For LSTM in pytorch, first I figure out that the size of LSTM hidden_states is [num_layers*num_direction, batch_size, hidden_features]. So how do I extract the hidden state of the sequence? Now my understanding of doing this is as follow, is that right? I cannot find any explanatory documents about this please help me.
# For multi-layer single directional LSTM network
rnn = nn.LSTM(input_size, hidden_features, num_layers, batch_first=True, bidirectional=False);
input = initializeInput();
# input.size() = [batch_size, seq_len, hidden_features]
hidden_cell = initializeHiddenCell();
# hidden_cell.size() = ([num_layers*num_direction, batch_size, hidden_features], [num_layers*num_direction, batch_size, hidden_features])
# input_size == hidden_features
output, hidden_cell = rnn(input, hidden_cell);
hidden_state for current timestep input is: hidden_cell[0][-1];
# For multi-layer bidirectional LSTM network
rnn = nn.LSTM(input_size, hidden_features, num_layers, batch_first=True, bidirectional=False);
input = initializeInput();
# input.size() = [batch_size, seq_len, hidden_features]
hidden_cell = initializeHiddenCell();
# hidden_cell.size() = ([num_layers*num_direction, batch_size, hidden_features], [num_layers*num_direction, batch_size, hidden_features])
# input_size == hidden_features
output, hidden_cell = rnn(input, hidden_cell);
forward and backward hidden_state for current timestep input is: hidden_cell[0][-1], hidden_cell[0][-2];