def build(self):
""" Construct internal trainable weights.
"""
self.W = []
n_size = self.n_size
for i in range(n_size):
if i < np.floor(n_size / 2):
W_effective = tf.Variable(tf.zeros(
(n_size, self.n_input_feat, self.n_output_feat)),
trainable=False)
else:
W_trainable = self.init(
[n_size - i, self.n_input_feat, self.n_output_feat])
W_before = tf.Variable(tf.zeros(
(i - int(np.floor(n_size / 2.)), self.n_input_feat,
self.n_output_feat)),
trainable=False)
W_after = tf.Variable(tf.zeros(
(int(np.floor(n_size / 2.)), self.n_input_feat,
self.n_output_feat)),
trainable=False)
W_effective = tf.concat([W_before, W_trainable, W_after], 0)
self.W.append(W_effective)
self.W = tf.stack(self.W, axis=0)
self.b = model_ops.zeros((self.n_output_feat, ))
self.trainable_weights = [self.W, self.b]
def build(self):
self.W_effective = self.init(
[self.n_size, self.n_input_feat, self.n_output_feat])
self.W = []
for i in range(self.n_size):
self.W.append(self.W_effective[i:i + 1, :, :])
if i < self.n_size - 1:
self.W.append(
tf.Variable(tf.zeros((self.rate - 1, self.n_input_feat,
self.n_output_feat)),
trainable=False))
self.W = tf.concat(self.W, 0)
self.b = model_ops.zeros((self.n_output_feat, ))
self.trainable_weights = [self.W, self.b]
def build(self):
self.W = self.init(
[self.n_size, self.n_size, self.n_output_feat, self.n_input_feat])
self.b = model_ops.zeros((self.n_output_feat, ))
self.trainable_weights = [self.W, self.b]