def __init__(self, config):
self.embed_dim = config.encoder_dim
self.config = config
self.input_noise = tf.placeholder(shape=[None, self.embed_dim],
dtype=tf.float32)
self.input_sequence = tf.placeholder(shape=[None, MAXLEN],
dtype=tf.int64)
self.uniform_initializer = tf.random_uniform_initializer(
minval=-self.config.uniform_init_scale,
maxval=self.config.uniform_init_scale)
self.cell = gru_cell.LayerNormGRUCell(
self.embed_dim,
w_initializer=tf.random_uniform_initializer(minval=-0.1,
maxval=0.1),
u_initializer=random_orthonormal_initializer,
b_initializer=tf.random_uniform_initializer(minval=-0.1,
maxval=0.1))
self.word_embedding = tf.get_variable(
'target_W',
shape=[self.config.vocab_size, self.config.word_embedding_dim],
initializer=self.uniform_initializer)
self.output_W = tf.get_variable(
'output_W',
shape=[self.embed_dim, self.config.vocab_size],
initializer=self.uniform_initializer)
self.output_b = tf.get_variable('output_b',
shape=[self.config.vocab_size],
initializer=self.uniform_initializer)
self.build_likelihood()
self.build_rollout()
def _initialize_gru_cell(self, num_units):
"""Initializes a GRU cell.
The Variables of the GRU cell are initialized in a way that exactly matches
the skip-thoughts paper: recurrent weights are initialized from random
orthonormal matrices and non-recurrent weights are initialized from random
uniform matrices.
Args:
num_units: Number of output units.
Returns:
cell: An instance of RNNCell with variable initializers that match the
skip-thoughts paper.
"""
return gru_cell.LayerNormGRUCell(
num_units,
w_initializer=self.uniform_initializer,
u_initializer=random_orthonormal_initializer,
b_initializer=tf.constant_initializer(0.0))
def _initialize_gru_cell(self, num_units):
return gru_cell.LayerNormGRUCell(
num_units,
w_initializer=self.uniform_initializer,
u_initializer=random_orthonormal_initializer,
b_initializer=tf.constant_initializer(0.0))