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/optim.py

@@ -0,0 +1,61 @@
+"""优化器与学习率调度。"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from typing import Any
+
+import torch
+from torch import nn
+
+
+def build_optimizer(model: nn.Module, cfg: Mapping[str, Any]) -> torch.optim.Optimizer:
+    """`cfg` 形如 `{name: sgd, lr: 0.01, momentum: 0.9, weight_decay: 1e-4}`。
+
+    BatchNorm 与 bias 不做 weight decay——小数据集上这项影响明显。
+    """
+    cfg = dict(cfg)
+    name = str(cfg.pop("name", "adamw")).lower()
+    weight_decay = float(cfg.pop("weight_decay", 0.0))
+
+    decay, no_decay = [], []
+    for module in model.modules():
+        for param_name, param in module.named_parameters(recurse=False):
+            if not param.requires_grad:
+                continue
+            if param_name == "bias" or isinstance(module, nn.modules.batchnorm._BatchNorm):
+                no_decay.append(param)
+            else:
+                decay.append(param)
+    groups = [
+        {"params": decay, "weight_decay": weight_decay},
+        {"params": no_decay, "weight_decay": 0.0},
+    ]
+
+    if name == "sgd":
+        cfg.setdefault("momentum", 0.9)
+        cfg.setdefault("nesterov", True)
+        return torch.optim.SGD(groups, **cfg)
+    if name == "adam":
+        return torch.optim.Adam(groups, **cfg)
+    if name == "adamw":
+        return torch.optim.AdamW(groups, **cfg)
+    raise ValueError(f"未知优化器: {name}")
+
+
+def build_scheduler(optimizer: torch.optim.Optimizer, cfg: Mapping[str, Any] | None, epochs: int):
+    """返回按 epoch 步进的 scheduler；`cfg` 为空则不使用。"""
+    if not cfg:
+        return None
+    cfg = dict(cfg)
+    name = str(cfg.pop("name", "cosine")).lower()
+    if name == "cosine":
+        cfg.setdefault("T_max", epochs)
+        return torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, **cfg)
+    if name == "step":
+        cfg.setdefault("step_size", max(epochs // 3, 1))
+        return torch.optim.lr_scheduler.StepLR(optimizer, **cfg)
+    if name == "multistep":
+        cfg.setdefault("milestones", [int(epochs * 0.6), int(epochs * 0.85)])
+        return torch.optim.lr_scheduler.MultiStepLR(optimizer, **cfg)
+    raise ValueError(f"未知调度器: {name}")