使用 LightningModule 定義模型

Scalable AI Models with PyTorch Lightning

Sergiy Tkachuk

Director, GenAI Productivity

聚焦 LightningModule

封裝模型架構
將訓練邏輯整理為單一且易管理的單元
為深度學習專案建立清晰有序的藍圖

PyTorch LightningModule 圖示

定義 init 方法

重點任務：

模型初始化
super():
- 自動處理訓練迴圈
- 紀錄日誌
- 檢查點儲存
初始化後定義模型層
模組化且易於維護

import lightning.pytorch as pl
import torch.nn as nn

class ClassificationModel(pl.LightningModule):
    def __init__(self, input_dim,
                 hidden_dim, num_class):
          # Initialize parent class
        super().__init__()

        # First layer
        self.layer1 = nn.Linear(input_dim,
                                hidden_dim)
        # Activation function
        self.relu = nn.ReLU()
        # Output layer
        self.layer2 = nn.Linear(hidden_dim,
                                num_class)

實作 forward 方法

關鍵步驟：

定義資料在網路中的流向
依序經過各層處理輸入
- 線性轉換
- 啟用函式
- 最後一層與輸出

import lightning.pytorch as pl
import torch.nn as nn

class ClassificationModel(pl.LightningModule):
    def __init__(self, input_dim,
                 hidden_dim, num_class):
          ...

    def forward(self, x):

        x = self.layer1(x) # Pass input

        x = nn.ReLU(x) # Apply activation

        x = self.layer2(x) # Compute output

        return x # Return result

範例：手寫數字分類

import lightning.pytorch as pl
from torch.utils.data import DataLoader
from torchvision.datasets import MNIST
from torchvision import transforms

transform = transforms.ToTensor()
train_ds = MNIST(root='.', train=True,  download=True, transform=transform)
test_ds  = MNIST(root='.', train=False, download=True, transform=transform)

train_loader = DataLoader(train_ds, batch_size=64, shuffle=True)
test_loader  = DataLoader(test_ds,  batch_size=64)


model = ClassificationModel(input_dim=28*28, hidden_dim=128, num_class=10)


trainer = pl.Trainer(max_epochs=3, accelerator='auto')
trainer.fit(model, train_loader, test_loader)

將模型用於分類任務的整合

聚焦分類情境
全流程封裝於 LightningModule
輸出提供軟性最大值前的原始分數
與 Lightning Trainer 整合

class ClassificationModel(pl.LightningModule):
  def __init__(self, input_dim, 
               hidden_dim, output_dim):
    super().__init__()

    self.hid = nn.Linear(input_dim, hidden_dim)
    self.out = nn.Linear(hidden_dim, output_dim)


  def forward(self, x):

    x = self.hidden(x)
    x = nn.ReLU(x)
    x = self.output(x)

    return x

一起來練習吧！

Scalable AI Models with PyTorch Lightning