def generate(model_path, reversed_dictionary, gen_word_count, predict_full_sentence=True,start_word=0):
    test_input = Input(batch_size=1, num_steps=num_steps, data=test_data)

    x = tf.placeholder(dtype=tf.int32, shape=[None, None])
    y = tf.placeholder(dtype=tf.int32, shape=[None, None])

    m = Model(x, y, input=test_input, is_training=False, vocab_size=vocabulary)
    saver = tf.train.Saver()
    with tf.Session() as sess:
        current_state_fw = np.zeros((NUM_LAYERS, 2, 1, HIDDEN_SIZE))
        current_state_bw = np.zeros((NUM_LAYERS, 2, 1, HIDDEN_SIZE))

        # restore the trained model
        saver.restore(sess, model_path)

        sentence = start_word    # 0 is 'the'
        sentence = np.reshape(sentence, (-1, 1))
        for step in range(gen_word_count):
            if predict_full_sentence:
                pred, current_state_fw, current_state_bw = sess.run([m.predict, m.state_fw, m.state_bw],
                                               feed_dict={m.init_state_fw: current_state_fw,
                                                            m.init_state_bw: current_state_bw,
                                                          x: sentence})
            else:

                pred, current_state_fw, current_state_bw = sess.run([m.predict, m.state_fw, m.state_bw],
                                               feed_dict={m.init_state_fw: current_state_fw,
                                                            m.init_state_bw: current_state_bw,
                                                          x: np.reshape(sentence[:,-1], (-1, 1))})
            sentence = np.append(sentence, pred[-1])
            sentence = np.reshape(sentence, (1, -1))
        print(" ".join([reversed_dictionary[x] for x in sentence[0]]))