simplify colab notebook parameters notebooks/colab_gas_valve_auto_label_train.py

notebooks/colab_gas_valve_auto_label_train.py

@@ -14,27 +14,46 @@
 # %% [markdown]
 # # Gas + valve auto-labeling and training
 #
-# This notebook expects the HuggingFace dataset repo to contain:
+# 这个 notebook 默认使用 HuggingFace 数据集仓库中的这些内容：
 #
-# - `images/train/` and `labels/train/`: existing gas-cylinder labels
-# - `raw/valve/`: valve images
-# - `raw/gas_extra/`: extra gas-cylinder images
-# - `reviewed_labels/valve/`: reviewed valve polygon labels, preferred for training
-# - `scripts/process_valve_to_polygon.py`: optional valve polygon auto-label script
+# - `images/train/` 和 `labels/train/`：原来的气瓶训练集
+# - `raw/valve/`：阀门原图
+# - `raw/gas_extra/`：补充气瓶原图
+# - `reviewed_labels/valve/`：已经检查过的阀门 polygon 标签，训练时优先使用它
+# - `scripts/process_valve_to_polygon.py`：阀门自动标注脚本，只作为备用实验模式
 #
-# Class mapping:
+# 类别编号：
 #
 # - `0`: `gas_cylinder`
 # - `1`: `valve`
 
 
+# %% [markdown]
+# ## 运行方式和参数说明
+#
+# 我已经按这个项目把默认参数设置好了。正常使用时，从上到下运行每个单元格即可。
+#
+# 你通常只需要填或改这几项：
+#
+# - Step 2 的 `HF_TOKEN`：你的 HuggingFace token，需要有写入权限。
+# - Step 9 的 `EXCLUDE_SAMPLE_KEYS`：审核时发现不想训练的样本，一行写一个。
+# - Step 9 的 `REVIEW_SOURCE`、`REVIEW_OFFSET`、`REVIEW_COUNT`：只控制预览哪一批图片，不影响标签本身。
+#
+# 其它参数我已经固定在代码里，正常训练不用改：
+#
+# - 阀门标签默认使用 HuggingFace 上的 `reviewed_labels/valve`。
+# - 之前失败的整图阀门自动标注模式不会默认参与训练。
+# - 训练默认使用 `yolo11x-seg.pt`、120 轮、图片尺寸 640、自动 batch、训练/验证 8:2 划分。
+# - 生成的标签和训练结果默认上传回同一个数据集仓库。
+
+
 # %% [markdown]
 # ---
-# ## Step 1 - Install dependencies
+# ## Step 1 - 安装依赖
 
 
 # %%
-# @title Step 1 - Install dependencies
+# @title Step 1 - 安装依赖
 
 get_ipython().system(
     "pip install -q ultralytics huggingface_hub opencv-python-headless "
@@ -73,15 +92,15 @@ if torch.cuda.is_available():
 
 # %% [markdown]
 # ---
-# ## Step 2 - HuggingFace authentication
+# ## Step 2 - HuggingFace 登录
 
 
 # %%
-# @title Step 2 - HuggingFace authentication
-# @markdown Token needs write permission if you want to upload generated labels and training results.
+# @title Step 2 - HuggingFace 登录
+# @markdown 这里只填 `HF_TOKEN`。数据集仓库已经固定为 `rick003/11111`。
 
 HF_TOKEN = ""  # @param {type:"string"}
-DATASET_REPO = "rick003/11111"  # @param {type:"string"}
+DATASET_REPO = "rick003/11111"
 
 if not HF_TOKEN:
     from getpass import getpass
@@ -101,19 +120,20 @@ print(f"Dataset: https://huggingface.co/datasets/{DATASET_REPO}")
 
 # %% [markdown]
 # ---
-# ## Step 3 - Download required files
+# ## Step 3 - 下载所需文件
 
 
 # %%
-# @title Step 3 - Download required files
-# @markdown Downloads only the files needed for joint labeling/training.
+# @title Step 3 - 下载所需文件
+# @markdown 下载原气瓶标签、新图片、已审核阀门标签和脚本。
+# @markdown 默认强制重新下载，保证 Colab 拿到 HuggingFace 上的最新文件。
 
