循环神经网络（RNN、LSTM、GRU）的结构及代码实现

RNN

 代码实现：

# 一般循环神经网络RNN
class ConvRNN(nn.Module):
    def __init__(self, inp_dim, oup_dim, kernel, dilation):
        super().__init__()
        pad_x = int(dilation * (kernel - 1) / 2)
        self.conv_x = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)

        pad_h = int((kernel - 1) / 2)
        self.conv_h = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)
        self.relu = nn.LeakyReLU(0.2)

    def forward(self, x, h=None):
        if h is None:
            h = F.tanh(self.conv_x(x))
        else:
            h = F.tanh(self.conv_x(x) + self.conv_h(h))

        h = self.relu(h)
        return h, h

---

 

LSTM  

参考：图解LSTM 结构逻辑_BruceJust的博客-CSDN博客_lstm结构图

 

前向传播公式：

 

代码实现：

 

# 参考LSTM结构图
# https://blog.csdn.net/weixin_42175217/article/details/106183682
class ConvLSTM(nn.Module):
    def __init__(self, inp_dim, oup_dim, kernel, dilation):
        super().__init__()
        pad_x = int(dilation * (kernel - 1) / 2)
        self.conv_xf = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)
        self.conv_xi = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)
        self.conv_xo = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)
        self.conv_xj = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)

        pad_h = int((kernel - 1) / 2)
        self.conv_hf = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)
        self.conv_hi = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)
        self.conv_ho = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)
        self.conv_hj = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)

        self.relu = nn.LeakyReLU(0.2)

    def forward(self, x, pair=None):
        if pair is None:
            i = F.sigmoid(self.conv_xi(x))
            o = F.sigmoid(self.conv_xo(x))
            j = F.tanh(self.conv_xj(x))
            c = i * j
            h = o * c
        else:
            h, c = pair
            f = F.sigmoid(self.conv_xf(x) + self.conv_hf(h))
            i = F.sigmoid(self.conv_xi(x) + self.conv_hi(h))
            o = F.sigmoid(self.conv_xo(x) + self.conv_ho(h))
            j = F.tanh(self.conv_xj(x) + self.conv_hj(h))
            c = f * c + i * j
            h = o * F.tanh(c)

        h = self.relu(h)
        return h, [h, c]

---

 GRU

 

前向传播公式：

 代码实现：

class ConvGRU(nn.Module):
    def __init__(self, inp_dim, oup_dim, kernel, dilation):
        super().__init__()
        pad_x = int(dilation * (kernel - 1) / 2)
        self.conv_xz = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)
        self.conv_xr = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)
        self.conv_xn = nn.Conv2d(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)

        pad_h = int((kernel - 1) / 2)
        self.conv_hz = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)
        self.conv_hr = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)
        self.conv_hn = nn.Conv2d(oup_dim, oup_dim, kernel, padding=pad_h)

        self.relu = nn.LeakyReLU(0.2)

    def forward(self, x, h=None):
        if h is None:
            z = F.sigmoid(self.conv_xz(x))
            f = F.tanh(self.conv_xn(x))
            h = z * f
        else:
            z = F.sigmoid(self.conv_xz(x) + self.conv_hz(h))
            r = F.sigmoid(self.conv_xr(x) + self.conv_hr(h))
            n = F.tanh(self.conv_xn(x) + self.conv_hn(r * h))
            h = (1 - z) * h + z * n

        h = self.relu(h)
        return h, h

 

 代码参考：DCSFN/model.py at master · Ohraincu/DCSFN · GitHub