nabu: sync optimizer problem

hi Vincent, I have another problem, I use 0 1 as training GPU and numbatches_to_aggregate=0 in default config standardtrainer.cfg , but I found 3 Start master session in log. Is this behavior right?

2018-08-21 07:20:16.556159: E tensorflow/core/grappler/clusters/utils.cc:127] Not found: TF GPU device with id 0 was not registered          [286/1991]
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gradients_impl.py:100: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gradients_impl.py:100: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
  "Converting sparse IndexedSlices to a dense Tensor of unknown shape. "
2018-08-21 07:20:22.265699: I tensorflow/core/distributed_runtime/master_session.cc:1136] Start master session 6066331584d0d493 with config: gpu_optios { allow_growth: true } allow_soft_placement: true
2018-08-21 07:20:22.711547: I tensorflow/core/distributed_runtime/master_session.cc:1136] Start master session 51400be6a5c9e9bc with config: gpu_optios { allow_growth: true } allow_soft_placement: true
WORKER 0: step 0/15600 loss: 3.890086, learning rate: 0.001000 
         time elapsed: 8.318348 sec
         peak memory usage: 21/22604 MB
WORKER 0: step 1/15600 loss: 3.776481, learning rate: 0.001000 
         time elapsed: 1.858370 sec
         peak memory usage: 688/22604 MB
WORKER 0: step 2/15600 loss: 3.670535, learning rate: 0.001000 
         time elapsed: 1.519805 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 3/15600 loss: 3.695373, learning rate: 0.001000 
         time elapsed: 1.608191 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 4/15600 loss: 3.627747, learning rate: 0.000999 
         time elapsed: 2.627351 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 5/15600 loss: 3.646843, learning rate: 0.000999 
         time elapsed: 1.744121 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 6/15600 loss: 3.629820, learning rate: 0.000999 
         time elapsed: 1.468782 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 7/15600 loss: 3.638589, learning rate: 0.000999 
         time elapsed: 1.179851 sec
         peak memory usage: 893/22604 MB
2018-08-21 07:20:52.899330: I tensorflow/core/distributed_runtime/master_session.cc:1136] Start master session c5b480c57e96eea5 with config: gpu_optios { allow_growth: true } allow_soft_placement: true
WORKER 0: step 8/15600 loss: 3.618140, learning rate: 0.000999 
         time elapsed: 1.060105 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 9/15600 loss: 3.596526, learning rate: 0.000999 
         time elapsed: 2.330892 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 10/15600 loss: 3.601762, learning rate: 0.000999 
         time elapsed: 1.179645 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 11/15600 loss: 3.615230, learning rate: 0.000998 
         time elapsed: 0.954928 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 12/15600 loss: 3.602469, learning rate: 0.000998 
         time elapsed: 1.251472 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 13/15600 loss: 3.601032, learning rate: 0.000998 
         time elapsed: 0.886364 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 14/15600 loss: 3.603386, learning rate: 0.000998 
         time elapsed: 1.565970 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 15/15600 loss: 3.622668, learning rate: 0.000998 
         time elapsed: 0.885060 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 16/15600 loss: 3.613049, learning rate: 0.000998 
         time elapsed: 1.336841 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 17/15600 loss: 3.610620, learning rate: 0.000997 
         time elapsed: 0.818247 sec
         peak memory usage: 893/22604 MB
WORKER 0: step 18/15600 loss: 3.618244, learning rate: 0.000997 
         time elapsed: 0.795765 sec
         peak memory usage: 893/22604 MB
WORKER 1: step 17/15600 loss: 3.590606, learning rate: 0.000997 
         time elapsed: 8.500966 sec
         peak memory usage: 893/22604 MB

In Contrast, When I set numbatches_to_aggregate=2 use sync replicas optimizer, there is a error msg like

Traceback (most recent call last):
  File "nabu/scripts/train.py", line 112, in <module>
    testing=False)
  File "nabu/scripts/train.py", line 90, in train
    tr.train(testing)
  File "/opt/cephfs1/asr/users/fanlu/mfs/nabu/nabu/neuralnetworks/trainers/trainer.py", line 607, in train
    outputs = self._create_graph()
  File "/opt/cephfs1/asr/users/fanlu/mfs/nabu/nabu/neuralnetworks/trainers/trainer.py", line 187, in _create_graph
    cluster=cluster)
  File "/opt/cephfs1/asr/users/fanlu/mfs/nabu/nabu/neuralnetworks/trainers/trainer.py", line 569, in _update
    name='apply_gradients')
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/sync_replicas_optimizer.py", line 238, in apply_gradients
    raise ValueError("Global step is required to check staleness")
ValueError: Global step is required to check staleness

So I add param global_step to apply_gradients_op in func _update

# opperation to apply the gradients
        apply_gradients_op = optimizer.apply_gradients(
            grads_and_vars=grads_and_vars,
            name='apply_gradients', global_step=global_step)

and start training, but there is no training log print anymore. How to set global_step to apply_gradients_op? @vrenkens