-DATASET_LOCAL = "/content/hf_11111"  # @param {type:"string"}
+DATASET_LOCAL = "/content/hf_11111"
 RAW_VALVE_PREFIX = "raw/valve"
 RAW_GAS_EXTRA_PREFIX = "raw/gas_extra"
 REVIEWED_VALVE_LABEL_PREFIX = "reviewed_labels/valve"
 PROCESS_SCRIPT_REPO_PATH = "scripts/process_valve_to_polygon.py"
-FORCE_REDOWNLOAD = False  # @param {type:"boolean"}
+FORCE_REDOWNLOAD = True
 
 base = Path(DATASET_LOCAL)
 if FORCE_REDOWNLOAD and base.exists():
@@ -144,7 +164,7 @@ RAW_GAS_EXTRA_DIR = base / RAW_GAS_EXTRA_PREFIX
 REVIEWED_VALVE_LABEL_DIR = base / REVIEWED_VALVE_LABEL_PREFIX
 PROCESS_SCRIPT_LOCAL = base / PROCESS_SCRIPT_REPO_PATH
 
-print("Download summary")
+print("下载结果统计")
 for p in [
     base / "images/train",
     base / "labels/train",
@@ -161,18 +181,18 @@ if not RAW_GAS_EXTRA_DIR.exists():
     raise FileNotFoundError(f"Missing {RAW_GAS_EXTRA_PREFIX} in HF dataset")
 if not PROCESS_SCRIPT_LOCAL.exists():
     print(
-        f"Warning: missing {PROCESS_SCRIPT_REPO_PATH}. "
-        "Reviewed valve labels can still be used; auto_no_bbox mode will fail."
+        f"警告：缺少 {PROCESS_SCRIPT_REPO_PATH}。"
+        "已审核阀门标签仍然可以使用，但 auto_no_bbox 实验模式不能运行。"
     )
 
 
 # %% [markdown]
 # ---
-# ## Step 4 - Shared helpers
+# ## Step 4 - 公共辅助函数
 
 
 # %%
-# @title Step 4 - Shared helpers
+# @title Step 4 - 公共辅助函数
 
 IMG_EXTS = {".png", ".jpg", ".jpeg", ".bmp", ".webp"}
 WORK_DIR = Path("/content/gas_valve_work")
@@ -264,11 +284,11 @@ print(f"Raw gas-extra images: {len(list_images(RAW_GAS_EXTRA_DIR))}")
 
 # %% [markdown]
 # ---
-# ## Step 5 - Gas-cylinder auto-labeler
+# ## Step 5 - 气瓶自动标注器
 
 
 # %%
-# @title Step 5 - Gas-cylinder auto-labeler
+# @title Step 5 - 气瓶自动标注器
 
 
 class GasCylinderSegmenter:
@@ -377,17 +397,17 @@ def auto_label_gas_folder(
 
 # %% [markdown]
 # ---
-# ## Step 6 - Auto-label new gas data
+# ## Step 6 - 自动标注补充气瓶数据
 
 
 # %%
-# @title Step 6 - Auto-label new gas data
-# @markdown Gas labels are generated for both extra gas images and valve images.
-# @markdown The valve-image gas labels prevent visible gas cylinders from being treated as background.
+# @title Step 6 - 自动标注补充气瓶数据
+# @markdown 会给补充气瓶图片生成气瓶标签，也会给阀门图片中的气瓶区域生成气瓶标签。
+# @markdown 这样训练时阀门图片里的气瓶不会被当成背景。
 
-GAS_LABEL_MODEL = "yolo11x-seg.pt"  # @param ["yolo11n-seg.pt","yolo11m-seg.pt","yolo11x-seg.pt"]
-REGENERATE_GAS_LABELS = False  # @param {type:"boolean"}
-GAS_LABEL_LIMIT = 0  # @param {type:"integer"}
+GAS_LABEL_MODEL = "yolo11x-seg.pt"  # 大模型，标注更稳，但比 n/m 慢。
+REGENERATE_GAS_LABELS = False       # 同一个 Colab 会话里强制重标时才改 True。
+GAS_LABEL_LIMIT = 0                 # 0 表示处理全部图片。
 
