deepseek.com 从入门到精通长短时记忆网络(LSTM),着重介绍的目标函数,损失函数,梯度下降 标量和矩阵形式的数学推导,pytorch真实能跑的代码案例以及模型,数据, 模型应用场景和优缺点,及如何改进解决及改进方法数据推导。

从入门到精通长短时记忆网络 (LSTM)

参考:长短时记忆网络(LSTM)在序列数据处理中的优缺点分析
LSTM


1. LSTM 核心机制

LSTM 通过门控机制(遗忘门、输入门、输出门)和细胞状态(Cell State)解决 RNN 的梯度消失问题。

核心公式(时间步 t t t):

  1. 遗忘门(Forget Gate):
    f t = σ ( W f [ h t − 1 , x t ] + b f ) \mathbf{f}_t = \sigma\left( \mathbf{W}_f [\mathbf{h}_{t-1}, \mathbf{x}_t] + \mathbf{b}_f \right) ft=σ(Wf[ht1,xt]+bf)
  2. 输入门(Input Gate):
    i t = σ ( W i [ h t − 1 , x t ] + b i ) \mathbf{i}_t = \sigma\left( \mathbf{W}_i [\mathbf{h}_{t-1}, \mathbf{x}_t] + \mathbf{b}_i \right) it=σ(Wi[ht1,xt]+bi)
    C ~ t = tanh ⁡ ( W C [ h t − 1 , x t ] + b C ) \tilde{\mathbf{C}}_t = \tanh\left( \mathbf{W}_C [\mathbf{h}_{t-1}, \mathbf{x}_t] + \mathbf{b}_C \right) C~t=tanh(WC[ht1,xt]+bC)
  3. 细胞状态更新
    C t = f t ⊙ C t − 1 + i t ⊙ C ~ t \mathbf{C}_t = \mathbf{f}_t \odot \mathbf{C}_{t-1} + \mathbf{i}_t \odot \tilde{\mathbf{C}}_t Ct=ftCt1+itC~t
  4. 输出门(Output Gate):
    o t = σ ( W o [ h t − 1 , x t ] + b o ) \mathbf{o}_t = \sigma\left( \mathbf{W}_o [\mathbf{h}_{t-1}, \mathbf{x}_t] + \mathbf{b}_o \right) ot=σ(Wo[ht1,xt]+bo)
    h t = o t ⊙ tanh ⁡ ( C t ) \mathbf{h}_t = \mathbf{o}_t \odot \tanh(\mathbf{C}_t) ht=ottanh(Ct)

2. 目标函数与损失函数
  • 目标函数:最小化预测与真实值的差异(监督学习)。
  • 损失函数(以分类任务交叉熵为例):
    L = − 1 T ∑ t = 1 T ∑ c = 1 C y ^ t , c log ⁡ ( y t , c ) L = -\frac{1}{T} \sum_{t=1}^T \sum_{c=1}^C \mathbf{\hat{y}}_{t,c} \log(\mathbf{y}_{t,c}) L=T1t=1Tc=1Cy^t,clog(yt,c)
    其中 C C C为类别数, y ^ \mathbf{\hat{y}} y^为真实标签的 one-hot 编码。

3. 梯度下降与数学推导

LSTM 的梯度反向传播通过细胞状态 C t \mathbf{C}_t Ct和门控机制稳定梯度流动。

标量形式推导(以遗忘门 f t \mathbf{f}_t ft为例):
∂ L ∂ f t = ∂ L ∂ h t ⋅ ∂ h t ∂ C t ⋅ ∂ C t ∂ f t \frac{\partial L}{\partial \mathbf{f}_t} = \frac{\partial L}{\partial \mathbf{h}_t} \cdot \frac{\partial \mathbf{h}_t}{\partial \mathbf{C}_t} \cdot \frac{\partial \mathbf{C}_t}{\partial \mathbf{f}_t} ftL=htLCthtftCt
其中:
∂ C t ∂ f t = C t − 1 ⊙ f t ⊙ ( 1 − f t ) \frac{\partial \mathbf{C}_t}{\partial \mathbf{f}_t} = \mathbf{C}_{t-1} \odot \mathbf{f}_t \odot (1 - \mathbf{f}_t) ftCt=Ct1ft(1ft)

矩阵形式推导(链式法则):
∂ L ∂ W f = ∑ t = 1 T ( δ f , t ⋅ [ h t − 1 , x t ] T ) \frac{\partial L}{\partial \mathbf{W}_f} = \sum_{t=1}^T \left( \delta_{f,t} \cdot [\mathbf{h}_{t-1}, \mathbf{x}_t]^T \right) WfL=t=1T(δf,t[ht1,xt]T)
其中 δ f , t \delta_{f,t} δf,t为遗忘门的梯度误差:
δ f , t = ∂ L ∂ f t ⊙ σ ′ ( ⋅ ) \delta_{f,t} = \frac{\partial L}{\partial \mathbf{f}_t} \odot \sigma'(\cdot) δf,t=ftLσ()


4. PyTorch 代码案例
import torch
import torch.nn as nn
import matplotlib.pyplot as plt

# 数据生成:正弦波 + 噪声
time = torch.arange(0, 100, 0.1)
data = torch.sin(time) + 0.1 * torch.randn(len(time))

# 转换为序列数据(窗口长度=20)
def create_sequences(data, seq_length=20):
    X, y = [], []
    for i in range(len(data)-seq_length):
        X.append(data[i:i+seq_length])
        y.append(data[i+seq_length])
    return torch.stack(X).unsqueeze(-1), torch.stack(y).unsqueeze(-1)

