I am attempting to use a Seq2Seq model to make forecasts of factory production data using an Encoder-Decoder model augmented with Attention. I have become a little stuck as the output of the model seems to be a constant and has the same size sequence length as the input, where in fact I would like to be able to specify that say I want to forecast 3 (or any number of) months into the future.

Here is 2 diagrams of the Seq2Seq architecture and the attention mechanism I am looking to construct:

Seq2Seq Architecture


The Target
To my understanding, I went to be predicting the production volume of a given material from this factory into the future. So its dimensionality is $1$ and it is of course an integer.

The Encoder
The encoder takes as input a sequence of length $168$, with each input being the $20$ previous days data, as well as $37$ factory-level features such as number of workers etc etc..

The Decoder
This is where I get confused and where I am running into issues with my code. Again, to my understanding the Decoder should be taking the previous time-steps production levels as input (meaning dimension $1$), as well as the previous hidden and cell state.


class EncoderRNN(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, p):
        super(EncoderRNN, self).__init__()
        self.lstm = nn.LSTM(input_size, hidden_size,
                            num_layers, dropout = p, 
                            bidirectional = True)

        self.fc_hidden = nn.Linear(hidden_size*2, hidden_size) 
        self.fc_cell = nn.Linear(hidden_size*2, hidden_size)

    def forward(self, input):
        print(f"Encoder input shape is {input.shape}")
        encoder_states, (hidden, cell_state) = self.lstm(input)

        print(f"Encoder Hidden: {hidden.shape}")
        print(f"Encoder Cell: {cell_state.shape}")

        hidden = self.fc_hidden(torch.cat((hidden[0:1], hidden[1:2]), dim = 2))
        cell = self.fc_cell(torch.cat((cell_state[0:1], cell_state[1:2]), dim = 2))

        print(f"Encoder Hidden: {hidden.shape}")
        print(f"Encoder Cell: {cell.shape}")
        return encoder_states, hidden, cell

class Decoder_LSTMwAttention(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, output_size, p):
        super(Decoder_LSTMwAttention, self).__init__()
        self.rnn = nn.LSTM(hidden_size*2 + input_size, hidden_size,

        self.energy = nn.Linear(hidden_size * 3, 1)
        self.fc = nn.Linear(hidden_size, output_size)
        self.softmax = nn.Softmax(dim=0)
        self.dropout = nn.Dropout(p)
        self.relu = nn.ReLU()  

        self.attention_combine = nn.Linear(hidden_size, hidden_size)

    def forward(self, input, encoder_states, hidden, cell):

        input = input.unsqueeze(0)
        input = input.unsqueeze(0)

        input = self.dropout(input)

        sequence_length = encoder_states.shape[0]
        h_reshaped = hidden.repeat(sequence_length, 1, 1)

        concatenated = torch.cat((h_reshaped, encoder_states), dim = 2)
        print(f"Concatenated size: {concatenated.shape}")

        energy = self.relu(self.energy(concatenated))
        attention = self.softmax(energy)
        attention = attention.permute(1, 0, 2)

        encoder_states = encoder_states.permute(1, 0, 2)

        context_vector = torch.einsum("snk,snl->knl", attention, encoder_states)
        rnn_input = torch.cat((context_vector, input), dim = 2)

        output, (hidden, cell) = self.rnn(rnn_input, hidden, cell)

        output = self.fc(output).squeeze(0)
        return output, hidden, cell

class Seq2Seq(nn.Module):
    def __init__(self, encoder, decoder):
        super(Seq2Seq, self).__init__()
        self.encoder = encoder
        self.decoder = decoder

    def forward(self, source, target, teacher_force_ratio=0.5):
        batch_size = source.shape[1]
        target_len = target.shape[0]
        #target_vocab_size = len(english.vocab)

        outputs = torch.zeros(target_len, batch_size).to(device)
        encoder_states, hidden, cell = self.encoder(source)

        # First input will be <SOS> token
        x = target[0]

        for t in range(1, target_len):
            # At every time step use encoder_states and update hidden, cell
            output, hidden, cell = self.decoder(x, encoder_states, hidden, cell)

            # Store prediction for current time step
            outputs[t] = output

            # Get the best word the Decoder predicted (index in the vocabulary)
            best_guess = output.argmax(1)

            # With probability of teacher_force_ratio we take the actual next word
            # otherwise we take the word that the Decoder predicted it to be.
            # Teacher Forcing is used so that the model gets used to seeing
            # similar inputs at training and testing time, if teacher forcing is 1
            # then inputs at test time might be completely different than what the
            # network is used to. This was a long comment.
            x = target[t] if random.random() < teacher_force_ratio else best_guess

        return outputs

Training Routine

def Seq2seq_trainer(model, optimizer, train_input, train_target,
                  test_input, test_target, criterion, num_epochs):

    train_losses = np.zeros(num_epochs)
    validation_losses = np.zeros(num_epochs)

    for it in range(num_epochs):
        # zero the parameter gradients

        # Forward pass
        outputs = model(train_input, train_target)  
        loss = criterion(outputs, train_target)

        # Back prop

        # Clip to avoid exploding gradient issues
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1)

        # Gradient descent step

        # Save losses
        train_losses[it] = loss.item()

        # Test loss
        test_outputs = model(test_input, test_target)
        validation_loss = loss_function(test_outputs, test_target)
        validation_losses[it] = validation_loss.item()
        if (it + 1) % 25 == 0:
            print(f'Epoch {it+1}/{num_epochs}, Train Loss: {loss.item():.4f}, Validation Loss: {validation_loss.item():.4f}')

