在 PyTorch 中，nn.Module 是神經網絡中最核心的基類，用于構建所有模型。理解并熟練使用 nn.Module 是掌握 PyTorch 的關鍵。
一、什么是 nn.Module
nn.Module 是 PyTorch 中所有神經網絡模塊的基類。可以把它看作是“神經網絡的容器”，它封裝了以下幾件事：
- 網絡層（如 Linear、Conv2d 等）
- 前向傳播邏輯（
forward函數） - 模型參數（自動注冊并可訓練）
- 可嵌套（可以包含多個子模塊）
- 便捷的模型保存 / 加載等工具函數
二、基礎用法
2.1 自定義模型類
import torch
import torch.nn as nnclass MyNet(nn.Module):def __init__(self):super().__init__()self.fc1 = nn.Linear(784, 128)self.relu = nn.ReLU()self.fc2 = nn.Linear(128, 10)def forward(self, x):x = self.fc1(x)x = self.relu(x)x = self.fc2(x)return x
2.2 實例化與調用
model = MyNet()
x = torch.randn(32, 784) # batch_size = 32
output = model(x) # 自動調用 forward
三、構造方法詳解
3.1 __init__()
- 定義子模塊、層等結構。
- 例如
self.conv1 = nn.Conv2d(...)會被自動注冊為模型參數。
3.2 forward()
- 定義前向傳播邏輯。
- 不能手動調用，應使用
model(x)形式。
四、常見模塊層
| 模塊名 | 作用 | 示例 |
|---|---|---|
nn.Linear | 全連接層 | nn.Linear(128, 64) |
nn.Conv2d | 卷積層 | nn.Conv2d(3, 16, 3) |
nn.ReLU | 激活函數 | nn.ReLU() |
nn.Sigmoid | 激活函數 | nn.Sigmoid() |
nn.BatchNorm2d | 批歸一化 | nn.BatchNorm2d(16) |
nn.Dropout | Dropout 層 | nn.Dropout(0.5) |
nn.LSTM | LSTM 層 | nn.LSTM(10, 20) |
nn.Sequential | 層的順序容器 | 見下文說明 |
五、模型嵌套結構（子模塊）
你可以將一個 nn.Module 作為另一個模塊的子模塊嵌套：
class Block(nn.Module):def __init__(self):super().__init__()self.layer = nn.Sequential(nn.Linear(64, 64),nn.ReLU())def forward(self, x):return self.layer(x)class Net(nn.Module):def __init__(self):super().__init__()self.block1 = Block()self.block2 = Block()self.output = nn.Linear(64, 10)def forward(self, x):x = self.block1(x)x = self.block2(x)return self.output(x)
六、內置方法和屬性
| 方法 / 屬性 | 說明 |
|---|---|
model.parameters() | 返回所有可訓練參數（用于優化器） |
model.named_parameters() | 返回帶名字的參數迭代器 |
model.children() | 返回子模塊迭代器 |
model.eval() | 設置為評估模式（Dropout、BN失效） |
model.train() | 設置為訓練模式 |
model.to(device) | 將模型轉移到 GPU/CPU |
model.state_dict() | 獲取模型參數字典（保存） |
model.load_state_dict() | 加載模型參數字典 |
七、使用 nn.Sequential
nn.Sequential 是一個順序容器，可以用來簡化網絡結構定義：
model = nn.Sequential(nn.Linear(784, 128),nn.ReLU(),nn.Linear(128, 10)
)
等價于手寫的自定義 nn.Module。適合前向傳播是線性“流動”的結構。
八、實戰完整示例：MNIST 分類網絡
class MNISTNet(nn.Module):def __init__(self):super().__init__()self.net = nn.Sequential(nn.Flatten(),nn.Linear(28*28, 256),nn.ReLU(),nn.Linear(256, 10))def forward(self, x):return self.net(x)# 實例化模型
model = MNISTNet()
print(model)# 配置訓練
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)# 示例訓練循環
for epoch in range(10):for images, labels in train_loader:output = model(images)loss = criterion(output, labels)optimizer.zero_grad()loss.backward()optimizer.step()
九、常見陷阱和建議
| 問題 | 說明 |
|---|---|
forward() 不起作用 | 應該使用 model(x)，而不是手動調用 model.forward(x) |
忘記 super().__init__() | 子模塊將不會被注冊 |
| 參數未注冊 | 層/模塊必須賦值為 self.xxx = ... |
| 訓練/測試模式混淆 | 注意 model.eval() 和 model.train() |
