def build(self, input_shape):
input_shape = input_shape.as_list()
input_hidden_size = input_shape[1]
kernel_shape = [input_hidden_size, self.num_outputs]
self.kernel = tf.get_variable("kernel",
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
dtype=self.dtype,
trainable=True)
if not self.log_alpha_initializer:
# default log alpha set s.t. \alpha / (\alpha + 1) = .1
self.log_alpha_initializer = tf.random_normal_initializer(
mean=2.197, stddev=0.01, dtype=self.dtype)
self.log_alpha = tf.get_variable(
"log_alpha",
shape=kernel_shape,
initializer=self.log_alpha_initializer,
dtype=self.dtype,
trainable=True)
layer_utils.add_variable_to_collection(
(self.kernel, self.log_alpha), [THETA_LOGALPHA_COLLECTION], None)
if self.use_bias:
self.bias = self.add_variable(name="bias",
shape=(self.num_outputs, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
self.built = True
def build(self, input_shape):
input_shape = input_shape.as_list()
input_hidden_size = input_shape[1]
kernel_shape = [input_hidden_size, self.num_outputs]
self.kernel = tf.get_variable("kernel",
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
dtype=tf.float32,
trainable=True)
if not self.log_sigma2_initializer:
self.log_sigma2_initializer = tf.constant_initializer(
value=-10, dtype=tf.float32)
self.log_sigma2 = tf.get_variable(
"log_sigma2",
shape=kernel_shape,
initializer=self.log_sigma2_initializer,
dtype=tf.float32,
trainable=True)
layer_utils.add_variable_to_collection(
(self.kernel, self.log_sigma2), [THETA_LOGSIGMA2_COLLECTION], None)
if self.use_bias:
self.bias = self.add_variable(name="bias",
shape=(self.num_outputs, ),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
self.built = True