best_practices

# 最佳实践 - PyTorch Lightning ## 代码组织 ### 1. 将研究与工程分离 **推荐：** python class MyModel(L.LightningModule): # 研究代码（模型的功能） def training_step(self, batch, batch_idx): loss = self.compute_loss(batch) return loss # 工程代码（如何训练） - 在 Trainer 中 trainer = L.Trainer( max_epochs=100, accelerator="gpu", devices=4, strategy="ddp" ) **不推荐：** python # 混合了研究和工程逻辑 class MyModel(L.LightningModule): def training_step(self, batch, batch_idx): loss = self.compute_loss(batch) # 不要手动进行设备管理 loss = loss.cuda() # 不要手动执行优化器步骤（除非是手动优化） self.optimizer.zero_grad() loss.backward() self.optimizer.step() return loss ### 2. 使用 LightningDataModule **推荐：** python class MyDataModule(L.LightningDataModule): def __init__(self, data_dir, batch_size): super().__init__() self.data_dir = data_dir self.batch_size = batch_size def prepare_data(self): # 下载数据一次 download_data(self.data_dir) def setup(self, stage): # 每个进程加载数据 self.train_dataset = MyDataset(self.data_dir, split='train') self.val_dataset = MyDataset(self.data_dir, split='val') def train_dataloader(self): return DataLoader(self.train_dataset, batch_size=self.batch_size, shuffle=True) # 可重用且易于共享 dm = MyDataModule("./data", batch_size=32) trainer.fit(model, datamodule=dm) **不推荐：** python # 分散的数据逻辑 train_dataset = load_data() val_dataset = load_data() train_loader = DataLoader(train_dataset, ...) val_loader = DataLoader(val_dataset, ...) trainer.fit(model, train_loader, val_loader) ### 3. 保持模型模块化 python class Encoder(nn.Module): def __init__(self): super().__init__() self.layers = nn.Sequential(...) def forward(self, x): return self.layers(x) class Decoder(nn.Module): def __init__(self): super().__init__() self.layers = nn.Sequential(...) def forward(self, x): return self.layers(x) class MyModel(L.LightningModule): def __init__(self): super().__init__() self.encoder = Encoder() self.decoder = Decoder() def forward(self, x): z = self.encoder(x) ret