十、總結
| 項目 | 說明 |
|---|---|
__init__() | 定義模型結構（子模塊、層） |
forward() | 定義前向傳播 |
| 自動注冊參數 | 所有 self.xxx = nn.XXX(...) 都會被追蹤 |
| 嵌套模塊 | 支持遞歸子模塊調用 |
| 便捷方法 | .parameters()、.to()、.eval() 等 |
十一、綜合示例
以下是基于 PyTorch nn.Module 封裝的三種經典深度學習架構（ResNet18、UNet、Transformer）的簡潔而完整的實現，適合初學者快速上手。
1、ResNet18 簡潔實現（適合圖像分類）
import torch
import torch.nn as nn
import torch.nn.functional as Fclass BasicBlock(nn.Module):expansion = 1def __init__(self, in_planes, planes, stride=1, downsample=None):super().__init__()self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)self.bn1 = nn.BatchNorm2d(planes)self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)self.bn2 = nn.BatchNorm2d(planes)self.downsample = downsampledef forward(self, x):identity = xif self.downsample:identity = self.downsample(x)out = F.relu(self.bn1(self.conv1(x)))out = self.bn2(self.conv2(out))out += identityreturn F.relu(out)class ResNet(nn.Module):def __init__(self, block, layers, num_classes=1000):super().__init__()self.in_planes = 64self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)self.bn1 = nn.BatchNorm2d(64)self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)self.layer1 = self._make_layer(block, 64, layers[0])self.layer2 = self._make_layer(block, 128, layers[1], stride=2)self.layer3 = self._make_layer(block, 256, layers[2], stride=2)self.layer4 = self._make_layer(block, 512, layers[3], stride=2)self.avgpool = nn.AdaptiveAvgPool2d((1, 1))self.fc = nn.Linear(512 * block.expansion, num_classes)def _make_layer(self, block, planes, blocks, stride=1):downsample = Noneif stride != 1 or self.in_planes != planes * block.expansion:downsample = nn.Sequential(nn.Conv2d(self.in_planes, planes * block.expansion,kernel_size=1, stride=stride, bias=False),nn.BatchNorm2d(planes * block.expansion))layers = [block(self.in_planes, planes, stride, downsample)]self.in_planes = planes * block.expansionfor _ in range(1, blocks):layers.append(block(self.in_planes, planes))return nn.Sequential(*layers)def forward(self, x):x = self.pool(F.relu(self.bn1(self.conv1(x))))x = self.layer1(x)x = self.layer2(x)x = self.layer3(x)x = self.layer4(x)x = self.avgpool(x).flatten(1)return self.fc(x)def ResNet18(num_classes=1000):return ResNet(BasicBlock, [2, 2, 2, 2], num_classes)
2、UNet（適合圖像分割）
class UNetBlock(nn.Module):def __init__(self, in_ch, out_ch):super().__init__()self.block = nn.Sequential(nn.Conv2d(in_ch, out_ch, 3, padding=1),nn.ReLU(inplace=True),nn.Conv2d(out_ch, out_ch, 3, padding=1),nn.ReLU(inplace=True))def forward(self, x):return self.block(x)class UNet(nn.Module):def __init__(self, in_channels=1, out_channels=1):super().