[파이토치] LSTM / GRU

• There are two other popular recurrent layer: LSTM and GRU.
• SimpleRNN is difficult to learn long-term dependencies.
  • This is due to the vanishing gradient problem.

• Long Short-Term Memory (LSTM)
  • proposed by Hochreiter and Schmidhuber in 1997
  • It adds a way to carry information across many timesteps.

Details of LSTM

 

LSTM diagram

Computations involved in LSTM

---

class LSTM(nn.Module):

    

    def __init__(self, input_size, hidden_size, sequence_length, num_layers, device):

        super(LSTM, self).__init__()

        self.device = device

        self.hidden_size = hidden_size

        self.num_layers = num_layers

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

        self.fc = nn.Linear(hidden_size*sequence_length, 1)

        

    def forward(self, x):

        h0 = torch.zeros(self.num_layers, x.size()[0], self.hidden_size).to(self.device)

        c0 = torch.zeros(self.num_layers, x.size()[0], self.hidden_size).to(self.device) # cell state가 추가되었다.

        out, _ = self.lstm(x, (h0, c0)) # output, (hn, cn): cell state와 hidden state만 반환 (순서쌍 형태로)

        out = out.reshape(out.shape[0], -1) # <- state 추가

        out = self.fc(out)

        return out

---

GRU &amp;nbsp;diagram

Computations involved in GRU

---

class GRU(nn.Module):

    

    def __init__(self, input_size, hidden_size, sequence_length, num_layers, device):

        super(GRU, self).__init__()

        self.device = device

        self.hidden_size = hidden_size

        self.num_layers = num_layers

        self.gru = nn.GRU(input_size, hidden_size, num_layers, batch_first=True)

        self.fc = nn.Linear(hidden_size*sequence_length, 1)

        

    def forward(self, x):

        h0 = torch.zeros(self.num_layers, x.size()[0], self.hidden_size).to(self.device)

        out, _ = self.gru(x, h0)

        out = out.reshape(out.shape[0], -1) # <- state 추가

        out = self.fc(out)

        return out

 

GRU 모델을 이용한 예측 그래프