Anna KaRNNa Name Scoped


Anna KaRNNa

In this notebook, I'll build a character-wise RNN trained on Anna Karenina, one of my all-time favorite books. It'll be able to generate new text based on the text from the book.

This network is based off of Andrej Karpathy's post on RNNs and implementation in Torch. Also, some information here at r2rt and from Sherjil Ozair on GitHub. Below is the general architecture of the character-wise RNN.


In [3]:
import time
from collections import namedtuple

import numpy as np
import tensorflow as tf

First we'll load the text file and convert it into integers for our network to use.


In [4]:
with open('anna.txt', 'r') as f:
    text=f.read()
vocab = set(text)
vocab_to_int = {c: i for i, c in enumerate(vocab)}
int_to_vocab = dict(enumerate(vocab))
chars = np.array([vocab_to_int[c] for c in text], dtype=np.int32)

In [5]:
text[:100]


Out[5]:
'Chapter 1\n\n\nHappy families are all alike; every unhappy family is unhappy in its own\nway.\n\nEverythin'

In [6]:
chars[:100]


Out[6]:
array([61,  2, 34, 18, 38, 50, 37, 25, 16, 72, 72, 72, 15, 34, 18, 18, 56,
       25, 26, 34, 42, 22, 17, 22, 50, 41, 25, 34, 37, 50, 25, 34, 17, 17,
       25, 34, 17, 22, 44, 50, 49, 25, 50, 69, 50, 37, 56, 25, 39, 35,  2,
       34, 18, 18, 56, 25, 26, 34, 42, 22, 17, 56, 25, 22, 41, 25, 39, 35,
        2, 34, 18, 18, 56, 25, 22, 35, 25, 22, 38, 41, 25, 11, 60, 35, 72,
       60, 34, 56, 73, 72, 72, 31, 69, 50, 37, 56, 38,  2, 22, 35], dtype=int32)

Now I need to split up the data into batches, and into training and validation sets. I should be making a test set here, but I'm not going to worry about that. My test will be if the network can generate new text.

Here I'll make both input and target arrays. The targets are the same as the inputs, except shifted one character over. I'll also drop the last bit of data so that I'll only have completely full batches.

The idea here is to make a 2D matrix where the number of rows is equal to the number of batches. Each row will be one long concatenated string from the character data. We'll split this data into a training set and validation set using the split_frac keyword. This will keep 90% of the batches in the training set, the other 10% in the validation set.


In [7]:
def split_data(chars, batch_size, num_steps, split_frac=0.9):
    """ 
    Split character data into training and validation sets, inputs and targets for each set.
    
    Arguments
    ---------
    chars: character array
    batch_size: Size of examples in each of batch
    num_steps: Number of sequence steps to keep in the input and pass to the network
    split_frac: Fraction of batches to keep in the training set
    
    
    Returns train_x, train_y, val_x, val_y
    """
    
    
    slice_size = batch_size * num_steps
    n_batches = int(len(chars) / slice_size)
    
    # Drop the last few characters to make only full batches
    x = chars[: n_batches*slice_size]
    y = chars[1: n_batches*slice_size + 1]
    
    # Split the data into batch_size slices, then stack them into a 2D matrix 
    x = np.stack(np.split(x, batch_size))
    y = np.stack(np.split(y, batch_size))
    
    # Now x and y are arrays with dimensions batch_size x n_batches*num_steps
    
    # Split into training and validation sets, keep the virst split_frac batches for training
    split_idx = int(n_batches*split_frac)
    train_x, train_y= x[:, :split_idx*num_steps], y[:, :split_idx*num_steps]
    val_x, val_y = x[:, split_idx*num_steps:], y[:, split_idx*num_steps:]
    
    return train_x, train_y, val_x, val_y

In [8]:
train_x, train_y, val_x, val_y = split_data(chars, 10, 200)

In [9]:
train_x.shape


Out[9]:
(10, 178400)

In [10]:
train_x[:,:10]


Out[10]:
array([[61,  2, 34, 18, 38, 50, 37, 25, 16, 72],
       [19, 35,  9, 25,  2, 50, 25, 42, 11, 69],
       [25, 23, 34, 38, 23,  2, 22, 35, 28, 25],
       [11, 38,  2, 50, 37, 25, 60, 11, 39, 17],
       [25, 38,  2, 50, 25, 17, 34, 35,  9, 12],
       [25, 13,  2, 37, 11, 39, 28,  2, 25, 17],
       [38, 25, 38, 11, 72,  9, 11, 73, 72, 72],
       [11, 25,  2, 50, 37, 41, 50, 17, 26, 63],
       [ 2, 34, 38, 25, 22, 41, 25, 38,  2, 50],
       [50, 37, 41, 50, 17, 26, 25, 34, 35,  9]], dtype=int32)

