def _weight_variable(scope_name,
name,
shape,
from_pretrain=False,
stddev=0.01):
# with tf.device('/gpu:3'):
if from_pretrain:
weights = get_pretrained_weights(scope_name, name, shape)
if weights is None:
if FLAGS.xavier_init:
return tf.get_variable(
name,
shape,
DTYPE,
initializer=tf.contrib.layers.xavier_initializer())
else:
return tf.get_variable(
name, shape, DTYPE,
tf.truncated_normal_initializer(stddev=stddev))
else:
init = tf.constant(weights)
return tf.get_variable(name, initializer=init)
else:
if FLAGS.xavier_init:
return tf.get_variable(
name,
shape,
DTYPE,
initializer=tf.contrib.layers.xavier_initializer())
else:
return tf.get_variable(
name, shape, DTYPE,
tf.truncated_normal_initializer(stddev=stddev))
def _bias_variable( scope_name, name, shape, from_pretrain=False, constant_value=0.01):
if from_pretrain:
bias = get_pretrained_weights(scope_name, name,shape)
if bias is None:
return tf.get_variable(name, shape, DTYPE, tf.constant_initializer(constant_value))
else:
init = tf.constant(bias)
return tf.get_variable(name, initializer=init)
else:
bias = tf.get_variable(name, shape, DTYPE, tf.constant_initializer(constant_value))
return bias