fix(configs): 修复 resnet18 训练配置无法构建，并补上对应的评估配置

- configs/train/resnet18.yaml 此前直接报错：它同时继承 tiny_conv 与 resnet18 的模型
  配置，深度合并后 backbone 混了两边字段。在 configs/model/resnet18.yaml 的 backbone
  上加 _replace_: true 整块替换
- _replace_ 标记此前不生效：_load_raw 对每个 base 都会先 deep_merge({}, data)，
  标记在参与替换判断之前就被剥掉了。改为合并期间保留、load_config 末尾统一剥除
- 新增 configs/eval/resnet18.yaml：评估必须与训练用同一套模型配置，否则构建出的
  骨架与权重对不上，而此前只有 baseline 一个评估配置
- 新增参数化测试：仓库里每个配置都要能真的构建出组件（片段自动跳过），
  且训练/评估配置对的 model 段必须一致
- README 补充骨架与配置对的对应表

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

README.md

@@ -89,6 +89,24 @@ uv run python scripts/export_onnx.py --config configs/train/baseline.yaml \
     --checkpoint runs/baseline/best.pth --output runs/baseline/model.onnx
 ```
 
+换骨架用对应的配置对（训练与评估必须成对，否则构建出的骨架与权重对不上）：
+
+| 骨架 | 训练 | 评估 | work_dir |
+| --- | --- | --- | --- |
+| `tiny_conv`（默认） | `configs/train/baseline.yaml` | `configs/eval/baseline.yaml` | `runs/baseline` |
+| `resnet18` | `configs/train/resnet18.yaml` | `configs/eval/resnet18.yaml` | `runs/resnet18` |
+
+```bash
+uv run python train.py --config configs/train/resnet18.yaml
+uv run python eval.py  --config configs/eval/resnet18.yaml
+uv run python scripts/export_onnx.py --config configs/train/resnet18.yaml \
+    --checkpoint runs/resnet18/best.pth --output runs/resnet18/model.onnx
+```
+
+要加别的骨架（mobilenet / shufflenet），照 [configs/model/resnet18.yaml](configs/model/resnet18.yaml)
+复制一份改 `arch` 即可——注意保留里面的 `_replace_: true`，否则 `tiny_conv` 的
+`widths/depths/block` 会残留下来传给新骨架。
+
 产物写在 `work_dir`（默认 `runs/<name>/`）：`train.log`、`config.yaml`、`last.pth`、`best.pth`、`swanlog/`。
 
 ## 多类 / 单类分类

configs/eval/resnet18.yaml

@@ -0,0 +1,13 @@
+# resnet18 版本的评估配置。
+# 必须和训练时用的模型配置一致，否则构建出来的骨架与权重对不上。
+_base_:
+  - ../train/resnet18.yaml
+
+device: auto
+checkpoint: runs/resnet18/best.pth
+
+eval:
+  split: val
+  score_threshold: 0.3
+  iou_threshold: 0.5
+  topk: 50

configs/model/resnet18.yaml

@@ -2,6 +2,9 @@
 model:
   name: defect_net
   backbone:
+    # 整块替换：不加这行的话，从 tiny_conv 继承来的 widths/depths/block 会残留，
+    # 传给 TorchvisionBackbone 直接报错。见 utils/config.py::deep_merge
+    _replace_: true
     name: torchvision
     arch: resnet18
     pretrained: true

tests/test_config.py

@@ -1,3 +1,5 @@
+from pathlib import Path
+
 import pytest
 
 pytest.importorskip("yaml")
@@ -83,3 +85,52 @@ def test_registry_reports_incompatible_fields_actionably():
 
     with pytest.raises(TypeError, match=r"不接受的字段.*widths"):
         BACKBONES.build({"name": "torchvision", "arch": "resnet18", "widths": [32]})
+
+
+@pytest.mark.parametrize(
+    "config_path",
+    sorted(str(p) for p in Path("configs").rglob("*.yaml")),
+)
+def test_every_shipped_config_builds_its_components(config_path):
+    """仓库里每个配置都必须能真的构建出组件。
+
+    回归测试：configs/train/resnet18.yaml 曾经无法使用——它同时继承 tiny_conv 与
+    resnet18 的模型配置，合并后 backbone 混了两边的字段，构建时直接报错。
+    """
+    torch = pytest.importorskip("torch")
+    from datasets.yolo import load_classes
+    from models import build_model
+    from models.registry import BACKBONES
+
+    cfg = load_config(config_path)
+
+    if "model" in cfg:
+        if "classes" not in cfg:
+            # configs/model/*.yaml 是片段，只能被 _base_ 组合，单独用没有 classes
+            pytest.skip(f"{config_path} 是配置片段，不含 classes")
+        model_cfg = dict(cfg["model"])
+        model_cfg.setdefault("num_classes", len(load_classes(cfg["classes"])))
+        # 单测里不下载 ImageNet 权重
+        if model_cfg.get("backbone", {}).get("name") == "torchvision":
+            model_cfg["backbone"] = {**model_cfg["backbone"], "pretrained": False}
+        model = build_model(model_cfg)
+        with torch.no_grad():
+            out = model(torch.randn(1, 3, 64, 64))
+        assert "heatmap" in out
+
+    simclr_backbone = cfg.get_path("simclr.backbone")
+    if simclr_backbone:
+        backbone_cfg = dict(simclr_backbone)
+        if backbone_cfg.get("name") == "torchvision":
+            backbone_cfg["pretrained"] = False
+        backbone = BACKBONES.build(backbone_cfg)
+        assert backbone.out_channels > 0
+
+
+def test_eval_config_matches_its_training_backbone():
+    """评估配置构建出的骨架必须与训练配置一致，否则权重根本加载不上。"""
+    for train_cfg, eval_cfg in [
+        ("configs/train/baseline.yaml", "configs/eval/baseline.yaml"),
+        ("configs/train/resnet18.yaml", "configs/eval/resnet18.yaml"),
+    ]:
+        assert load_config(train_cfg)["model"] == load_config(eval_cfg)["model"], eval_cfg

utils/config.py

@@ -64,21 +64,25 @@ def deep_merge(base: Mapping, override: Mapping) -> dict:
     merged = dict(base)
     for key, value in override.items():
         if isinstance(value, Mapping) and value.get(REPLACE_KEY):
-            merged[key] = _strip_replace_key(value)
+            merged[key] = dict(value)  # 整块替换；标记本身留到最后统一剥掉
         elif key in merged and isinstance(merged[key], Mapping) and isinstance(value, Mapping):
             merged[key] = deep_merge(merged[key], value)
         else:
-            merged[key] = _strip_replace_key(value) if isinstance(value, Mapping) else value
+            merged[key] = value
     return merged
 
 
-def _strip_replace_key(value: Mapping) -> dict:
-    """去掉 `_replace_` 标记本身，避免它漏进组件的构造参数。"""
-    return {
-        k: (_strip_replace_key(v) if isinstance(v, Mapping) else v)
-        for k, v in value.items()
-        if k != REPLACE_KEY
-    }
+def strip_replace_keys(value: Any) -> Any:
+    """剥掉所有 `_replace_` 标记，避免它漏进组件的构造参数。
+
+    必须等**全部合并做完**再剥：`_load_raw` 对每个 base 都会先 `deep_merge({}, data)`
+    一次，若在那时就剥掉，标记还没来得及参与真正的替换判断就没了。
+    """
+    if isinstance(value, Mapping):
+        return {k: strip_replace_keys(v) for k, v in value.items() if k != REPLACE_KEY}
+    if isinstance(value, list):
+        return [strip_replace_keys(v) for v in value]
+    return value
 
 
 def _load_raw(path: Path, seen: set[Path]) -> dict:
@@ -132,5 +136,6 @@ def apply_overrides(cfg: dict, overrides: Iterable[str]) -> dict:
 def load_config(path: str | Path, overrides: Iterable[str] = ()) -> Config:
     cfg = _load_raw(Path(path), set())
     cfg = apply_overrides(cfg, overrides)
+    cfg = strip_replace_keys(cfg)
     cfg.setdefault("config_path", str(Path(path).resolve()))
     return Config(cfg)