How to use Tensorflow's PTB model example?

Question

I'm trying out Tensorflow's rnn example. With some problems at the start I could run the example in order to train the ptb and now I have a model trained.

How do I exactly use the model now to create sentences without having to train every time again?

I'm running it with a command like python ptb_word_lm.py --data_path=/home/data/ --model medium --save_path=/home/medium

Is there a example somewhere on how to use the trained model to make sentences?

Answer 1

1.Add the following code at the last line of PTBModel:__init__() function:

self._output_probs = tf.nn.softmax(logits)

2.Add the following function in PTBModel:

@property
def output_probs(self):
    return self._output_probs

3.Try to run the following code:

raw_data = reader.ptb_raw_data(FLAGS.data_path)
train_data, valid_data, test_data, vocabulary, word_to_id, id_to_word = raw_data

eval_config = get_config()
eval_config.batch_size = 1
eval_config.num_steps = 1

sess = tf.Session()

initializer = tf.random_uniform_initializer(-eval_config.init_scale,
                                            eval_config.init_scale)
with tf.variable_scope("model", reuse=None, initializer=initializer):
    mtest = PTBModel(is_training=False, config=eval_config)

sess.run(tf.initialize_all_variables())

saver = tf.train.Saver()

ckpt = tf.train.get_checkpoint_state('/home/medium')  # __YOUR__MODEL__SAVE__PATH__
if ckpt and gfile.Exists(ckpt.model_checkpoint_path):
    msg = 'Reading model parameters from %s' % ckpt.model_checkpoint_path
    print(msg)
    saver.restore(sess, ckpt.model_checkpoint_path)

def pick_from_weight(weight, pows=1.0):
    weight = weight**pows
    t = np.cumsum(weight)
    s = np.sum(weight)
    return int(np.searchsorted(t, np.random.rand(1) * s))

while True:
    number_of_sentences = 10  # generate 10 sentences one time
    sentence_cnt = 0
    text = '\n'
    end_of_sentence_char = word_to_id['<eos>']
    input_char = np.array([[end_of_sentence_char]])
    state = sess.run(mtest.initial_state)
    while sentence_cnt < number_of_sentences:
        feed_dict = {mtest.input_data: input_char,
                     mtest.initial_state: state}
        probs, state = sess.run([mtest.output_probs, mtest.final_state],
                                       feed_dict=feed_dict)
        sampled_char = pick_from_weight(probs[0])
        if sampled_char == end_of_sentence_char:
            text += '.\n'
            sentence_cnt += 1
        else:
            text += ' ' + id_to_word[sampled_char]
        input_char = np.array([[sampled_char]])
    print(text)
    raw_input('press any key to continue ...')

Answer 2

This website has an answer with an alteration on the PTB word script, working currently with Tensorflow version 1.0

http://deeplearningathome.com/2016/10/Text-generation-using-deep-recurrent-neural-networks.html