Stacked two LSTMs with different hidden layers

Hi,

I would like to create LSTM layers which contain different hidden layers to predict time series data, for the 1st layer of LSTM_1 contains 10 hidden layers, LSTM_2 contains 1 hidden layer, the proposed neural network architecture is illustrated following
f1

def __init__(self, nb_features=1, hidden_size_1=100, hidden_size_2=100, nb_layers_1 =10, nb_layers_2 = 1, dropout=0.5): #(self, nb_features=1, hidden_size=100, nb_layers=10, dropout=0.5):      
    super(Sequence, self).__init__()
    self.nb_features = nb_features
    self.hidden_size_1 = hidden_size_1
    self.hidden_size_2 = hidden_size_2
    self.nb_layers_1 =nb_layers_1
    self.nb_layers_2 = nb_layers_2
    self.lstm_1 = nn.LSTM(self.nb_features, self.hidden_size_1, self.nb_layers_1, dropout=dropout)  
    self.lstm_2 = nn.LSTM(self.hidden_size_1, self.hidden_size_2, self.nb_layers_2, dropout=dropout)
    self.lin = nn.Linear(self.hidden_size_2, 1)



def forward(self, input):  
    h_t1 = Variable(torch.zeros(self.nb_layers_1, input.size()[1], self.hidden_size_1))
    c_t1 = Variable(torch.zeros(self.nb_layers_1, input.size()[1], self.hidden_size_1))
    h_t2 = Variable(torch.zeros(self.nb_layers_2, input.size()[1], self.hidden_size_2))
    c_t2 = Variable(torch.zeros(self.nb_layers_2, input.size()[1], self.hidden_size_2))
    outputs = []

    for i, input_t in enumerate(input.chunk(input.size(1))):
        h_t1, c_t1 = self.lstm_1(input_t, (h_t1, c_t1))
        h_t2, _ = self.lstm_2(h_t1, (h_t2, c_t2))
        output = self.lin(h_t2[-1])
        outputs += [output]

    outputs = torch.stack(outputs, 1).squeeze(2)

    return outputs

However, I got runtime error for the above code “RuntimeError: Expected hidden[0] size (3, 60, 100), got (1, 60, 100)”. is it correct for the above code, Could you please give some suggestions? Many thanks

3 Likes

Quick reply, I didn’t try your code but it seems that you are ignoring the lstm output

i.e. these two lines:

h_t1, c_t1 = self.lstm_1(input_t, (h_t1, c_t1))
h_t2, _ = self.lstm_2(h_t1, (h_t2, c_t2))

should read:

_, (h_t1, c_t1) = self.lstm_1(input_t, (h_t1, c_t1))
_, (h_t2, _) = self.lstm_2(h_t1, (h_t2, c_t2))

if you are not interested in the lstm output at all and just want to use the hidden states.

Check the LSTM output here https://pytorch.org/docs/stable/nn.html#lstm

3 Likes