__init__()self.enc1 = UNetBlock(in_channels, 64)self.enc2 = UNetBlock(64, 128)self.enc3 = UNetBlock(128, 256)self.enc4 = UNetBlock(256, 512)self.pool = nn.MaxPool2d(2)self.bottleneck = UNetBlock(512, 1024)self.upconv4 = nn.ConvTranspose2d(1024, 512, 2, stride=2)self.dec4 = UNetBlock(1024, 512)self.upconv3 = nn.ConvTranspose2d(512, 256, 2, stride=2)self.dec3 = UNetBlock(512, 256)self.upconv2 = nn.ConvTranspose2d(256, 128, 2, stride=2)self.dec2 = UNetBlock(256, 128)self.upconv1 = nn.ConvTranspose2d(128, 64, 2, stride=2)self.dec1 = UNetBlock(128, 64)self.final = nn.Conv2d(64, out_channels, kernel_size=1)def forward(self, x):e1 = self.enc1(x)e2 = self.enc2(self.pool(e1))e3 = self.enc3(self.pool(e2))e4 = self.enc4(self.pool(e3))b = self.bottleneck(self.pool(e4))d4 = self.upconv4(b)d4 = self.dec4(torch.cat([d4, e4], dim=1))d3 = self.upconv3(d4)d3 = self.dec3(torch.cat([d3, e3], dim=1))d2 = self.upconv2(d3)d2 = self.dec2(torch.cat([d2, e2], dim=1))d1 = self.upconv1(d2)d1 = self.dec1(torch.cat([d1, e1], dim=1))return self.final(d1)
3、簡化版 Transformer 編碼器（適合序列建模）
class TransformerBlock(nn.Module):def __init__(self, embed_dim, heads, ff_hidden_dim, dropout=0.1):super().__init__()self.attn = nn.MultiheadAttention(embed_dim, heads, dropout=dropout, batch_first=True)self.ff = nn.Sequential(nn.Linear(embed_dim, ff_hidden_dim),nn.ReLU(),nn.Linear(ff_hidden_dim, embed_dim))self.norm1 = nn.LayerNorm(embed_dim)self.norm2 = nn.LayerNorm(embed_dim)self.dropout = nn.Dropout(dropout)def forward(self, x, mask=None):attn_out, _ = self.attn(x, x, x, attn_mask=mask)x = self.norm1(x + self.dropout(attn_out))ff_out = self.ff(x)x = self.norm2(x + self.dropout(ff_out))return xclass TransformerEncoder(nn.Module):def __init__(self, vocab_size, embed_dim=512, n_heads=8, ff_dim=2048, num_layers=6, max_len=512):super().__init__()self.embedding = nn.Embedding(vocab_size, embed_dim)self.pos_encoding = self._generate_positional_encoding(max_len, embed_dim)self.layers = nn.ModuleList([TransformerBlock(embed_dim, n_heads, ff_dim)for _ in range(num_layers)])self.dropout = nn.Dropout(0.1)def _generate_positional_encoding(self, max_len, d_model):pos = torch.arange(0, max_len).unsqueeze(1)i = torch.arange(0, d_model, 2)angle_rates = 1 / torch.pow(10000, (i / d_model))pos_enc = torch.zeros(max_len, d_model)pos_enc[:, 0::2] = torch.sin(pos * angle_rates)pos_enc[:, 1::2] = torch.cos(pos * angle_rates)return pos_enc.unsqueeze(0)def forward(self, x):B, T = x.shapex = self.embedding(x) + self.pos_encoding[:, :T].to(x.device)x = self.dropout(x)for layer in self.layers:x = layer(x)return x
4、 總結對比
| 模型類型 | 場景 | 特點 |
|---|---|---|
| ResNet18 | 圖像分類 | 深殘差網絡結構，適合遷移學習 |
| UNet | 圖像分割 | 對稱結構，編碼 + 解碼 + skip |
| Transformer | NLP / 序列建模 | 全注意力機制，無卷積無循環 |