I'll write another function to grab batches out of the arrays made by split data. Here each batch will be a sliding window on these arrays with size batch_size X num_steps. For example, if we want our network to train on a sequence of 100 characters, num_steps = 100. For the next batch, we'll shift this window the next sequence of num_steps characters. In this way we can feed batches to the network and the cell states will continue through on each batch.


In [11]:
def get_batch(arrs, num_steps):
    batch_size, slice_size = arrs[0].shape
    
    n_batches = int(slice_size/num_steps)
    for b in range(n_batches):
        yield [x[:, b*num_steps: (b+1)*num_steps] for x in arrs]

In [12]:
def build_rnn(num_classes, batch_size=50, num_steps=50, lstm_size=128, num_layers=2,
              learning_rate=0.001, grad_clip=5, sampling=False):
        
    if sampling == True:
        batch_size, num_steps = 1, 1

    tf.reset_default_graph()
    
    # Declare placeholders we'll feed into the graph
    with tf.name_scope('inputs'):
        inputs = tf.placeholder(tf.int32, [batch_size, num_steps], name='inputs')
        x_one_hot = tf.one_hot(inputs, num_classes, name='x_one_hot')
    
    with tf.name_scope('targets'):
        targets = tf.placeholder(tf.int32, [batch_size, num_steps], name='targets')
        y_one_hot = tf.one_hot(targets, num_classes, name='y_one_hot')
        y_reshaped = tf.reshape(y_one_hot, [-1, num_classes])
    
    keep_prob = tf.placeholder(tf.float32, name='keep_prob')
    
    # Build the RNN layers
    with tf.name_scope("RNN_layers"):
        lstm = tf.contrib.rnn.BasicLSTMCell(lstm_size)
        drop = tf.contrib.rnn.DropoutWrapper(lstm, output_keep_prob=keep_prob)
        cell = tf.contrib.rnn.MultiRNNCell([drop] * num_layers)
    
    with tf.name_scope("RNN_init_state"):
        initial_state = cell.zero_state(batch_size, tf.float32)

    # Run the data through the RNN layers
    with tf.name_scope("RNN_forward"):
        rnn_inputs = [tf.squeeze(i, squeeze_dims=[1]) for i in tf.split(x_one_hot, num_steps, 1)]
        outputs, state = tf.contrib.rnn.static_rnn(cell, rnn_inputs, initial_state=initial_state)
    
    final_state = state
    
    # Reshape output so it's a bunch of rows, one row for each cell output
    with tf.name_scope('sequence_reshape'):
        seq_output = tf.concat(outputs, axis=1,name='seq_output')
        output = tf.reshape(seq_output, [-1, lstm_size], name='graph_output')
    
    # Now connect the RNN putputs to a softmax layer and calculate the cost
    with tf.name_scope('logits'):
        softmax_w = tf.Variable(tf.truncated_normal((lstm_size, num_classes), stddev=0.1),
                               name='softmax_w')
        softmax_b = tf.Variable(tf.zeros(num_classes), name='softmax_b')
        logits = tf.matmul(output, softmax_w) + softmax_b

    with tf.name_scope('predictions'):
        preds = tf.nn.softmax(logits, name='predictions')
    
    
    with tf.name_scope('cost'):
        loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y_reshaped, name='loss')
        cost = tf.reduce_mean(loss, name='cost')

    # Optimizer for training, using gradient clipping to control exploding gradients
    with tf.name_scope('train'):
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), grad_clip)
        train_op = tf.train.AdamOptimizer(learning_rate)
        optimizer = train_op.apply_gradients(zip(grads, tvars))
    
    # Export the nodes 
    export_nodes = ['inputs', 'targets', 'initial_state', 'final_state',
                    'keep_prob', 'cost', 'preds', 'optimizer']
    Graph = namedtuple('Graph', export_nodes)
    local_dict = locals()
    graph = Graph(*[local_dict[each] for each in export_nodes])
    
    return graph

Hyperparameters

Here I'm defining the hyperparameters for the network. The two you probably haven't seen before are lstm_size and num_layers. These set the number of hidden units in the LSTM layers and the number of LSTM layers, respectively. Of course, making these bigger will improve the network's performance but you'll have to watch out for overfitting. If your validation loss is much larger than the training loss, you're probably overfitting. Decrease the size of the network or decrease the dropout keep probability.


