feat(engine): 训练评估循环、CenterNet 损失与指标

- trainer: AMP、梯度裁剪、cosine/step 调度、last/best 检查点，按 monitor 指标挑选 best
- losses: penalty-reduced focal loss（heatmap）+ 掩码 L1（wh/offset）+ 分类损失
  （多类 BCE / 单类 CE，随 label_mode 切换）
- metrics: 定位 IoU 匹配下的 P/R/F1 与 mean IoU；分类多标签 micro-F1 与单标签 accuracy，
  两种模式都输出 cls_f1，故 train.monitor 无需跟着改
- optim: BatchNorm 与 bias 不做 weight decay
- visualizer: SwanLab 记录损失/学习率/验证指标/显存峰值，未安装或离线时降级为警告

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

engine/evaluator.py

@@ -0,0 +1,87 @@
+"""评估：跑一遍验证集，输出分类与定位指标。"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import torch
+from torch import nn
+from torch.utils.data import DataLoader
+
+from engine.metrics import AverageMeter, DetectionMetric, MultiLabelMetric, SingleLabelMetric
+from models.postprocess import decode_heatmap
+from utils.labels import MULTILABEL
+from utils.logging import get_logger
+
+logger = get_logger("engine.evaluator")
+
+
+class Evaluator:
+    def __init__(
+        self,
+        model: nn.Module,
+        dataloader: DataLoader,
+        device: torch.device,
+        criterion: nn.Module | None = None,
+        score_threshold: float = 0.3,
+        iou_threshold: float = 0.5,
+        topk: int = 50,
+    ) -> None:
+        self.model = model
+        self.dataloader = dataloader
+        self.device = device
+        self.criterion = criterion
+        self.score_threshold = score_threshold
+        self.topk = topk
+        self.det_metric = DetectionMetric(iou_threshold)
+        # 分类指标按模型的 label_mode 选择：多类用多标签指标，单类用互斥分类指标。
+        self.label_mode = getattr(model, "label_mode", MULTILABEL)
+        if self.label_mode == MULTILABEL:
+            self.cls_metric: MultiLabelMetric | SingleLabelMetric = MultiLabelMetric()
+        else:
+            num_outputs = model.cls_head.num_outputs if model.cls_head is not None else 1
+            self.cls_metric = SingleLabelMetric(num_outputs)
+
+    @torch.no_grad()
+    def evaluate(self) -> dict[str, float]:
+        self.model.eval()
+        self.det_metric.reset()
+        self.cls_metric.reset()
+        loss_meter = AverageMeter()
+
+        stride = getattr(self.model, "out_stride", 4)
+        for batch in self.dataloader:
+            images = batch["images"].to(self.device, non_blocking=True)
+            outputs = self.model(images)
+
+            if self.criterion is not None:
+                batch_on_device = _to_device(batch, self.device)
+                losses = self.criterion(outputs, batch_on_device)
+                loss_meter.update(float(losses["total"]), images.shape[0])
+
+            detections = decode_heatmap(
+                outputs["heatmap"],
+                outputs["wh"],
+                outputs["offset"],
+                stride=stride,
+                topk=self.topk,
+                score_threshold=self.score_threshold,
+            )
+            for i, det in enumerate(detections):
+                self.det_metric.update(det["boxes"], det["labels"], batch["boxes"][i], batch["labels"][i])
+
+            if "logits" in outputs:
+                self.cls_metric.update(outputs["logits"], batch["image_labels"])
+
+        metrics: dict[str, float] = {**self.det_metric.compute(), **self.cls_metric.compute()}
+        if self.criterion is not None:
+            metrics["loss"] = loss_meter.avg
+        return metrics
+
+
+def _to_device(batch: dict, device: torch.device) -> dict:
+    moved: dict[str, Any] = dict(batch)
+    moved["images"] = batch["images"].to(device, non_blocking=True)
+    moved["target"] = {k: v.to(device, non_blocking=True) for k, v in batch["target"].items()}
+    moved["image_labels"] = batch["image_labels"].to(device, non_blocking=True)
+    return moved