In this project, we'll generate our own Simpsons TV scripts using RNNs. We'll be using part of the Simpsons dataset of scripts from 27 seasons. The Neural Network you'll build will generate a new TV script for a scene at Moe's Tavern.
We'll be using a subset of the original dataset. It consists of only the scenes in Moe's Tavern. This doesn't include other versions of the tavern, like "Moe's Cavern", "Flaming Moe's", "Uncle Moe's Family Feed-Bag", etc..
In [1]:
import helper
data_dir = './data/simpsons/moes_tavern_lines.txt'
text = helper.load_data(data_dir)
# Ignore notice, since we don't use it for analysing the data
text = text[81:]
In [3]:
view_sentence_range = (20, 30)
import numpy as np
print('Dataset Stats')
print('Roughly the number of unique words: {}'.format(len({word: None for word in text.split()})))
scenes = text.split('\n\n')
print('Number of scenes: {}'.format(len(scenes)))
sentence_count_scene = [scene.count('\n') for scene in scenes]
print('Average number of sentences in each scene: {}'.format(np.average(sentence_count_scene)))
sentences = [sentence for scene in scenes for sentence in scene.split('\n')]
print('Number of lines: {}'.format(len(sentences)))
word_count_sentence = [len(sentence.split()) for sentence in sentences]
print('Average number of words in each line: {}'.format(np.average(word_count_sentence)))
print()
print('The sentences {} to {}:'.format(*view_sentence_range))
print('\n'.join(text.split('\n')[view_sentence_range[0]:view_sentence_range[1]]))
The first thing to do to any dataset is preprocessing. Implement the following preprocessing functions below:
To create a word embedding, you first need to transform the words to ids. In this function, create two dictionaries:
vocab_to_intint_to_vocabReturn these dictionaries in the following tuple (vocab_to_int, int_to_vocab)
In [4]:
import numpy as np
import problem_unittests as tests
from collections import Counter
def create_lookup_tables(text):
"""
Create lookup tables for vocabulary
:param text: The text of tv scripts split into words
:return: A tuple of dicts (vocab_to_int, int_to_vocab)
"""
# Implement Function
word_counts = Counter(text)
sorted_vocab = sorted(word_counts, key=word_counts.get, reverse=True)
int_to_vocab = {index: text for index, text in enumerate(sorted_vocab)}
vocab_to_int = {text: index for index, text in int_to_vocab.items()}
return vocab_to_int, int_to_vocab
tests.test_create_lookup_tables(create_lookup_tables)
We'll be splitting the script into a word array using spaces as delimiters. However, punctuations like periods and exclamation marks make it hard for the neural network to distinguish between the word "bye" and "bye!".
Implement the function token_lookup to return a dict that will be used to tokenize symbols like "!" into "||Exclamation_Mark||". Create a dictionary for the following symbols where the symbol is the key and value is the token:
This dictionary will be used to token the symbols and add the delimiter (space) around it. This separates the symbols as it's own word, making it easier for the neural network to predict on the next word. Make sure you don't use a token that could be confused as a word. Instead of using the token "dash", try using something like "||dash||".
In [5]:
def token_lookup():
"""
Generate a dict to turn punctuation into a token.
:return: Tokenize dictionary where the key is the punctuation and the value is the token
"""
# Implement Function
token_dict = {
'.': '<<period>>',
',': '<<comma>>',
'"': '<<quotation_mark>>',
';': '<<semicolon>>',
'!': '<<exclamation_mark>>',
'?': '<<question_mark>>',
'(': '<<left_parentheses>>',
')': '<<right_parentheses>>',
'--': '<<dash>>',
'\n': '<<return>>',
}
return token_dict
tests.test_tokenize(token_lookup)
In [6]:
# Preprocess Training, Validation, and Testing Data
helper.preprocess_and_save_data(data_dir, token_lookup, create_lookup_tables)
In [7]:
import helper
import numpy as np
import problem_unittests as tests
int_text, vocab_to_int, int_to_vocab, token_dict = helper.load_preprocess()
In [8]:
from distutils.version import LooseVersion
import warnings
import tensorflow as tf
# Check TensorFlow Version
assert LooseVersion(tf.__version__) >= LooseVersion('1.0'), 'Please use TensorFlow version 1.0 or newer'
print('TensorFlow Version: {}'.format(tf.__version__))
# Check for a GPU
if not tf.test.gpu_device_name():
warnings.warn('No GPU found. Please use a GPU to train your neural network.')
else:
print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))
In [9]:
def get_inputs():
"""
Create TF Placeholders for input, targets, and learning rate.
:return: Tuple (input, targets, learning rate)
"""
# Implement Function
inputs = tf.placeholder(tf.int32, [None, None], name='input')
targets = tf.placeholder(tf.int32, [None, None], name='targets')
learning_rate = tf.placeholder(tf.float32, name='learning_rate')
return inputs, targets, learning_rate
tests.test_get_inputs(get_inputs)
Stack one or more BasicLSTMCells in a MultiRNNCell.
