True
-UPLOAD_REPO = DATASET_REPO
-UPLOAD_REPO_TYPE = "dataset"
-
-if UPLOAD_TRAINING_RESULTS:
-    upload_items = [
-        "weights/best.pt",
-        "weights/last.pt",
-        "results.png",
-        "results.csv",
-        "args.yaml",
-        "confusion_matrix_normalized.png",
-        "confusion_matrix.png",
-        "BoxP_curve.png",
-        "BoxR_curve.png",
-        "BoxF1_curve.png",
-        "BoxPR_curve.png",
-        "MaskP_curve.png",
-        "MaskR_curve.png",
-        "MaskF1_curve.png",
-        "MaskPR_curve.png",
-    ]
-
-    operations = []
-    for rel in upload_items:
-        local_file = RUN_DIR / rel
-        if not local_file.exists():
-            continue
-        operations.append(CommitOperationAdd(
-            path_in_repo=f"runs/{RUN_NAME}/{rel}",
-            path_or_fileobj=local_file.read_bytes(),
-        ))
-        print(f"Prepared {rel}")
+if "RUN_DIR" not in globals() or "RUN_NAME" not in globals():
+    raise RuntimeError("还没有训练结果。请先运行 Step 12。")
 
-    operations.append(CommitOperationAdd(
-        path_in_repo=f"runs/{RUN_NAME}/dataset.yaml",
-        path_or_fileobj=YAML_PATH.read_bytes(),
-    ))
-
-    if operations:
-        api.create_commit(
-            repo_id=UPLOAD_REPO,
-            repo_type=UPLOAD_REPO_TYPE,
-            token=HF_TOKEN,
-            operations=operations,
-            commit_message=f"Training_results_{RUN_NAME}",
-        )
-        print(f"Uploaded results to https://huggingface.co/datasets/{UPLOAD_REPO}/tree/main/runs/{RUN_NAME}")
-    else:
-        print("No result files found to upload")
-else:
-    print("Skipping result upload")
+upload_training_outputs(RUN_DIR, RUN_NAME)
 
 
 # %% [markdown]