In [13]:
batch_size = 100
num_steps = 100
lstm_size = 512
num_layers = 2
learning_rate = 0.001

Write out the graph for TensorBoard


In [14]:
model = build_rnn(len(vocab), 
                  batch_size=batch_size,
                  num_steps=num_steps,
                  learning_rate=learning_rate,
                  lstm_size=lstm_size,
                  num_layers=num_layers)

with tf.Session() as sess:
    
    sess.run(tf.global_variables_initializer())
    file_writer = tf.summary.FileWriter('./logs/3', sess.graph)

Training

Time for training which is is pretty straightforward. Here I pass in some data, and get an LSTM state back. Then I pass that state back in to the network so the next batch can continue the state from the previous batch. And every so often (set by save_every_n) I calculate the validation loss and save a checkpoint.


In [12]:
!mkdir -p checkpoints/anna

In [15]:
epochs = 10
save_every_n = 200
train_x, train_y, val_x, val_y = split_data(chars, batch_size, num_steps)

model = build_rnn(len(vocab), 
                  batch_size=batch_size,
                  num_steps=num_steps,
                  learning_rate=learning_rate,
                  lstm_size=lstm_size,
                  num_layers=num_layers)

saver = tf.train.Saver(max_to_keep=100)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    
    # Use the line below to load a checkpoint and resume training
    #saver.restore(sess, 'checkpoints/anna20.ckpt')
    
    n_batches = int(train_x.shape[1]/num_steps)
    iterations = n_batches * epochs
    for e in range(epochs):
        
        # Train network
        new_state = sess.run(model.initial_state)
        loss = 0
        for b, (x, y) in enumerate(get_batch([train_x, train_y], num_steps), 1):
            iteration = e*n_batches + b
            start = time.time()
            feed = {model.inputs: x,
                    model.targets: y,
                    model.keep_prob: 0.5,
                    model.initial_state: new_state}
            batch_loss, new_state, _ = sess.run([model.cost, model.final_state, model.optimizer], 
                                                 feed_dict=feed)
            loss += batch_loss
            end = time.time()
            print('Epoch {}/{} '.format(e+1, epochs),
                  'Iteration {}/{}'.format(iteration, iterations),
                  'Training loss: {:.4f}'.format(loss/b),
                  '{:.4f} sec/batch'.format((end-start)))
        
            
            if (iteration%save_every_n == 0) or (iteration == iterations):
                # Check performance, notice dropout has been set to 1
                val_loss = []
                new_state = sess.run(model.initial_state)
                for x, y in get_batch([val_x, val_y], num_steps):
                    feed = {model.inputs: x,
                            model.targets: y,
                            model.keep_prob: 1.,
                            model.initial_state: new_state}
                    batch_loss, new_state = sess.run([model.cost, model.final_state], feed_dict=feed)
                    val_loss.append(batch_loss)

                print('Validation loss:', np.mean(val_loss),
                      'Saving checkpoint!')
                saver.save(sess, "checkpoints/anna/i{}_l{}_{:.3f}.ckpt".format(iteration, lstm_size, np.mean(val_loss)))


