So after running the training process for few hours on GPU, you can see that the accuracy has reached the value 1.0 and loss has significantly reduced to 0.00045. Let's see how the model performs when we ask some generic questions.

To make predictions we will create a predict() which will take the raw text of any size as input and returns the response to the question that we asked. We did a quick fix to handle the out of vocab (OOV) words by replacing them with the PAD.

def predict(sentence):
    X_in = []
    for word in sentence.split():
        try:
            X_in.append(dictionary_from[word])
        except:
            X_in.append(PAD)
            pass
        
    test, seq_x = pad_sentence_batch([X_in], PAD)
    input_batch = np.zeros([batch_size,seq_x[0]])
    input_batch[0] =test[0] 
        
    log = sess.run(tf.argmax(model.logits,2), 
                                      feed_dict={
                                              model.X:input_batch,
                                              model.X_seq_len:seq_x,
                                              model.Y_seq_len:seq_x
                                              }
                                      )
    
    result=' '.join(rev_dictionary_to[i] for i in log[0])
    return result

When the model was trained for first 50 epochs, we had the following result:

>> predict('where do you live')
>> i PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD

>> print predict('how are you ?')
>> i am PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD

When the model was trained for 1136 epochs:

>> predict('where do you live')
>> miami florida PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD

>> print predict('how are you ?')
>> i am fine thank you PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD PAD

Well!! That's impressive right. Now your model is not just able to understand the context but can also generate answers words by word.