Skip to content

Instantly share code, notes, and snippets.

@v-i-s-h

v-i-s-h/conda-tf.md

Created April 17, 2020 05:51
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save v-i-s-h/eb32bf83d479b391457204df2e817107 to your computer and use it in GitHub Desktop.
Save v-i-s-h/eb32bf83d479b391457204df2e817107 to your computer and use it in GitHub Desktop.
Download ZIP
Tensorflow performance tips
Raw
conda-tf.md

Notes for performance optimization

Training in local CPU (When using conda tensorflow)

Consider setting the following optimization flags while training.

  • KMP_BLOCKTIME=0
  • KMP_AFFINITY=granularity=fine,verbose,compact,1,0
  • KMP_SETTINGS=1
  • OMP_NUM_THREADS=8 Also, use the following code after importing tf.
import os
if 'OMP_NUM_THREADS' in os.environ:
    NUM_THREADS = int(os.environ['OMP_NUM_THREADS'])
    tf.config.threading.set_inter_op_parallelism_threads(NUM_THREADS)
    tf.config.threading.set_intra_op_parallelism_threads(NUM_THREADS)
    print(f"Using {NUM_THREADS} threads")
else:
    print("Running with default configuration. Consider setting optimization flags while starting jupyter.")

Example:

  • When starting jupyter-lab, use
    KMP_BLOCKTIME=0 KMP_AFFINITY=granularity=fine,verbose,compact,1,0 KMP_SETTINGS=1 OMP_NUM_THREADS=8 jupyter-lab
  • When starting a python script, use
    KMP_BLOCKTIME=0 KMP_AFFINITY=granularity=fine,verbose,compact,1,0 KMP_SETTINGS=1 OMP_NUM_THREADS=8 python <script_to_run.py>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment