PyTorch训练LSTM时遭遇梯度消失问题求助

我正在用PyTorch训练一个简单LSTM神经网络做股票价格预测，但模型完全无法拟合——损失爆炸、R2值为负，训练过程毫无改善。代码里肯定有致命错误，试了多种方法都没解决。

以下是我的代码：

class LSTMModel(nn.Module):
    def __init__(self, features):
        super(LSTMModel, self).__init__()
        self.lstm1 = nn.LSTM(input_size=features, hidden_size=16, batch_first=True)
        self.dense2 = nn.Linear(16, 1)

        self._init_weights()

    def forward(self, x):
        x, _ = self.lstm1(x)
        # x, _ = self.lstm2(x)
        # Flatten the output for Dense layer input
        x = x[:, -1, :]
        # x = self.dense1(x)
        x = self.dense2(x)
        return x

    def _init_weights(self):
        for name, param in self.named_parameters():
            if 'weight' in name:
                nn.init.xavier_uniform_(param)
            elif 'bias' in name:
                nn.init.zeros_(param)


# Initialize the model
model = LSTMModel(len(feature_cols))
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=0.01)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.95)


def train_model(num_epochs):
    for epoch in range(num_epochs):
        model.train()
        total_loss = 0
        for data, target in train_loader:
            optimizer.zero_grad()
            output = model(data)
            loss = criterion(output.reshape(len(output), ), target)
            loss.backward()

            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)

            optimizer.step()
            total_loss += loss.item()
        scheduler.step()

        model.eval()
        val_loss = 0
        pred = []
        with torch.no_grad():
            for data, target in test_loader:
                output = model(data)
                # print(data, output.reshape(len(output), ), target)
                val_loss += criterion(output.reshape(len(output), ), target).item()
                pred += list(output.reshape(len(output), ))

        val_loss /= len(test_loader)
        r2 = r2_score(test_y, pred)
        print(f'Epoch {epoch + 1}, Train Loss: {total_loss / len(train_loader)}, Val Loss: {val_loss}, val r2: {r2}')

我已尝试的方法：

• 梯度裁剪（代码中已实现），无效
• 修改batch size，无效
• 查看网络权重，发现LSTM层权重均接近0，而全连接层权重正常，疑似梯度消失问题
• 自定义权重初始化（代码中已实现），无效
• 修改模型超参数（包括隐藏层数量、学习率、hidden_size等），无效
• 修改输入特征数量，无效
• 调整时间序列数据的滑动窗口大小，无效

备注：

• 输入数据特征已使用MinMaxScaler进行归一化
• 数据集包含约4000条观测数据

内容的提问来源于stack exchange，提问作者王一诺