Prevent TensorFlow from accessing the GPU?

Question

Is there a way to run TensorFlow purely on the CPU. All of the memory on my machine is hogged by a separate process running TensorFlow. I have tried setting the per_process_memory_fraction to 0, unsuccessfully.

Answer 1

Have a look to this question or this answer.

To summarise you can add this piece of code:

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
import tensorflow as tf

Playing with the CUDA_VISIBLE_DEVICES environment variable is one of if not the way to go whenever you have GPU-tensorflow installed and you don't want to use any GPUs.

You to want either export CUDA_VISIBLE_DEVICES= or alternatively use a virtualenv with a non-GPU installation of TensorFlow.

Answer 2

You can use only CPUs by openning a session with a GPU limit of 0:

sess = tf.Session(config=tf.ConfigProto(device_count={'GPU': 0}))

See https://www.tensorflow.org/api_docs/python/tf/ConfigProto for more details.

A proof that it works for @Nicolas:

In Python, write:

import tensorflow as tf
sess_cpu = tf.Session(config=tf.ConfigProto(device_count={'GPU': 0}))

Then in a terminal:

nvidia-smi

You will see something like:

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0     24869    C   /.../python                 99MiB                     |
+-----------------------------------------------------------------------------+

Then repeat the process: In Python, write:

import tensorflow as tf
sess_gpu = tf.Session()

Then in a terminal:

nvidia-smi

You will see something like:

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|    0     25900    C   /.../python                                   5775MiB |
+-----------------------------------------------------------------------------+