def call(self, inputs):
"""Runs the model forward to generate a sequence of productions.
Args:
inputs: Unused.
Returns:
productions: Tensor of shape [1, num_productions, num_production_rules].
Slices along the `num_productions` dimension represent one-hot vectors.
"""
del inputs # unused
latent_code = ed.MultivariateNormalDiag(loc=tf.zeros(self.latent_size),
sample_shape=1,
name="latent_code")
state = self.lstm.zero_state(1, dtype=tf.float32)
t = 0
productions = []
stack = [self.grammar.start_symbol]
while stack:
symbol = stack.pop()
net, state = self.lstm(latent_code, state)
logits = (self.output_layer(net) +
self.grammar.mask(symbol, on_value=0., off_value=-1e9))
production = ed.OneHotCategorical(logits=logits,
name="production_" + str(t))
_, rhs = self.grammar.production_rules[tf.argmax(production,
axis=-1)]
for symbol in rhs:
if symbol in self.grammar.nonterminal_symbols:
stack.append(symbol)
productions.append(production)
t += 1
return tf.stack(productions, axis=1)
def latent_dirichlet_allocation(concentration, topics_words):
"""Latent Dirichlet Allocation in terms of its generative process.
The model posits a distribution over bags of words and is parameterized by
a concentration and the topic-word probabilities. It collapses per-word
topic assignments.
Args:
concentration: A Tensor of shape [1, num_topics], which parameterizes the
Dirichlet prior over topics.
topics_words: A Tensor of shape [num_topics, num_words], where each row
(topic) denotes the probability of each word being in that topic.
Returns:
bag_of_words: A random variable capturing a sample from the model, of shape
[1, num_words]. It represents one generated document as a bag of words.
"""
topics = ed.Dirichlet(concentration=concentration, name="topics")
word_probs = tf.matmul(topics, topics_words)
# The observations are bags of words and therefore not one-hot. However,
# log_prob of OneHotCategorical computes the probability correctly in
# this case.
bag_of_words = ed.OneHotCategorical(probs=word_probs, name="bag_of_words")
return bag_of_words