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/losses.py

@@ -0,0 +1,79 @@
+"""损失函数。"""
+
+from __future__ import annotations
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from utils.labels import MULTILABEL, check_label_mode
+
+
+def gaussian_focal_loss(
+    pred_logits: torch.Tensor,
+    gt_heatmap: torch.Tensor,
+    alpha: float = 2.0,
+    beta: float = 4.0,
+    eps: float = 1e-4,
+) -> torch.Tensor:
+    """CenterNet 的 penalty-reduced focal loss，按正样本数归一化。"""
+    pred = pred_logits.sigmoid().clamp(eps, 1 - eps)
+    pos_mask = gt_heatmap.eq(1.0).float()
+    neg_mask = 1.0 - pos_mask
+
+    pos_loss = -((1 - pred) ** alpha) * torch.log(pred) * pos_mask
+    neg_loss = -((1 - gt_heatmap) ** beta) * (pred**alpha) * torch.log(1 - pred) * neg_mask
+
+    num_pos = pos_mask.sum().clamp(min=1.0)
+    return (pos_loss.sum() + neg_loss.sum()) / num_pos
+
+
+def masked_l1_loss(pred: torch.Tensor, target: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
+    """只在中心点位置计算 L1。mask: (B, H, W)。"""
+    mask = mask.unsqueeze(1).expand_as(pred)
+    loss = F.l1_loss(pred * mask, target * mask, reduction="sum")
+    return loss / mask.sum().clamp(min=1.0)
+
+
+class DetectionLoss(nn.Module):
+    """heatmap + wh + offset (+ 图像级分类) 的加权和。
+
+    分类分支按 `label_mode` 切换：
+        multilabel（多类）—— 多热标签，逐类 BCE；
+        multiclass（单类）—— 互斥下标，softmax 交叉熵。
+    """
+
+    def __init__(
+        self,
+        heatmap_weight: float = 1.0,
+        wh_weight: float = 0.1,
+        offset_weight: float = 1.0,
+        cls_weight: float = 1.0,
+        label_mode: str = MULTILABEL,
+    ) -> None:
+        super().__init__()
+        self.label_mode = check_label_mode(label_mode)
+        self.heatmap_weight = heatmap_weight
+        self.wh_weight = wh_weight
+        self.offset_weight = offset_weight
+        self.cls_weight = cls_weight
+
+    def forward(self, outputs: dict, batch: dict) -> dict[str, torch.Tensor]:
+        target = batch["target"]
+        losses = {
+            "heatmap": self.heatmap_weight * gaussian_focal_loss(outputs["heatmap"], target["heatmap"]),
+            "wh": self.wh_weight * masked_l1_loss(outputs["wh"], target["wh"], target["mask"]),
+            "offset": self.offset_weight
+            * masked_l1_loss(outputs["offset"], target["offset"], target["mask"]),
+        }
+        if "logits" in outputs and self.cls_weight > 0:
+            losses["cls"] = self.cls_weight * self._classification_loss(
+                outputs["logits"], batch["image_labels"]
+            )
+        losses["total"] = sum(losses.values())
+        return losses
+
+    def _classification_loss(self, logits: torch.Tensor, targets: torch.Tensor) -> torch.Tensor:
+        if self.label_mode == MULTILABEL:
+            return F.binary_cross_entropy_with_logits(logits, targets.float())
+        return F.cross_entropy(logits, targets.long())