TensorFlow : CPU warnings

tensorflow

#1

Each time I run a TF programme it runs fine but starts out warning me in about 20 lines of messages that I really should be putting my Intel CPU to better use by using the MKL instructions. So I put several hours into trying to build a custom version of TensorFlow and succeeded in compiing a wheel file it but it fails to import properly at this point so I will probably have to revert. My question of the Julia community is whether there is a Julia way to suppress these warnings or whether they are fixed in the TF code and I just have to ignore them.

The reason for the failure of TF to load into python3 appears to be related to numpy versions; the numpy I am using regularly is not the same version that the TF build used, but that is not a question for this forum unless anyone else saw the same thing.


#2

Well I managed to sort out the numpy thing, now my version of TensorFlow seems to be loadable in python3 and julia, but I am still getting the warnings. So I guess the MKL libraries did not compile in despite following instructions very carefully, ENV vars set and so on. Quite a few hours wasted.


#3

It seems the warnings can be killed with:
export TF_CPP_MIN_LOG_LEVEL=2 on Linux.
Seems to work well, but fundamental problem not resolved.


#4

If you’re talking about TensorFlow.jl, here are the instructions for building a custom TensorFlow binary optimized for your CPU.


#5

Thanks, yes I was talking about Julia tensorflow. It is easy to get the impression that Julia just runs what Python runs, but it seems that is not the case. I have now built the .so and it runs fine without the warning messages.