LSTM模型在整个训练过程中都持续损失

Question

暹罗模型实施可以在下面找到：

class ClassifierSiameseLSTM(nn.Module):
    def __init__(self, num_sensors=2, hidden_units=16):
        super().__init__()
        self.num_sensors = num_sensors  # this is the number of features
        self.hidden_units = hidden_units
        self.num_layers = 1

        self.lstm = nn.LSTM(
          input_size=num_sensors,
          hidden_size=hidden_units,
          batch_first=True,
          num_layers=self.num_layers)

        self.fc = nn.Linear(in_features=self.hidden_units, out_features=256)

    def forward_once(self, x):
        batch_size = x.shape[0]
        h0 = torch.zeros(
          self.num_layers, batch_size, self.hidden_units,
          dtype=torch.double).to(device).requires_grad_()
        c0 = torch.zeros(self.num_layers, batch_size, self.hidden_units, 
          dtype=torch.double).to(device).requires_grad_()

        output, (hn, cn) = self.lstm(x, (h0, c0))
        out = self.fc(output[:, -1, :])
        return out

    def forward(self,x1,x2):
        output1 = self.forward_once(x1)
        output2 = self.forward_once(x2)

        return output1,output2

当我尝试训练模型时，我得到以下损失值：

Epoch : 0
Train Loss: 0.090 | Accuracy: 64.000

Epoch : 1
Train Loss: 0.091 | Accuracy: 64.000

Epoch : 2
Train Loss: 0.090 | Accuracy: 64.000

Epoch : 3
Train Loss: 0.090 | Accuracy: 64.000

Epoch : 4
Train Loss: 0.091 | Accuracy: 64.000

Epoch : 5
Train Loss: 0.090 | Accuracy: 64.000

损失值几乎是恒定的，并且模型根本没有学习。有人知道解决方案还是对问题的原因有想法？