Apply to speak at TensorFlow World. Deadline April 23rd. Propose talk


Class ElasticAverageCustomGetter

Defined in tensorflow/contrib/opt/python/training/

Custom_getter class is used to do: 1. Change trainable variables to local collection and place them at worker device 2. Generate global variables(global center variables) 3. Generate local variables(local center variables) which record the global variables and place them at worker device Notice that the class should be used with tf.replica_device_setter, so that the global center variables and global step variable can be placed at ps device. Besides, use 'tf.get_variable' instead of 'tf.Variable' to use this custom getter.

For example, ea_custom_getter = ElasticAverageCustomGetter(worker_device) with tf.device( tf.train.replica_device_setter( worker_device=worker_device, ps_device="/job:ps", cluster=cluster)), tf.variable_scope('',custom_getter=ea_custom_getter): ... create your model here ... with tf.device(worker_device): opt = tf.train.MomentumOptimizer(...) optimizer = ElasticAverageOptimizer( opt, num_worker=2, moving_rate=0.01, # or use default value communication_period=20, ea_custom_getter=ea_custom_getter) ... train_op = optimizer.apply_gradients( grads_vars, global_step=global_step) ... hooks = [optimizer.make_session_run_hook(is_chief, task_index)] ... with tf.train.MonitoredTrainingSession(, is_chief=is_chief, checkpoint_dir=("...), save_checkpoint_secs=600, hooks=hooks) as mon_sess:



Create a new ElasticAverageCustomGetter.


  • worker_device: String. Name of the worker job.




Call self as a function.