This step is really similar to the training one. The first difference is that we don't make any evaluation of our predictions, but instead use the input to generate the results. The second difference is that we use the already trained set of variables to yield this result. You will see how it is done later in this chapter. 

To make things clearer, we first initialize a new sequence-to-sequence model. Its purpose is to use the already trained weights and biases and make predictions based on different sets of inputs. We only have an encoder and decoder sequence, where the encoder one is an input sentence and the decoder sequence is fed one word at a time. We define the new model as follows:

encode_seqs2 = tf.placeholder(dtype=tf.int64, shape=[1, None], name="encode_seqs")
decode_seqs2 = tf.placeholder(dtype=tf.int64, shape=[1, None], name...