X, y = create_sequences(data)
X_train, y_train = X[:800], y[:800]  # 划分训练集和测试集
X_test, y_test = X[800:], y[800:]

# 定义 LSTM 模型
class LSTMModel(nn.Module):
    def __init__(self, input_size=1, hidden_size=64, output_size=1):
        super().__init__()
        self.lstm = nn.LSTM(input_size, hidden_size, batch_first=True)
        self.fc = nn.Linear(hidden_size, output_size)
    
    def forward(self, x):
        out, (h_n, c_n) = self.lstm(x)  # out: (batch, seq_len, hidden_size)
        out = self.fc(out[:, -1, :])    # 取最后一个时间步
        return out

model = LSTMModel()
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

# 训练
epochs = 100
train_loss = []
for epoch in range(epochs):
    optimizer.zero_grad()
    outputs = model(X_train)
    loss = criterion(outputs, y_train)
    loss.backward()
    nn.utils.clip_grad_norm_(model.parameters(), 0.5)  # 梯度裁剪
    optimizer.step()
    train_loss.append(loss.item())

# 可视化训练损失
plt.plot(train_loss)
plt.title("Training Loss")
plt.show()

# 预测
model.eval()
with torch.no_grad():
    train_pred = model(X_train)
    test_pred = model(X_test)

# 绘制结果
plt.figure(figsize=(12, 5))
plt.plot(data.numpy(), label="True Data")
plt.plot(range(20, 820), train_pred.numpy(), label="Train Predictions")
plt.plot(range(820, len(data)), test_pred.numpy(), label="Test Predictions")
plt.legend()
plt.show()

5. 应用场景与优缺点
  • 应用场景
    • 时间序列预测(股票价格、天气)
    • 自然语言处理(文本生成、机器翻译)
    • 语音识别
  • 优点
    • 解决长程依赖问题
    • 通过门控机制稳定梯度流动
    • 可处理变长序列
  • 缺点
    • 计算复杂度高(参数多)
    • 对短序列可能过拟合
    • 训练时间较长

6. 改进方法及数学推导
  1. GRU(门控循环单元)
    简化 LSTM,合并遗忘门和输入门:
    z t = σ ( W z [ h t − 1 , x t ] ) \mathbf{z}_t = \sigma(\mathbf{W}_z [\mathbf{h}_{t-1}, \mathbf{x}_t]) zt=σ(Wz[ht1,xt])
    r t = σ ( W r [ h t − 1 , x t ] ) \mathbf{r}_t = \sigma(\mathbf{W}_r [\mathbf{h}_{t-1}, \mathbf{x}_t]) rt=σ(Wr[ht1,xt])
    h ~ t = tanh ⁡ ( W [ r t ⊙ h t − 1 , x t ] ) \tilde{\mathbf{h}}_t = \tanh(\mathbf{W} [\mathbf{r}_t \odot \mathbf{h}_{t-1}, \mathbf{x}_t]) h~t=tanh(W[rtht1,xt])
    h t = ( 1 − z t ) ⊙ h t − 1 + z t ⊙ h ~ t \mathbf{h}_t = (1 - \mathbf{z}_t) \odot \mathbf{h}_{t-1} + \mathbf{z}_t \odot \tilde{\mathbf{h}}_t ht=(1zt)ht1+zth~t

  2. 双向 LSTM(Bi-LSTM)
    同时捕捉前向和后向依赖:
    h t → = LSTM ( x t , h t − 1 → ) \overrightarrow{\mathbf{h}_t} = \text{LSTM}(\mathbf{x}_t, \overrightarrow{\mathbf{h}_{t-1}}) ht =LSTM(xt,ht1 )
    h t ← = LSTM ( x t , h t + 1 ← ) \overleftarrow{\mathbf{h}_t} = \text{LSTM}(\mathbf{x}_t, \overleftarrow{\mathbf{h}_{t+1}}) ht =LSTM(xt,ht+1 )
    h t = [ h t → , h t ← ] \mathbf{h}_t = [\overrightarrow{\mathbf{h}_t}, \overleftarrow{\mathbf{h}_t}] ht=[ht ,ht ]

  3. 注意力机制
    增强对关键时间步的关注:
    α t = softmax ( v T tanh ⁡ ( W h h t + W s s ) ) \alpha_t = \text{softmax}(\mathbf{v}^T \tanh(\mathbf{W}_h \mathbf{h}_t + \mathbf{W}_s \mathbf{s})) αt=softmax(vTtanh(Whht+Wss))
    c = ∑ t = 1 T α t h t \mathbf{c} = \sum_{t=1}^T \alpha_t \mathbf{h}_t c=t=1Tαtht


7. 关键改进的数学验证(以 GRU 为例)
  • 梯度稳定性
    GRU 的更新门 z t \mathbf{z}_t zt控制历史信息的保留比例,梯度可沿两条路径传播:
    ∂ h t ∂ h t − 1 = ( 1 − z t ) + z t ⊙ ∂ h ~ t ∂ h t − 1 \frac{\partial \mathbf{h}_t}{\partial \mathbf{h}_{t-1}} = (1 - \mathbf{z}_t) + \mathbf{z}_t \odot \frac{\partial \tilde{\mathbf{h}}_t}{\partial \mathbf{h}_{t-1}} ht1ht=(1zt)+ztht1h~t
    避免传统 RNN 的连乘梯度。

通过上述内容,您可全面掌握 LSTM 的理论基础、实际实现及优化方法。

Logo

欢迎加入DeepSeek 技术社区。在这里,你可以找到志同道合的朋友,共同探索AI技术的奥秘。

更多推荐