tensorflow: training network on one GPU but occupy the space of the other

Question

I have two GPUs, and I just would like to use one GPU to train a network by tensorflow. When I train it, the code use all the memories of two GPUs, but only one GPU is working:

I do not know why and how to solve this problem.

do `export CUDA_VISIBLE_DEVICES=0` before running your script — Yaroslav Bulatov, Dec 14 '16 at 21:22

score 2 · Answer 1 · edited May 23 '17 at 12:13

2

Try setting:

config = tf.ConfigProto()
config.gpu_options.allow_growth=True
sess = tf.Session(config=config)

as in question.

edited May 23 '17 at 12:13

Community

1
1

answered Dec 14 '16 at 19:50

sygi

4,437
2
33
52

score 0 · Accepted Answer · edited May 23 '17 at 12:08

0

By default tensorflow will occupy consume the memory of all available GPUs. You can either set allow growth as per @sygi's answer or make only one of the GPUs visible to tensorflow as per Yaroslav's comment. And as per this question.

edited May 23 '17 at 12:08

Community

1
1

answered Dec 15 '16 at 04:22

Russell

1,074
12
19

Finally, I use both ways. – karl_TUM Dec 15 '16 at 08:42

tensorflow: training network on one GPU but occupy the space of the other

2 Answers2