Specify GPU Card(s) to run on about prettytensor

[eiderman]

Pretty Tensor is completely compatible with the device scoping mechanism in TF.

If you wrap your model with:

with tf.device('/gpu:1'):
  build_my_model()

then it will assign the appropriate device. When you explicitly specify a gpu device, no attempt is made to put cpu-only ops on a cpu, so for those ops (such as lookup_embedding) you would either need to be in a nested device context or built before the gpu device specification.

Intelligently guessing the correct device is outside the scope of PT, but support for explicit assignments is definitely in scope.

[mschonwe]

Could you provide an example, in the shakespeare.py, where would the with tf.device('/gpu:1') go?

Presumably the change you made (Issue #1) for embedding lookup would provide the necessary assignment for that Op?
with tf.device('/cpu:0'):
embedded = text_in.embedding_lookup(CHARS, [EMBEDDING_SIZE])

[eiderman]

It requires 2 changes and you've alerted me to a few more bugs with device placement.

Before entering the gpu scope, please create the global variable (either on the default device or explicitly on cpu). I added it on line 150:

global_step = pt.global_step()

I surrounded 63-73 in the gpu device context.

Since I don't think that this should be as hard as it is, I've opened a new issue to track this.