# Feed the target as the next input
for di in range(target_length):

    # The ouput of different steps
    decoder_output, decoder_hidden, decoder_attention = decoder(
        decoder_input, decoder_hidden, encoder_outputs)

    # Calculate the NLLLoss fro each step
    loss += criterion(decoder_output, summary[:, di])

    decoder_input = summary[:, di]  # Supervised

# Calculate the gradient
loss.backward()