Training procedure for LSTM model

train.py

# Define training procedure
def train(sequence, target, device):
    # Move tensors to device
    hidden = rnn.initHidden(device)
    sequence = sequence.to(device)
    target = target.to(device)

    rnn.zero_grad()

    # Forward step
    for i in range(sequence.size()[0]):
        output, hidden = rnn(sequence[i], hidden)
        
    output, hidden = rnn(sequence[i], hidden)
    
    loss = criterion(output, indexFromTensor(target).to(device))
    loss.backward()

    # Add parameters' gradients to their values, multiplied by learning rate
    for p in rnn.parameters():
        p.data.add_(-learning_rate, p.grad.data)

    return output, loss.item()