use_session_with_seed is deprecated in Keras 2.6.0

[boxydog]

I tried

library(keras)
keras::use_session_with_seed(
    SEED,
    disable_gpu = FALSE,
    disable_parallel_cpu = FALSE)

on Kaggle, and it says:

Warning message in keras::use_session_with_seed(SEED, disable_gpu = FALSE, disable_parallel_cpu = FALSE):
“use_session_with_seed will be deprecated in the future. use set_random_seed instead.”
Error in keras::use_session_with_seed(SEED, disable_gpu = FALSE, disable_parallel_cpu = FALSE): use_session_with_seed will be deprecated in the future. use set_random_seed instead.
Traceback:

1. keras::use_session_with_seed(SEED, disable_gpu = FALSE, disable_parallel_cpu = FALSE)
2. stop(msg)

However, set_random_seed does not look to be exported:

keras::set_random_seed(SEED)
Error: 'set_random_seed' is not an exported object from 'namespace:keras'

set_random_seed looks exported from tensorflow, but I'm not sure if it has the same functionality.

Versions:

packageVersion('keras')
[1] ‘2.6.0’

foo <- reticulate::py_run_string('import keras;  vers=keras.__version__')
foo$vers
'2.6.0'

[t-kalinowski]

Hi, thanks for filing. set_random_seed is exported by the {tensorflow} package. You can access it with tensorflow::set_random_seed()

[boxydog]

But does it do the same thing as use_session_with_seed? That's what I'm unclear on.

The docs for use_session_with_seed say:

The provided seed value will establish a new random seed for R, Python, NumPy, and TensorFlow. GPU computations and CPU parallelism will also be disabled by default.

If it's the one to use, shouldn't the rstudio documentation point to set_random_seed instead of this deprecated function?

I'm not just asking about a function, but the expected and documented behavior.

[t-kalinowski]

use_session_with_seed is a hold-over from TF v1, when users were expected to manage Sessions explicitly. It's no longer relevant in TF v2, and set_random_seed should be all you need.

You are right, the docs needs to be updated. Thank you!

[boxydog]

Thanks for your help.