Epoch 1/10  Iteration 1/1780 Training loss: 4.4195 1.3313 sec/batch
Epoch 1/10  Iteration 2/1780 Training loss: 4.3756 0.1287 sec/batch
Epoch 1/10  Iteration 3/1780 Training loss: 4.2069 0.1276 sec/batch
Epoch 1/10  Iteration 4/1780 Training loss: 4.5396 0.1185 sec/batch
Epoch 1/10  Iteration 5/1780 Training loss: 4.4190 0.1206 sec/batch
Epoch 1/10  Iteration 6/1780 Training loss: 4.3547 0.1233 sec/batch
Epoch 1/10  Iteration 7/1780 Training loss: 4.2792 0.1188 sec/batch
Epoch 1/10  Iteration 8/1780 Training loss: 4.2018 0.1170 sec/batch
Epoch 1/10  Iteration 9/1780 Training loss: 4.1251 0.1187 sec/batch
Epoch 1/10  Iteration 10/1780 Training loss: 4.0558 0.1174 sec/batch
Epoch 1/10  Iteration 11/1780 Training loss: 3.9946 0.1190 sec/batch
Epoch 1/10  Iteration 12/1780 Training loss: 3.9451 0.1193 sec/batch
Epoch 1/10  Iteration 13/1780 Training loss: 3.9011 0.1210 sec/batch
Epoch 1/10  Iteration 14/1780 Training loss: 3.8632 0.1185 sec/batch
Epoch 1/10  Iteration 15/1780 Training loss: 3.8275 0.1199 sec/batch
Epoch 1/10  Iteration 16/1780 Training loss: 3.7945 0.1211 sec/batch
Epoch 1/10  Iteration 17/1780 Training loss: 3.7649 0.1215 sec/batch
Epoch 1/10  Iteration 18/1780 Training loss: 3.7400 0.1214 sec/batch
Epoch 1/10  Iteration 19/1780 Training loss: 3.7164 0.1247 sec/batch
Epoch 1/10  Iteration 20/1780 Training loss: 3.6933 0.1212 sec/batch
Epoch 1/10  Iteration 21/1780 Training loss: 3.6728 0.1203 sec/batch
Epoch 1/10  Iteration 22/1780 Training loss: 3.6538 0.1207 sec/batch
Epoch 1/10  Iteration 23/1780 Training loss: 3.6359 0.1200 sec/batch
Epoch 1/10  Iteration 24/1780 Training loss: 3.6198 0.1229 sec/batch
Epoch 1/10  Iteration 25/1780 Training loss: 3.6041 0.1204 sec/batch
Epoch 1/10  Iteration 26/1780 Training loss: 3.5904 0.1202 sec/batch
Epoch 1/10  Iteration 27/1780 Training loss: 3.5774 0.1189 sec/batch
Epoch 1/10  Iteration 28/1780 Training loss: 3.5642 0.1214 sec/batch
Epoch 1/10  Iteration 29/1780 Training loss: 3.5522 0.1231 sec/batch
Epoch 1/10  Iteration 30/1780 Training loss: 3.5407 0.1199 sec/batch
Epoch 1/10  Iteration 31/1780 Training loss: 3.5309 0.1180 sec/batch
Epoch 1/10  Iteration 32/1780 Training loss: 3.5207 0.1179 sec/batch
Epoch 1/10  Iteration 33/1780 Training loss: 3.5109 0.1224 sec/batch
Epoch 1/10  Iteration 34/1780 Training loss: 3.5021 0.1206 sec/batch
Epoch 1/10  Iteration 35/1780 Training loss: 3.4931 0.1241 sec/batch
Epoch 1/10  Iteration 36/1780 Training loss: 3.4850 0.1169 sec/batch
Epoch 1/10  Iteration 37/1780 Training loss: 3.4767 0.1204 sec/batch
Epoch 1/10  Iteration 38/1780 Training loss: 3.4688 0.1202 sec/batch
Epoch 1/10  Iteration 39/1780 Training loss: 3.4611 0.1213 sec/batch
---------------------------------------------------------------------------
KeyboardInterrupt                         Traceback (most recent call last)
<ipython-input-15-09fa3beeed23> in <module>()
     33                     model.initial_state: new_state}
     34             batch_loss, new_state, _ = sess.run([model.cost, model.final_state, model.optimizer], 
---> 35                                                  feed_dict=feed)
     36             loss += batch_loss
     37             end = time.time()

/home/mat/miniconda3/envs/tf-gpu/lib/python3.5/site-packages/tensorflow/python/client/session.py in run(self, fetches, feed_dict, options, run_metadata)
    765     try:
    766       result = self._run(None, fetches, feed_dict, options_ptr,
--> 767                          run_metadata_ptr)
    768       if run_metadata:
    769         proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)

/home/mat/miniconda3/envs/tf-gpu/lib/python3.5/site-packages/tensorflow/python/client/session.py in _run(self, handle, fetches, feed_dict, options, run_metadata)
    963     if final_fetches or final_targets:
    964       results = self._do_run(handle, final_targets, final_fetches,
--> 965                              feed_dict_string, options, run_metadata)
    966     else:
    967       results = []

/home/mat/miniconda3/envs/tf-gpu/lib/python3.5/site-packages/tensorflow/python/client/session.py in _do_run(self, handle, target_list, fetch_list, feed_dict, options, run_metadata)
   1013     if handle is None:
   1014       return self._do_call(_run_fn, self._session, feed_dict, fetch_list,
-> 1015                            target_list, options, run_metadata)
   1016     else:
   1017       return self._do_call(_prun_fn, self._session, handle, feed_dict,

