def distributed_initializer(self):
bcast_ops = []
for v in self.variables():
bcast_ops.append(tf.assign(v, broadcast(v)))
with tf.control_dependencies(bcast_ops):
with tf.control_dependencies([self._save_model_op]):
return barrier()
def _distributed_initializer(self):
bcast_ops = [tf.assign(v, broadcast(v)) for v in tf.global_variables()]
# FIXME: tf.trainable_variables() will return a TFOptimizer/iterations:0 if with Keras
# I think we need to find a better way to identify trainable variables?
# TODO: Can we decouple distribuetd_initilizer and the tensor store init?
variables = tf.trainable_variables()
variables = [
v for v in variables if 'TFOptimizer/iterations' not in v.name
]
with tf.control_dependencies(bcast_ops):
with tf.control_dependencies([self._build_save_op(variables)]):
return barrier()
def _distributed_initializer(self):
ops = [tf.assign(v, broadcast(v)) for v in tf.global_variables()]
return tf.group(ops)