I am using Pytorch for an LSTM encoder-decoder sequence-to-sequence prediction problem. As a first step, I would like to forecast 2D trajectories (trajectory x, trajectory y) from multivariate input - 2-D or more (trajectory x, trajectory y, speed, rotation, etc.)
I am following the below notebook (link):
Here excerpts (encoder, decoder, attention):
class EncoderRNN(nn.Module):
def __init__(self, input_size, hidden_size, n_layers=1, dropout=0.1):
super(EncoderRNN, self).__init__()
self.input_size = input_size
self.hidden_size = hidden_size
self.n_layers = n_layers
self.dropout = dropout
self.embedding = nn.Embedding(input_size, hidden_size)
self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=self.dropout, bidirectional=True)
def forward(self, input_seqs, input_lengths, hidden=None):
# Note: we run this all at once (over multiple batches of multiple sequences)
embedded = self.embedding(input_seqs)
packed = torch.nn.utils.rnn.pack_padded_sequence(embedded, input_lengths)
outputs, hidden = self.gru(packed, hidden)
outputs, output_lengths = torch.nn.utils.rnn.pad_packed_sequence(outputs) # unpack (back to padded)
outputs = outputs[:, :, :self.hidden_size] + outputs[:, : ,self.hidden_size:] # Sum bidirectional outputs
return outputs, hidden
class LuongAttnDecoderRNN(nn.Module):
def __init__(self, attn_model, hidden_size, output_size, n_layers=1, dropout=0.1):
super(LuongAttnDecoderRNN, self).__init__()
# Keep for reference
self.attn_model = attn_model
self.hidden_size = hidden_size
self.output_size = output_size
self.n_layers = n_layers
self.dropout = dropout
# Define layers
self.embedding = nn.Embedding(output_size, hidden_size)
self.embedding_dropout = nn.Dropout(dropout)
self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=dropout)
self.concat = nn.Linear(hidden_size * 2, hidden_size)
self.out = nn.Linear(hidden_size, output_size)
# Choose attention model
if attn_model != 'none':
self.attn = Attn(attn_model, hidden_size)
def forward(self, input_seq, last_hidden, encoder_outputs):
# Note: we run this one step at a time
# Get the embedding of the current input word (last output word)
batch_size = input_seq.size(0)
embedded = self.embedding(input_seq)
embedded = self.embedding_dropout(embedded)
embedded = embedded.view(1, batch_size, self.hidden_size) # S=1 x B x N
# Get current hidden state from input word and last hidden state
rnn_output, hidden = self.gru(embedded, last_hidden)
# Calculate attention from current RNN state and all encoder outputs;
# apply to encoder outputs to get weighted average
attn_weights = self.attn(rnn_output, encoder_outputs)
context = attn_weights.bmm(encoder_outputs.transpose(0, 1)) # B x S=1 x N
# Attentional vector using the RNN hidden state and context vector
# concatenated together (Luong eq. 5)
rnn_output = rnn_output.squeeze(0) # S=1 x B x N -> B x N
context = context.squeeze(1) # B x S=1 x N -> B x N
concat_input = torch.cat((rnn_output, context), 1)
concat_output = F.tanh(self.concat(concat_input))
# Finally predict next token (Luong eq. 6, without softmax)
output = self.out(concat_output)
# Return final output, hidden state, and attention weights (for visualization)
return output, hidden, attn_weights
For calculating attention in the decoder stage, the encoder hidden state and encoder outputs are input and used as below:
class Attn(nn.Module):
def __init__(self, method, hidden_size):
super(Attn, self).__init__()
self.method = method
self.hidden_size = hidden_size
if self.method == 'general':
self.attn = nn.Linear(self.hidden_size, hidden_size)
elif self.method == 'concat':
self.attn = nn.Linear(self.hidden_size * 2, hidden_size)
self.v = nn.Parameter(torch.FloatTensor(1, hidden_size))
def forward(self, hidden, encoder_outputs):
max_len = encoder_outputs.size(0)
this_batch_size = encoder_outputs.size(1)
# Create variable to store attention energies
attn_energies = Variable(torch.zeros(this_batch_size, max_len)) # B x S
if USE_CUDA:
attn_energies = attn_energies.cuda()
# For each batch of encoder outputs
for b in range(this_batch_size):
# Calculate energy for each encoder output
for i in range(max_len):
attn_energies[b, i] = self.score(hidden[:, b], encoder_outputs[i, b].unsqueeze(0))
# Normalize energies to weights in range 0 to 1, resize to 1 x B x S
return F.softmax(attn_energies).unsqueeze(1)
def score(self, hidden, encoder_output):
if self.method == 'dot':
energy = hidden.dot(encoder_output)
return energy
elif self.method == 'general':
energy = self.attn(encoder_output)
energy = hidden.dot(energy)
return energy
elif self.method == 'concat':
energy = self.attn(torch.cat((hidden, encoder_output), 1))
energy = self.v.dot(energy)
return energy
My actual goal is to extend the method by adding further information to be fed into the decoder, such as image data at each input time step. Technically, I want to use two (or more) encoders, one for the trajectories as in the link above, and one separate one for image data (convolutional encoder).
I do this by concatenating embeddings produced by the trajectory encoder and the convolutional encoder (as well as the cell states etc.) and feeding the concatenated tensors to the decoder.
For example, image embedding (256-length tensor) concatenated with trajectory data embedding (256-length tensor) yields a 512-length embedding.
My question is: is it a problem for the attention calculation if I use a concatenated encoder hidden state, concatenated encoder cell state, and concatenated encoder output coming from those different sources rather than hidden states, cells, outputs coming from a single source?
What are the caveats or pre-processing that should happen to make this work?
Thank you very much in advance.