 GAS_EXTRA_LABEL_DIR = LABEL_WORK_DIR / "gas_extra"
 VALVE_GAS_LABEL_DIR = LABEL_WORK_DIR / "valve_gas"
@@ -411,19 +431,19 @@ valve_gas_stats = auto_label_gas_folder(
 
 # %% [markdown]
 # ---
-# ## Step 7 - Auto-label valve polygons
+# ## Step 7 - 准备阀门 polygon 标签
 
 
 # %%
-# @title Step 7 - Prepare valve polygon labels
-# @markdown `reviewed_hf` uses reviewed labels from HuggingFace and is the default for training.
-# @markdown `auto_no_bbox` runs the whole-image heuristic labeler; use it only for experiments and manual review.
+# @title Step 7 - 准备阀门 polygon 标签
+# @markdown 默认使用 HuggingFace 上已经审核过的阀门标签。
+# @markdown `auto_no_bbox` 是之前失败的整图启发式自动标注，只保留给实验排查，不建议训练。
 
-VALVE_LABEL_SOURCE = "reviewed_hf"  # @param ["reviewed_hf","auto_no_bbox"]
-VALVE_LABEL_MODEL = "yolo11x-seg.pt"  # @param ["yolo11n-seg.pt","yolo11m-seg.pt","yolo11x-seg.pt"]
-REGENERATE_VALVE_LABELS = False  # @param {type:"boolean"}
-SAVE_VALVE_VISUALIZATIONS = False  # @param {type:"boolean"}
-VALVE_LABEL_LIMIT = 0  # @param {type:"integer"}
+VALVE_LABEL_SOURCE = "reviewed_hf"  # 训练保持这个默认值。
+VALVE_LABEL_MODEL = "yolo11x-seg.pt"
+REGENERATE_VALVE_LABELS = False
+SAVE_VALVE_VISUALIZATIONS = False
+VALVE_LABEL_LIMIT = 0
 
 VALVE_LABEL_DIR = LABEL_WORK_DIR / "valve"
 VALVE_VIS_DIR = VIS_WORK_DIR / "valve"
@@ -438,8 +458,8 @@ if VALVE_LABEL_SOURCE == "reviewed_hf":
     if not reviewed_labels:
         raise FileNotFoundError(
             f"No reviewed valve labels found at {REVIEWED_VALVE_LABEL_PREFIX}. "
-            "Upload reviewed YOLO-seg valve txt files there, or switch "
-            "VALVE_LABEL_SOURCE to auto_no_bbox for an experimental run."
+            "请先把审核过的 YOLO-seg 阀门 txt 标签上传到这里；"
+            "或者把 VALVE_LABEL_SOURCE 改成 auto_no_bbox 做实验性自动标注。"
         )
     for src in reviewed_labels:
         shutil.copy2(src, VALVE_LABEL_DIR / src.name)
@@ -451,7 +471,7 @@ elif VALVE_LABEL_SOURCE == "auto_no_bbox":
     if not PROCESS_SCRIPT_LOCAL.exists():
         raise FileNotFoundError(
             f"Missing {PROCESS_SCRIPT_REPO_PATH}. Upload process_valve_to_polygon.py "
-            "before using auto_no_bbox mode."
+            "后才能使用 auto_no_bbox 模式。"
         )
     device = "0" if torch.cuda.is_available() else "cpu"
     cmd = [
@@ -492,15 +512,15 @@ print(f"Valve non-empty labels: {sum(1 for p in valve_labels if p.stat().st_size
 
 # %% [markdown]
 # ---
-# ## Step 8 - Optionally upload generated labels
+# ## Step 8 - 上传本次标签
 
 
 # %%
-# @title Step 8 - Optionally upload generated labels
-# @markdown Uploads generated labels under `auto_labels/gas_valve_v1/`.
+# @title Step 8 - 上传本次标签
+# @markdown 默认把本次生成或选用的标签上传到 `auto_labels/gas_valve_v1/`。
 