    return train_losses, validation_losses

Results I get

The issue seems to be the decoder is predicting a constant value each time and does not pick up on the noise in the data

enter image description here

  • $\begingroup$ It seems you have a couple of different questions, if you are able to more clearly outline them that would be help us assist you more. Can you post your training routine as well? $\endgroup$
    – hH1sG0n3
    Jun 22, 2021 at 15:46
  • $\begingroup$ Hi, yes so my main focus is being able to derive a working Seq2Seq model with arbitrary input size and an output size that can be specified by me (or anyone else). So my issues are that I cannot get a custom output size e.g. 3 months forecast and that the current forecasts seem to just be the same value at every time step. $\endgroup$ Jun 22, 2021 at 16:28

1 Answer 1


Breaking down a number of questions, firstly

I want to forecast 3 months into the future.

You need at least 3+ months worth of data to do this task. That means, your "forecast horizon" needs to be a subset of your data set in which you can define how far ahead you want to be making predictions. See for example the image below: enter image description here

the Decoder should be taking the previous time-steps production levels as input (meaning dimension 1), as well as the previous hidden and cell state.

I think you are mixing up attention pooling architecture with self-attention/transformer architecture concepts.

  1. In typical encoder/decoder rnn, the decoder needs to consume the hidden state output from each LSTM cell/timestep. Unless you are trying a more experimental output, the cell state as well as the input data should not be passed to the decoder. If you are using Luong/additive attention pooling, again only the hidden state is needed for its calculation. I am not sure your attention method is entirely correct. I paste below an example of hoe your decoder with additive attention should look like:
class AttnDecoderRNN(nn.Module):
    Courtesy of https://pytorch.org/tutorials/intermediate/seq2seq_translation_tutorial.html

    def __init__(self, hidden_size, output_size, dropout_p=0.1, max_length=MAX_LENGTH):
        super(AttnDecoderRNN, self).__init__()
        self.hidden_size = hidden_size
        self.output_size = output_size
        self.dropout_p = dropout_p
        self.max_length = max_length

        self.attn = nn.Linear(self.hidden_size * 2, self.max_length)
        self.attn_combine = nn.Linear(self.hidden_size * 2, self.hidden_size)
        self.dropout = nn.Dropout(self.dropout_p)
        self.gru = nn.GRU(self.hidden_size, self.hidden_size)
        self.out = nn.Linear(self.hidden_size, self.output_size)

    def forward(self, input, hidden, encoder_outputs):

        # Here is the attention pooling calculation
        attn_weights = F.softmax(
            self.attn(torch.cat((embedded[0], hidden[0]), 1)), dim=1)
        attn_applied = torch.bmm(attn_weights.unsqueeze(0),

        output = torch.cat((embedded[0], attn_applied[0]), 1)
        output = self.attn_combine(output).unsqueeze(0)

        output = F.relu(output)
        output, hidden = self.gru(output, hidden)

        output = F.log_softmax(self.out(output[0]), dim=1)
        return output, hidden, attn_weights

    def initHidden(self):
        return torch.zeros(1, 1, self.hidden_size, device=device)
  1. Rightfully so, you claim that input and hidden state needs to be passed to the decoder and I presume you refer to a self-attention architecture. For this model, there are no decoder/decoder units but rather each input is abstracted into a query, key and value latent spaces, as in the following diagram: enter image description here You can read more about the self-attention architecture here https://towardsdatascience.com/self-attention-5b95ea164f61, before you decide how you want to proceed attacking your problem.
  • $\begingroup$ Hi @hH1sG0n3 , thanks for responding. Addressing each comment you have made, 1. I have about 20 years of data so that shouldn't be an issue and the 3 months was just arbitrary, I want the architecture to be flexible so that I can forecast x months into the future. 2. I am not using this for NLP tasks so how is the embedding layer relevant? I am passing numbers directly to the model. 3. Overall, my encoder should yield a context vector which is used at each time step of the decoder with attention so to make the most accurate prediction $\endgroup$ Jun 23, 2021 at 10:30
  • $\begingroup$ Oh i see - sorry you say "The encoder takes as input a sequence of length 168, with each input being the 20 previous days data" and then define a 3 month horizon which is misleading. You are correct about the embedding in the decoder, I will remove this mistake. On three, "yield a context vector which is used at each time step of the decoder with attention" only hidden states from each timestep are required for the calculation of the context vector and not the inputs, unless self attention. $\endgroup$
    – hH1sG0n3
    Jun 23, 2021 at 10:42
  • $\begingroup$ Apologies for the confusion but essentially, my input data is of the form (seq length, batch size, number of features), so in my example my sequence length was 168, with 20 months previous data (batch) and 37 features. Please also see a diagram that I have uploaded of how I have visualised the architecture, I think it better explains my point of a context vector and inputs. During training, the decoder needs to take the output of the previous LSTM cell as input, alongside the context vector to make a prediction $\endgroup$ Jun 23, 2021 at 10:52
  • $\begingroup$ Hi, @hH1sG0n3, please see my comment and the edits in the original post $\endgroup$ Jun 23, 2021 at 17:10
  • $\begingroup$ Thank you, makes more sense now. I will try to post a reply as soon as i find sometime. $\endgroup$
    – hH1sG0n3
    Jun 24, 2021 at 8:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.