setting OMP_NUM_THREADS

[myoung3]

How do I set OMP_NUM_THREADS? I'm using high-core cpu and I'm trying to implement intel's recommendations here: https://github.com/IntelAI/models/blob/master/docs/general/tensorflow/GeneralBestPractices.md

As far as I can tell, I can set inter_op_parallelism_threads via tf$config.threading$set_intra_op_parallelism_threads(32L)

and channels first via k_set_image_data_format('channels_first') but since OMP_NUM_THREADS is an environment variable I'm not sure how to set it such that the python shell inherits it.

[myoung3]

this works (it successfully caused a major speedup)

n_physical <- 32
reticulate::py_run_string('import os')
reticulate::py_run_string(paste0('os.environ["OMP_NUM_THREADS"] = str(',n_physical,')'))

Please add this to the documentation somewhere.

[t-kalinowski]

The python session will inherit the environment variables that are set at the time the python session is initialized.
Additional environment variables set in R after python has initialized will not be visible to python; at that time, if you want to set environment variables in the python session, you have to modify the os$environ object directly.

Here is a short snippet demonstrating:

# ------- start of new R session ---------------
Sys.setenv("foo" = "bar")
reticulate::import("os")$environ["foo"] # python session initialized here
# "bar"

# ------- start of new R session ---------------
reticulate:::initialize_python()
Sys.setenv("foo" = "bar")
reticulate::import("os")$environ["foo"]
# Error in py_get_item_impl(x, key, FALSE) : KeyError: 'foo'

os <- reticulate::import("os")
reticulate::py_set_item(os$environ, "foo", "bar")
reticulate::import("os")$environ["foo"]
# "bar"