-UPLOAD_GENERATED_LABELS = True  # @param {type:"boolean"}
-GENERATED_LABELS_REPO_DIR = "auto_labels/gas_valve_v1"  # @param {type:"string"}
+UPLOAD_GENERATED_LABELS = True
+GENERATED_LABELS_REPO_DIR = "auto_labels/gas_valve_v1"
 
 if UPLOAD_GENERATED_LABELS:
     print(f"Uploading generated labels to {GENERATED_LABELS_REPO_DIR}")
@@ -512,32 +532,37 @@ if UPLOAD_GENERATED_LABELS:
         path_in_repo=GENERATED_LABELS_REPO_DIR,
         commit_message="upload_generated_gas_valve_labels",
     )
-    print("Generated labels uploaded")
+    print("本次标签已上传")
 else:
     print("Skipping generated label upload")
 
 
 # %% [markdown]
 # ---
-# ## Step 9 - Review and cleaning controls
+# ## Step 9 - 审核和清理控制
 
 
 # %%
-# @title Step 9 - Review and cleaning controls
-# @markdown The manifest uses `source/stem` keys, for example `valve/Pic_阀门_60` or `gas_extra/Pic_气瓶_125`.
-# @markdown Put one key per line in `EXCLUDE_SAMPLE_KEYS` to remove those samples from the training dataset.
+# @title Step 9 - 审核和清理控制
+# @markdown 这一步只做预览和清单统计，不会自动修改标签。
+# @markdown `source_key` 示例：`valve/Pic_阀门_60`、`gas_extra/Pic_气瓶_125`、`old_gas/Pic_gas_1`。
+# @markdown 把不想参与训练的 `source_key` 填到 `EXCLUDE_SAMPLE_KEYS`，一行一个。
 
 import pandas as pd
 
 EXCLUDE_SAMPLE_KEYS = """
 """  # @param {type:"string"}
 
+# 预览哪一类样本：阀门、补充气瓶、旧气瓶，或者全部。
 REVIEW_SOURCE = "valve"  # @param ["valve", "gas_extra", "old_gas", "all"]
+# 从第几条开始预览。想看下一页，就加上 REVIEW_COUNT。
 REVIEW_OFFSET = 0  # @param {type:"integer"}
+# 本次显示多少张预览图。
 REVIEW_COUNT = 9  # @param {type:"integer"}
 
-DELETE_EXCLUDED_FROM_HF = False  # @param {type:"boolean"}
-CONFIRM_HF_DELETE = ""  # @param {type:"string"}
+# 正常清理只需要排除训练样本，不直接删 HuggingFace 远端文件。
+DELETE_EXCLUDED_FROM_HF = False
+CONFIRM_HF_DELETE = ""
 
 
 def parse_exclude_keys(text: str) -> set[str]:
@@ -730,25 +755,26 @@ if DELETE_EXCLUDED_FROM_HF:
 
 # %% [markdown]
 # ---
-# ## Step 10 - Build the two-class training dataset
+# ## Step 10 - 构建两类训练集
 
 
 # %%
-# @title Step 10 - Build the two-class training dataset
+# @title Step 10 - 构建两类训练集
 
 COMBINED_DATASET = Path("/content/gas_valve_dataset")
-VAL_RATIO = 0.20  # @param {type:"number"}
-RANDOM_SEED = 42  # @param {type:"integer"}
-INCLUDE_OLD_GAS_DATA = True  # @param {type:"boolean"}
-ALLOW_AUTO_NO_BBOX_VALVE_TRAINING = False  # @param {type:"boolean"}
+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。
+ALLOW_AUTO_NO_BBOX_VALVE_TRAINING = False
 
 if (
     globals().get("VALVE_LABEL_SOURCE") == "auto_no_bbox"
     and not ALLOW_AUTO_NO_BBOX_VALVE_TRAINING
 ):
     raise RuntimeError(
-        "Valve labels came from auto_no_bbox mode. Review Step 9 first. "
-        "Only set ALLOW_AUTO_NO_BBOX_VALVE_TRAINING=True if the overlays are correct."
+        "当前阀门标签来自 auto_no_bbox 实验模式。请先在 Step 9 审核。"
+        "只有确认可视化结果正确后，才可以把 "
+        "ALLOW_AUTO_NO_BBOX_VALVE_TRAINING 改成 True。"
     )
 
 if COMBINED_DATASET.exists():
@@ -869,7 +895,7 @@ YAML_PATH.write_text(
 train_label_files = list((COMBINED_DATASET / "labels/train").glob("*.txt"))
 val_label_files = list((COMBINED_DATASET / "labels/val").glob("*.txt"))
 
