How to explicitly run Tensor Flow Estimator on GPU

Question

I've found the information that so as to use Estimator Model on GPU I need the following code:

# Create a tf.estimator.RunConfig to ensure the model is run on CPU, which
# trains faster than GPU for this model.
run_config = tf.estimator.RunConfig().replace(
      session_config=tf.ConfigProto(device_count={'GPU': 0},
                                    inter_op_parallelism_threads=inter_op,
intra_op_parallelism_threads=intra_op))

Source of code: https://github.com/tensorflow/models/blob/master/official/wide_deep/census_main.py

Is this correct? Because I got the error:

NameError: name 'inter_op' is not defined

UPD: What is inter_op supposed to be like? How to choose its value?

You need to define inter_op. It should be the number of parallel threads you want. — Ruzihm, Sep 21 '18 at 16:37
Possible duplicate of [Meaning of inter\_op\_parallelism\_threads and intra\_op\_parallelism\_threads](https://stackoverflow.com/questions/41233635/meaning-of-inter-op-parallelism-threads-and-intra-op-parallelism-threads) — Ruzihm, Sep 21 '18 at 16:43
My question is not exact dupliate because the main problem is to run on GPU rather than what is the meaning of threads variables — James Flash, Sep 21 '18 at 17:19
if all you want to do is run on the GPU, then `config = tf.ConfigProto()` and then `sess = tf.Session( config = config)` will run that session on your GPU by default. You don't need to set any ConfigProto parameters manually. — Ruzihm, Sep 21 '18 at 17:29
I can't use tf.Session because I employ tf.Estimator. Nevertheless, thank you very much for help — James Flash, Sep 21 '18 at 19:21

Ruzihm · Accepted Answer · 2018-09-22T01:50:28.737

3

You need to define the variables before you use them, as well as set the GPU count to a non-zero number.

inter_op = 10
intra_op = 10
session_config=tf.ConfigProto(device_count={'GPU': 1},
                              inter_op_parallelism_threads=inter_op,
                              intra_op_parallelism_threads=intra_op))

edited Sep 22 '18 at 01:50

answered Sep 21 '18 at 16:40

Ruzihm

19,749
5
36
48

Why are they equal to 10s? – James Flash Sep 21 '18 at 16:42
Task manager shows that my GPU is used only at 2 %. It seems my method doesn't work – James Flash Sep 21 '18 at 17:14
@JamesFlash I edited my answer to select how many GPUs you want. – Ruzihm Sep 22 '18 at 01:50

How to explicitly run Tensor Flow Estimator on GPU

1 Answers1