/home/mat/miniconda3/envs/tf-gpu/lib/python3.5/site-packages/tensorflow/python/client/session.py in _do_call(self, fn, *args)
   1020   def _do_call(self, fn, *args):
   1021     try:
-> 1022       return fn(*args)
   1023     except errors.OpError as e:
   1024       message = compat.as_text(e.message)

/home/mat/miniconda3/envs/tf-gpu/lib/python3.5/site-packages/tensorflow/python/client/session.py in _run_fn(session, feed_dict, fetch_list, target_list, options, run_metadata)
   1002         return tf_session.TF_Run(session, options,
   1003                                  feed_dict, fetch_list, target_list,
-> 1004                                  status, run_metadata)
   1005 
   1006     def _prun_fn(session, handle, feed_dict, fetch_list):

KeyboardInterrupt: 

In [35]:
tf.train.get_checkpoint_state('checkpoints/anna')


Out[35]:
model_checkpoint_path: "checkpoints/anna/i3560_l512_1.122.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i200_l512_2.432.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i400_l512_1.980.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i600_l512_1.750.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i800_l512_1.595.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i1000_l512_1.484.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i1200_l512_1.407.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i1400_l512_1.349.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i1600_l512_1.292.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i1800_l512_1.255.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i2000_l512_1.224.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i2200_l512_1.204.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i2400_l512_1.187.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i2600_l512_1.172.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i2800_l512_1.160.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i3000_l512_1.148.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i3200_l512_1.137.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i3400_l512_1.129.ckpt"
all_model_checkpoint_paths: "checkpoints/anna/i3560_l512_1.122.ckpt"

Sampling

Now that the network is trained, we'll can use it to generate new text. The idea is that we pass in a character, then the network will predict the next character. We can use the new one, to predict the next one. And we keep doing this to generate all new text. I also included some functionality to prime the network with some text by passing in a string and building up a state from that.

The network gives us predictions for each character. To reduce noise and make things a little less random, I'm going to only choose a new character from the top N most likely characters.


In [17]:
def pick_top_n(preds, vocab_size, top_n=5):
    p = np.squeeze(preds)
    p[np.argsort(p)[:-top_n]] = 0
    p = p / np.sum(p)
    c = np.random.choice(vocab_size, 1, p=p)[0]
    return c

In [41]:
def sample(checkpoint, n_samples, lstm_size, vocab_size, prime="The "):
    prime = "Far"
    samples = [c for c in prime]
    model = build_rnn(vocab_size, lstm_size=lstm_size, sampling=True)
    saver = tf.train.Saver()
    with tf.Session() as sess:
        saver.restore(sess, checkpoint)
        new_state = sess.run(model.initial_state)
        for c in prime:
            x = np.zeros((1, 1))
            x[0,0] = vocab_to_int[c]
            feed = {model.inputs: x,
                    model.keep_prob: 1.,
                    model.initial_state: new_state}
            preds, new_state = sess.run([model.preds, model.final_state], 
                                         feed_dict=feed)

        c = pick_top_n(preds, len(vocab))
        samples.append(int_to_vocab[c])

        for i in range(n_samples):
            x[0,0] = c
            feed = {model.inputs: x,
                    model.keep_prob: 1.,
                    model.initial_state: new_state}
            preds, new_state = sess.run([model.preds, model.final_state], 
                                         feed_dict=feed)

            c = pick_top_n(preds, len(vocab))
            samples.append(int_to_vocab[c])
        
    return ''.join(samples)

In [44]:
checkpoint = "checkpoints/anna/i3560_l512_1.122.ckpt"
samp = sample(checkpoint, 2000, lstm_size, len(vocab), prime="Far")
print(samp)


Farlathit that if had so
like it that it were. He could not trouble to his wife, and there was
anything in them of the side of his weaky in the creature at his forteren
to him.

"What is it? I can't bread to those," said Stepan Arkadyevitch. "It's not
my children, and there is an almost this arm, true it mays already,
and tell you what I have say to you, and was not looking at the peasant,
why is, I don't know him out, and she doesn't speak to me immediately, as
you would say the countess and the more frest an angelembre, and time and
things's silent, but I was not in my stand that is in my head. But if he
say, and was so feeling with his soul. A child--in his soul of his
soul of his soul. He should not see that any of that sense of. Here he
had not been so composed and to speak for as in a whole picture, but
all the setting and her excellent and society, who had been delighted
and see to anywing had been being troed to thousand words on them,
we liked him.

