How to force tensorflow to forget about the GPU, after it has been found?

Question

I have a tensorflow-based code which I am running on various computers, some with CPUs and some with both CPUs & GPUs.

If a GPU is available on the machine, I would like to give the user the option of using the CPU instead.

The code from this answer works fine:

import os
import tensorflow as tf

os.environ['CUDA_VISIBLE_DEVICES'] = '-1'

if tf.test.gpu_device_name():
    print('GPU found')
else:
    print("No GPU found")

# No GPU found

However, I would like to check if a GPU is available first, and then disable it.

I tried:

import tensorflow as tf

if tf.test.gpu_device_name():
    import os
    os.environ['CUDA_VISIBLE_DEVICES'] = '-1'

if tf.test.gpu_device_name():
    print('GPU found')

# GPU found

But it does not work. Once I use tf.test.gpu_device_name(), it always remembers that the system has a GPU.

I also tried del tf, importlib.reload(tf), to no avail.
The only thing that does work is to quit the interpreter and run the first script above.

How can I make the code "forget" about the GPU once it has been found?

Please don't artificially "tag" your titles with languages and frameworks. That's what actual _tags_ are for. — ChrisGPT was on strike, May 21 '20 at 23:55
If it fits naturally into the title it's fine. It's things like "(python) How do I do foo?" or "Why does this give an AttributeError [python][django]" that's a problem. (The _real_ Python tag is fine. It has already been edited back.) — ChrisGPT was on strike, May 21 '20 at 23:59
tldr; instead of writing a title like "(language) question?" or "language: question?", write it like "question?" and add the language [such as python] as a tag. — user2864740, May 22 '20 at 00:02

score 2 · Accepted Answer · answered May 22 '20 at 00:33

2

I don't understand why you need to let the TensorFlow forget. You have GPU that doesn't mean you have to use GPU.

You can use tf.device to specify the underlying device.

For example:

# Place tensors on the CPU
with tf.device('/CPU:0'):
  a = tf.constant([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]])
  b = tf.constant([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]])

c = tf.matmul(a, b)
print(c)

So even though you have GPU, the program will still use CPU.

answered May 22 '20 at 00:33

Sraw

18,892
11
54
87

Is there a way to set the CPU as the “global” device for all tensorflow calls in the interpreter? – SuperCiocia May 22 '20 at 00:59
Hum, if you modify your script that should be easy to implement. Just pass a variable to switch. – Sraw May 22 '20 at 01:38
1

My slight issue is that I have entire classes and methods inside them that are using tensorflow all over the place. I wanted the user to be able to set a Boolean when creating an instance of the class. Boolean that sets whether the GPU or the CPU will be used for all tensorflow methods – SuperCiocia May 22 '20 at 06:59

How to force tensorflow to forget about the GPU, after it has been found?

1 Answers1