No documentation on how to profile the memory usage for TF


URL(s) with the issue:

Description of issue (what needs changing):

Clear description

TF profiler only outputs the CPU information with the default configuration, but I was not able to find neither in the documentation nor on the StackOverflow etc. how to set it up for monitoring the memory usage.

I get only this message on the “memory_profile” tools page:
“There is no memory profile to display because there were no memory activity data in the captured duration. “

TF and tensorboard versions are both 2.3.0.
Profiling is run on the machine without GPU, under the non-root user, in Jupyter Notebook.

Usage example

No code example for the memory profiling is provided, hence it is not clear whether some setting should be provided to the profiler.

1 possible answer(s) on “No documentation on how to profile the memory usage for TF

  1. This tool reports memory consumption on the device (i.e. GPU or TPU); not on the host (CPU).

    For the GPU issue (no breakdown to TF Ops), this is a bug that we found and fixed after TF 2.3 release. You will need to get the latest TF ( You don’t need a newer profiler plugin (i.e. the TF 2.3 release is okay).