Return the cell and initial state in the following tuple (Cell, InitialState)
In [22]:
def get_init_cell(batch_size, rnn_size):
"""
Create an RNN Cell and initialize it.
:param batch_size: Size of batches
:param rnn_size: Size of RNNs
:return: Tuple (cell, initialize state)
"""
# Implement Function
lstm_layers = 2
keep_prob = 0.5
# Use a basic LSTM cell
lstm = tf.contrib.rnn.BasicLSTMCell(rnn_size)
# Add dropout to the cell
drop = tf.contrib.rnn.DropoutWrapper(lstm, output_keep_prob=keep_prob)
# Stack up multiple LSTM layers, for deep learning
cell = tf.contrib.rnn.MultiRNNCell([drop] * lstm_layers)
initial_state = tf.identity(cell.zero_state(batch_size, tf.float32), name="initial_state")
return cell, initial_state
tests.test_get_init_cell(get_init_cell)
In [23]:
def get_embed(input_data, vocab_size, embed_dim):
"""
Create embedding for <input_data>.
:param input_data: TF placeholder for text input.
:param vocab_size: Number of words in vocabulary.
:param embed_dim: Number of embedding dimensions
:return: Embedded input.
"""
# Implement Function
#with graph.as_default():
embedding = tf.Variable(tf.random_uniform((vocab_size, embed_dim), -1, 1))
embed = tf.nn.embedding_lookup(embedding, input_data)
return embed
tests.test_get_embed(get_embed)
In [24]:
def build_rnn(cell, inputs):
"""
Create a RNN using a RNN Cell
:param cell: RNN Cell
:param inputs: Input text data
:return: Tuple (Outputs, Final State)
"""
# Implement Function
#embed_size = 100
#vocab_size = 100
#embed = get_embed(inputs, vocab_size, embed_size)
#outputs, final_state = tf.nn.dynamic_rnn(cell, embed, initial_state=initial_state)
outputs, state = tf.nn.dynamic_rnn(cell, inputs, dtype=tf.float32)
final_state = tf.identity(state, name = 'final_state')
return outputs, final_state
tests.test_build_rnn(build_rnn)
Apply the functions you implemented above to:
input_data using your get_embed(input_data, vocab_size, embed_dim) function.cell and your build_rnn(cell, inputs) function.vocab_size as the number of outputs.Return the logits and final state in the following tuple (Logits, FinalState)
In [25]:
def build_nn(cell, rnn_size, input_data, vocab_size):
"""
Build part of the neural network
:param cell: RNN cell
:param rnn_size: Size of rnns
:param input_data: Input data
:param vocab_size: Vocabulary size
:return: Tuple (Logits, FinalState)
"""
# Implement Function
build_rnn_input = get_embed(input_data, vocab_size, rnn_size)
outputs, final_state = build_rnn(cell, build_rnn_input)
logits = tf.contrib.layers.fully_connected(outputs, vocab_size, activation_fn=None)
return logits, final_state
tests.test_build_nn(build_nn)
Implement get_batches to create batches of input and targets using int_text. The batches should be a Numpy array with the shape (number of batches, 2, batch size, sequence length). Each batch contains two elements:
[batch size, sequence length][batch size, sequence length]
In [26]:
def get_batches(int_text, batch_size, seq_length):
"""
Return batches of input and target
:param int_text: Text with the words replaced by their ids
:param batch_size: The size of batch
:param seq_length: The length of sequence
:return: Batches as a Numpy array
"""
# Implement Function
n_batches = int(len(int_text) / (batch_size * seq_length))
# Drop the last few characters to make only full batches
xdata = np.array(int_text[: n_batches * batch_size * seq_length])
ydata = np.array(int_text[1: n_batches * batch_size * seq_length + 1])
x_batches = np.split(xdata.reshape(batch_size, -1), n_batches, 1)
y_batches = np.split(ydata.reshape(batch_size, -1), n_batches, 1)
return np.asarray(list(zip(x_batches, y_batches)))
tests.test_get_batches(get_batches)
Tune the following parameters:
num_epochs to the number of epochs.batch_size to the batch size.rnn_size to the size of the RNNs.seq_length to the length of sequence.learning_rate to the learning rate.show_every_n_batches to the number of batches the neural network should print progress.
