Limit number of cores used in Keras

Question

I have a shared machine with 64 cores on which I have a big pipeline of Keras functions that I want to run. The thing is that it seems that Keras automatically uses all the cores available and I can't do that.

I use Python and I want to run 67 neural networks in a for loop. I would like to use half of the available cores.

I can't find any way of limiting the number of cores in Keras... Do you have any clue?

Try adding `from keras import backend as K; K.set_session(K.tf.Session(config=K.tf.ConfigProto(intra_op_parallelism_threads=32, inter_op_parallelism_threads=32)))` before running the models. — Yu-Yang, Sep 26 '17 at 08:27
The number of cores is controlled in TensorFlow, not in Keras. — Yu-Yang, Sep 26 '17 at 08:27
This doesn't seem to work... The thing is that my keras models are in a python class. to fit the model i use a function i implemented in this class so im not sure where I am supposed to put this line ? — Mohamed AL ANI, Sep 26 '17 at 08:42
Normally I put it in the beginning of a file with other `import` statements. — Yu-Yang, Sep 26 '17 at 08:58
Thank you it worked ! I just put this line before each fit I do — Mohamed AL ANI, Sep 26 '17 at 09:22
@Yu-Yang sorry if this is a dumb question, but don't you risk having 32 threads for intra-op and 32 for inter-op, so 64 in total ? — gdelab, Sep 26 '17 at 09:32
Good point. I think the bottleneck will be at the intra-op pool, since most of the operations in Keras are multithreaded matrix operations. — Yu-Yang, Sep 26 '17 at 10:46

score 23 · Accepted Answer · edited May 23 '22 at 00:49

23

As @Yu-Yang suggested, I used these lines before each fit:

from keras import backend as K
K.set_session(K.tf.Session(config=K.tf.ConfigProto(intra_op_parallelism_threads=32,
                                                   inter_op_parallelism_threads=32)))

Check the CPU usage (htop) : htop example

edited May 23 '22 at 00:49

Alpha

7,586
8
59
92

answered Sep 26 '17 at 09:26

Mohamed AL ANI

2,012
1
12
29

1

For information, in order to complete the answer, here you are the tf ConfigProto, containing all the options https://github.com/tensorflow/tensorflow/blob/r1.4/tensorflow/core/protobuf/config.proto – Filippo Mazza Jan 18 '18 at 11:22
3

This doesn't seem to work for inference when loading a saved model. – Steven Dec 13 '18 at 01:41

score 2 · Answer 2 · answered Jul 31 '20 at 12:34

As mentioned in this solution, (https://stackoverflow.com/a/54832345/5568660)

if you want to use this using Tensforflow or Tensorflow_gpu, you can directly use the tf.config and feed it to the session:

config = tf.ConfigProto(intra_op_parallelism_threads=32, 
                        inter_op_parallelism_threads=32, 
                        allow_soft_placement=True)

session = tf.Session(config=config)

Limit number of cores used in Keras

2 Answers2

Linked