That set in her money at the table, he came into the party. The capable
of his she could not be as an old composure.

"That's all something there will be down becime by throe is
such a silent, as in a countess, I should state it out and divorct.
The discussion is not for me. I was that something was simply they are
all three manshess of a sensitions of mind it all."

"No," he thought, shouted and lifting his soul. "While it might see your
honser and she, I could burst. And I had been a midelity. And I had a
marnief are through the countess," he said, looking at him, a chosing
which they had been carried out and still solied, and there was a sen that
was to be completely, and that this matter of all the seconds of it, and
a concipation were to her husband, who came up and conscaously, that he
was not the station. All his fourse she was always at the country,,
to speak oft, and though they were to hear the delightful throom and
whether they came towards the morning, and his living and a coller and
hold--the children. 

In [43]:
checkpoint = "checkpoints/anna/i200_l512_2.432.ckpt"
samp = sample(checkpoint, 1000, lstm_size, len(vocab), prime="Far")
print(samp)


Farnt him oste wha sorind thans tout thint asd an sesand an hires on thime sind thit aled, ban thand and out hore as the ter hos ton ho te that, was tis tart al the hand sostint him sore an tit an son thes, win he se ther san ther hher tas tarereng,.

Anl at an ades in ond hesiln, ad hhe torers teans, wast tar arering tho this sos alten sorer has hhas an siton ther him he had sin he ard ate te anling the sosin her ans and
arins asd and ther ale te tot an tand tanginge wath and ho ald, so sot th asend sat hare sother horesinnd, he hesense wing ante her so tith tir sherinn, anded and to the toul anderin he sorit he torsith she se atere an ting ot hand and thit hhe so the te wile har
ens ont in the sersise, and we he seres tar aterer, to ato tat or has he he wan ton here won and sen heren he sosering, to to theer oo adent har herere the wosh oute, was serild ward tous hed astend..

I's sint on alt in har tor tit her asd hade shithans ored he talereng an soredendere tim tot hees. Tise sor and 

In [46]:
checkpoint = "checkpoints/anna/i600_l512_1.750.ckpt"
samp = sample(checkpoint, 1000, lstm_size, len(vocab), prime="Far")
print(samp)


Fard as astice her said he celatice of to seress in the raice, and to be the some and sere allats to that said to that the sark and a cast a the wither ald the pacinesse of her had astition, he said to the sount as she west at hissele. Af the cond it he was a fact onthis astisarianing.


"Or a ton to to be that's a more at aspestale as the sont of anstiring as
thours and trey.

The same wo dangring the
raterst, who sore and somethy had ast out an of his book. "We had's beane were that, and a morted a thay he had to tere. Then to
her homent andertersed his his ancouted to the pirsted, the soution for of the pirsice inthirgest and stenciol, with the hard and and
a colrice of to be oneres,
the song to this anderssad.
The could ounterss the said to serom of
soment a carsed of sheres of she
torded
har and want in their of hould, but
her told in that in he tad a the same to her. Serghing an her has and with the seed, and the camt ont his about of the
sail, the her then all houg ant or to hus to 

In [47]:
checkpoint = "checkpoints/anna/i1000_l512_1.484.ckpt"
samp = sample(checkpoint, 1000, lstm_size, len(vocab), prime="Far")
print(samp)


Farrat, his felt has at it.

"When the pose ther hor exceed
to his sheant was," weat a sime of his sounsed. The coment and the facily that which had began terede a marilicaly whice whether the pose of his hand, at she was alligated herself the same on she had to
taiking to his forthing and streath how to hand
began in a lang at some at it, this he cholded not set all her. "Wo love that is setthing. Him anstering as seen that."

"Yes in the man that say the mare a crances is it?" said Sergazy Ivancatching. "You doon think were somether is ifficult of a mone of
though the most at the countes that the
mean on the come to say the most, to
his feesing of
a man she, whilo he
sained and well, that he would still at to said. He wind at his for the sore in the most
of hoss and almoved to see him. They have betine the sumper into at he his stire, and what he was that at the so steate of the
sound, and shin should have a geest of shall feet on the conderation to she had been at that imporsing the dre