-print("Combined dataset built")
+print("混合训练集已构建")
 print(f"  train images: {len(list_images(COMBINED_DATASET / 'images/train'))}")
 print(f"  train labels: {len(train_label_files)}")
 print(f"  val images: {len(list_images(COMBINED_DATASET / 'images/val'))}")
@@ -881,11 +907,11 @@ print(YAML_PATH.read_text(encoding="utf-8"))
 
 # %% [markdown]
 # ---
-# ## Step 11 - Visual sanity check
+# ## Step 11 - 可视化抽查训练集
 
 
 # %%
-# @title Step 11 - Visual sanity check
+# @title Step 11 - 可视化抽查训练集
 
 def draw_label_overlay(img_path: Path, label_path: Path) -> np.ndarray:
     img = read_image(img_path)
@@ -932,27 +958,28 @@ if preview_imgs:
     plt.tight_layout()
     plt.show()
 else:
-    print("No preview images found")
+    print("没有找到可预览图片")
 
 
 # %% [markdown]
 # ---
-# ## Step 12 - Train YOLO-seg
+# ## Step 12 - 训练 YOLO-seg
 
 
 # %%
-# @title Step 12 - Train YOLO-seg
+# @title Step 12 - 训练 YOLO-seg
+# @markdown 训练参数已经按本项目设置好：高精度模型、120 轮、图片尺寸 640、自动 batch。
 
 from ultralytics import YOLO
 
-TRAIN_MODEL = "yolo11x-seg.pt"  # @param ["yolo11n-seg.pt","yolo11m-seg.pt","yolo11x-seg.pt"]
-EPOCHS = 120  # @param {type:"integer"}
-IMG_SIZE = 640  # @param {type:"integer"}
-BATCH = -1  # @param {type:"integer"}
-PATIENCE = 35  # @param {type:"integer"}
-RUN_NAME = "gas_valve_seg_v1"  # @param {type:"string"}
+TRAIN_MODEL = "yolo11x-seg.pt"  # 优先保证精度。
+EPOCHS = 120                    # 最多训练 120 轮，早停可能提前结束。
+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')}"
 
-print("Training config")
+print("训练参数")
 print(f"  model: {TRAIN_MODEL}")
 print(f"  data: {YAML_PATH}")
 print(f"  epochs: {EPOCHS}")
@@ -991,18 +1018,18 @@ train_results = model.train(
 
 RUN_DIR = Path(train_results.save_dir)
 elapsed = time.time() - t_start
-print(f"Training finished in {elapsed / 60:.1f} min")
+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'}")
 
 
 # %% [markdown]
 # ---
-# ## Step 13 - Review results
+# ## Step 13 - 查看训练结果
 
 
 # %%
-# @title Step 13 - Review results
+# @title Step 13 - 查看训练结果
 
 import pandas as pd
 import matplotlib.image as mpimg
@@ -1029,15 +1056,15 @@ for name in ["results.png", "confusion_matrix_normalized.png", "confusion_matrix
 
 # %% [markdown]
 # ---
-# ## Step 14 - Upload training results
+# ## Step 14 - 上传训练结果
 
 
 # %%
-# @title Step 14 - Upload training results
+# @title Step 14 - 上传训练结果
 
-UPLOAD_TRAINING_RESULTS = True  # @param {type:"boolean"}
-UPLOAD_REPO = DATASET_REPO  # @param {type:"string"}
-UPLOAD_REPO_TYPE = "dataset"  # @param ["dataset", "model"]
+UPLOAD_TRAINING_RESULTS = True
+UPLOAD_REPO = DATASET_REPO
+UPLOAD_REPO_TYPE = "dataset"
 
 if UPLOAD_TRAINING_RESULTS:
     upload_items = [