In [30]:
# Number of Epochs
num_epochs = 200
# Batch Size
batch_size = 64
# RNN Size
rnn_size = 512
# Sequence Length
seq_length = 32
# Learning Rate
learning_rate = 0.001
# Show stats for every n number of batches
show_every_n_batches = 15
save_dir = './save'
In [31]:
from tensorflow.contrib import seq2seq
train_graph = tf.Graph()
with train_graph.as_default():
vocab_size = len(int_to_vocab)
input_text, targets, lr = get_inputs()
input_data_shape = tf.shape(input_text)
cell, initial_state = get_init_cell(input_data_shape[0], rnn_size)
logits, final_state = build_nn(cell, rnn_size, input_text, vocab_size)
# Probabilities for generating words
probs = tf.nn.softmax(logits, name='probs')
# Loss function
cost = seq2seq.sequence_loss(
logits,
targets,
tf.ones([input_data_shape[0], input_data_shape[1]]))
# Optimizer
optimizer = tf.train.AdamOptimizer(lr)
# Gradient Clipping
gradients = optimizer.compute_gradients(cost)
capped_gradients = [(tf.clip_by_value(grad, -1., 1.), var) for grad, var in gradients]
train_op = optimizer.apply_gradients(capped_gradients)
In [32]:
batches = get_batches(int_text, batch_size, seq_length)
with tf.Session(graph=train_graph) as sess:
sess.run(tf.global_variables_initializer())
for epoch_i in range(num_epochs):
state = sess.run(initial_state, {input_text: batches[0][0]})
for batch_i, (x, y) in enumerate(batches):
feed = {
input_text: x,
targets: y,
initial_state: state,
lr: learning_rate}
train_loss, state, _ = sess.run([cost, final_state, train_op], feed)
# Show every <show_every_n_batches> batches
if (epoch_i * len(batches) + batch_i) % show_every_n_batches == 0:
print('Epoch {:>3} Batch {:>4}/{} train_loss = {:.3f}'.format(
epoch_i,
batch_i,
len(batches),
train_loss))
# Save Model
saver = tf.train.Saver()
saver.save(sess, save_dir)
print('Model Trained and Saved')
In [33]:
# Save parameters for checkpoint
helper.save_params((seq_length, save_dir))
In [34]:
import tensorflow as tf
import numpy as np
import helper
import problem_unittests as tests
_, vocab_to_int, int_to_vocab, token_dict = helper.load_preprocess()
seq_length, load_dir = helper.load_params()
Get tensors from loaded_graph using the function get_tensor_by_name().
Return the tensors in the following tuple (InputTensor, InitialStateTensor, FinalStateTensor, ProbsTensor)
In [35]:
def get_tensors(loaded_graph):
"""
Get input, initial state, final state, and probabilities tensor from <loaded_graph>
:param loaded_graph: TensorFlow graph loaded from file
:return: Tuple (InputTensor, InitialStateTensor, FinalStateTensor, ProbsTensor)
"""
# Implement Function
InputTensor = loaded_graph.get_tensor_by_name('input:0')
InitialStateTensor = loaded_graph.get_tensor_by_name('initial_state:0')
FinalStateTensor = loaded_graph.get_tensor_by_name('final_state:0')
ProbsTensor = loaded_graph.get_tensor_by_name('probs:0')
return InputTensor, InitialStateTensor, FinalStateTensor, ProbsTensor
tests.test_get_tensors(get_tensors)
In [36]:
def pick_word(probabilities, int_to_vocab):
"""
Pick the next word in the generated text
:param probabilities: Probabilites of the next word
:param int_to_vocab: Dictionary of word ids as the keys and words as the values
:return: String of the predicted word
"""
# Implement Function
#int_words = [int_to_vocab[word] for word in text]
#train_words = [word for word in int_to_vocab if np.any(probabilities < random.random())]
index = np.argmax(probabilities)
train_words = int_to_vocab[index]
return train_words
tests.test_pick_word(pick_word)
In [37]:
gen_length = 200
# homer_simpson, moe_szyslak, or Barney_Gumble
prime_word = 'moe_szyslak'
loaded_graph = tf.Graph()
with tf.Session(graph=loaded_graph) as sess:
# Load saved model
loader = tf.train.import_meta_graph(load_dir + '.meta')
loader.restore(sess, load_dir)
# Get Tensors from loaded model
input_text, initial_state, final_state, probs = get_tensors(loaded_graph)
# Sentences generation setup
gen_sentences = [prime_word + ':']
prev_state = sess.run(initial_state, {input_text: np.array([[1]])})
# Generate sentences
for n in range(gen_length):
# Dynamic Input
dyn_input = [[vocab_to_int[word] for word in gen_sentences[-seq_length:]]]
dyn_seq_length = len(dyn_input[0])
# Get Prediction
probabilities, prev_state = sess.run(
[probs, final_state],
{input_text: dyn_input, initial_state: prev_state})
pred_word = pick_word(probabilities[dyn_seq_length-1], int_to_vocab)
gen_sentences.append(pred_word)
# Remove tokens
tv_script = ' '.join(gen_sentences)
for key, token in token_dict.items():
ending = ' ' if key in ['\n', '(', '"'] else ''
tv_script = tv_script.replace(' ' + token.lower(), key)
tv_script = tv_script.replace('\n ', '\n')
tv_script = tv_script.replace('( ', '(')
print(tv_script)
It's ok if the TV script doesn't make any sense. We trained on less than a megabyte of text. In order to get good results, we'll have to use a smaller vocabulary or get more data. This was a subset of another dataset. We can use it all!