Fail to run word embedding example in tensorflow tutorial with GPUs

Question

I am trying to run the word embedding example code at https://github.com/tensorflow/tensorflow/tree/master/tensorflow/g3doc/tutorials/word2vec (installed with GPU version of tensorflow under Ubuntu 14.04), but it returns the following error message:

Found and verified text8.zip
Data size 17005207
Most common words (+UNK) [['UNK', 418391], ('the', 1061396), ('of', 593677), ('and', 416629), ('one', 411764)]
Sample data [5239, 3084, 12, 6, 195, 2, 3137, 46, 59, 156]
3084 -> 12
originated -> as
3084 -> 5239
originated -> anarchism
12 -> 3084
as -> originated
12 -> 6
as -> a
6 -> 12
a -> as
6 -> 195
a -> term
195 -> 6
term -> a
195 -> 2
term -> of
I tensorflow/core/common_runtime/local_device.cc:25] Local device intra op parallelism threads: 12
I tensorflow/core/common_runtime/gpu/gpu_init.cc:88] Found device 0 with properties: 
name: GeForce GTX TITAN X
major: 5 minor: 2 memoryClockRate (GHz) 1.076
pciBusID 0000:03:00.0
Total memory: 12.00GiB
Free memory: 443.32MiB
I tensorflow/core/common_runtime/gpu/gpu_init.cc:88] Found device 1 with properties: 
name: GeForce GTX TITAN X
major: 5 minor: 2 memoryClockRate (GHz) 1.076
pciBusID 0000:05:00.0
Total memory: 12.00GiB
Free memory: 451.61MiB
I tensorflow/core/common_runtime/gpu/gpu_init.cc:112] DMA: 0 1 
I tensorflow/core/common_runtime/gpu/gpu_init.cc:122] 0:   Y Y 
I tensorflow/core/common_runtime/gpu/gpu_init.cc:122] 1:   Y Y 
I tensorflow/core/common_runtime/gpu/gpu_device.cc:643] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX TITAN X, pci bus id: 0000:03:00.0)
I tensorflow/core/common_runtime/gpu/gpu_device.cc:643] Creating TensorFlow device (/gpu:1) -> (device: 1, name: GeForce GTX TITAN X, pci bus id: 0000:05:00.0)
I tensorflow/core/common_runtime/gpu/gpu_region_allocator.cc:47] Setting region size to 254881792
I tensorflow/core/common_runtime/gpu/gpu_region_allocator.cc:47] Setting region size to 263835648
I tensorflow/core/common_runtime/local_session.cc:45] Local session inter op parallelism threads: 12
Initialized
Traceback (most recent call last):
  File "word2vec_basic.py", line 171, in <module>
    _, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 345, in run
    results = self._do_run(target_list, unique_fetch_targets, feed_dict_string)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 419, in _do_run
    e.code)
tensorflow.python.framework.errors.InvalidArgumentError: Cannot assign a device to node 'GradientDescent/update_Variable_2/ScatterSub': Could not satisfy explicit device specification '' because the node was colocated with a group of nodes that required incompatible device '/job:localhost/replica:0/task:0/GPU:0'
     [[Node: GradientDescent/update_Variable_2/ScatterSub = ScatterSub[T=DT_FLOAT, Tindices=DT_INT64, use_locking=false](Variable_2, gradients/concat_1, GradientDescent/update_Variable_2/mul)]]
Caused by op u'GradientDescent/update_Variable_2/ScatterSub', defined at:
  File "word2vec_basic.py", line 145, in <module>
    optimizer = tf.train.GradientDescentOptimizer(1.0).minimize(loss)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 167, in minimize
    name=name)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/training/optimizer.py", line 256, in apply_gradients
    update_ops.append(self._apply_sparse(grad, var))
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/training/gradient_descent.py", line 40, in _apply_sparse
    return var.scatter_sub(delta, use_locking=self._use_locking)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/ops/variables.py", line 324, in scatter_sub
    use_locking=use_locking)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/ops/gen_state_ops.py", line 227, in scatter_sub
    name=name)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/ops/op_def_library.py", line 633, in apply_op
    op_def=op_def)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1710, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/home/chentingpc/anaconda/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 988, in __init__
    self._traceback = _extract_stack()

When I run the code in CPU version tensorflow, it works just fine. But not for GPU version. I also tried to use tf.device('/cpu:0') to force it using CUP instead of GPU, but it produces the same output.

Is there any function in this example cannot be run in GPUs? And how do I switch to CPU without reinstalling CPU version of tensorflow since tf.device('/cpu:0') not working?

Answer 1

It seems a whole bunch of operations used in this example aren't supported on a GPU. A quick workaround is to restrict operations such that only matrix muls are ran on the GPU.

There's an example in the docs: http://tensorflow.org/api_docs/python/framework.md

See the section on tf.Graph.device(device_name_or_function)

I was able to get it working with the following:

def device_for_node(n):
  if n.type == "MatMul":
    return "/gpu:0"
  else:
    return "/cpu:0"

with graph.as_default():
  with graph.device(device_for_node):
    ...