|
import tensorflow as tf
def layer_norm_1(component, net):
ops = component.ops
scope = ops.generate_name()
with tf.variable_scope(scope, reuse=ops._reuse):
net = tf.contrib.layers.layer_norm(net, scope=scope, center=True, scale=True, variables_collections=tf.GraphKeys.LOCAL_VARIABLES)
vars = lookup_vars(scope)
if not ops._reuse:
ops.add_weights(vars)
return net
def batch_norm_1(component, net):
config = component.config
ops = component.ops
dtype = ops.dtype
shape = ops.shape(net)
epsilon = config.epsilon or 0.001
batch_norm_gamma_stddev = config.batch_norm_gamma_stddev or 0.02
decay = config.batch_norm_decay or 0.999
center = config.batch_norm_center or True
scale = config.batch_norm_scale or False
epsilon = config.batch_norm_epsilon or 0.001
scope = ops.generate_name()
with tf.variable_scope(scope, reuse=ops._reuse):
net = tf.contrib.layers.batch_norm(net,
decay = decay,
center = center,
scale = scale,
epsilon = epsilon,
is_training = True,
scope=scope
)
vars = lookup_vars(scope)
if not ops._reuse:
ops.add_weights(vars)
return net
def lookup_vars(name):
vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
filtered = []
for var in vars:
if var.name.startswith(name):
filtered.append(var)
return filtered
|
