训练指南
=========

基本训练
--------

EMCAD 项目使用 Trainer 类封装完整的训练流程。

.. code-block:: python

   from src.utils.config import Config
   from src.core.networks import EMCADNet
   from src.utils.dataloader import get_loader
   from src.utils.trainer import Trainer

   config = Config.from_yaml("configs/default.yaml")

   train_loader = get_loader(
       config.dataset.root_path,
       config.dataset.list_dir,
       config.dataset.volume_path,
       split="train",
       img_size=config.dataset.img_size,
       batch_size=config.training.batch_size,
   )

   model = EMCADNet(
       num_classes=config.model.num_classes,
       encoder=config.model.encoder,
       kernel_sizes=config.model.kernel_sizes,
       expansion_factor=config.model.expansion_factor,
   )

   trainer = Trainer(
       model=model,
       config=config,
       train_loader=train_loader,
   )

   trainer.train(epochs=config.training.max_epochs)

自定义训练
----------

早停配置
~~~~~~~~

在配置文件中启用早停：

.. code-block:: yaml

   training:
     early_stopping:
       enabled: true
       patience: 50
       min_delta: 0.001

学习率调度
~~~~~~~~~~

支持多种学习率调度器：

.. code-block:: yaml

   training:
     lr_scheduler:
       type: "CosineAnnealingLR"
       T_max: 300
       eta_min: 1e-6

监督策略
--------

EMCAD 支持三种监督策略：

1. **mutation**: 多尺度监督
2. **deep_supervision**: 深度监督
3. **last_layer**: 仅最后一层监督

.. code-block:: yaml

   training:
     supervision: "mutation"