What is the format for device filters in TensorFlow?

Question

So the Session config proto has a device_filters option, with the comment:

// When any filters are present sessions will ignore all devices which do not
  // match the filters. Each filter can be partially specified, e.g. "/job:ps"
  // "/job:worker/replica:3", etc.

Does anyone have concrete explanation of the format? For example, I want to exclude /gpu:0 as an option because I use it for running other models.

I've tried

config = tf.ConfigProto()
config.device_filters.append('/gpu:1')
config.device_filters.append('/cpu:0')
with tf.Session(config=config):
    # Do stuff

But I'm still getting ops allocated to gpu 0. I don't want to override the devices on a per-op basis.

Answer 1

The ConfigProto.device_filters field is currently ignored by TensorFlow, although it is intended to support your use case in future. If you want to achieve the same end of running ops on /gpu:1 and /cpu:0, you can do that as follows, using "soft placement":

with tf.device("/gpu:1"):
  # Build your model in this with context. All nodes will get the
  # device "/gpu:1".

with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)):
  # Execute your mode in this with context.
  # Soft placement will use /gpu:1 for GPU-compatible ops, and /cpu:0
  # for CPU-only ops.