Tensorflow-Metal Errors

Hi i am trying to set up tensorflow-metal as instructed by https://developer.apple.com/metal/tensorflow-plugin/

when running line (python -m pip install tensorflow-metal) I get the following error: ERROR: Could not find a version that satisfies the requirement tensorflow-metal (from versions: none) ERROR: No matching distribution found for tensorflow-metal

According to the troubleshooting section: "Check that the Python version used in the environment is supported (Python 3.8, Python 3.9, Python 3.10)." My current version is Python 3.9.12.

Any insight would be great!

Up vote post of LF_Alberts Down vote post of LF_Alberts
547 views
Posted by
LF_Alberts
Reply
Add a Comment

Replies

Hello,

Thank you for reaching out about this issue. I am unable to reproduce the error by following the linked instructions & using the specified python version. Could you provide the full list of commands executed, if possible?

Posted by
Apple Staff
Up vote reply of Apple Staff Down vote reply of Apple Staff
Add a Comment

I have just found out the version of tensorflow-metal==1.1.0 must be along with tensorflow-macos==2.14.0. So, If you use "pip install tensorflow-macos" alone. You will get 2.15.0 or the latest. So you must install the version which tensorflow-metal is supported. For now is like I mention earlier. Let try condo environment to mess with the system safely. Believe me it takes me some time and some space (garbage) to achieve.

Posted by
jcnwrk
Up vote reply of jcnwrk Down vote reply of jcnwrk
Add a Comment

Thanks for the tip as this is helpful...something I definitely missed.

However, now I have python 3.10, tensorflow 2.14 and tensorflow-metal 1.1 installed and no errors messages, but I get non-sensical solutions to my problems. I remove tensorflow-metal and everything works as expected.

With no error messages this is difficult to troubleshoot and understand where the issue is coming from. If anyone has any tips to dig in and troubleshoot please share as I suspect many will benefit from your experience.

Posted by
wpilgri
Up vote reply of wpilgri Down vote reply of wpilgri
Add a Comment

@wpilgri I have also faced unexpected results using tensorflow-metal. After some investigations it seems linked to the reLu fonction that seems buggy in the version 1.1.0. Changing the activation function to tanh fixed it. Unfortunately, in some cases the reLu function gives better far better results !

Posted by
maitreGui
Up vote reply of maitreGui Down vote reply of maitreGui

  • @maitreGui Thanks for the tip! I tried out your suggestion and can confirm that changing from reLu to tanh for the activation function fixed by problem! Excellent debugging on your part and thank you very much for sharing. I can only hope this is fixed in the next version.

    —
    wpilgri
Add a Comment

if you are running Conda, type the following prior to the pip command. export SYSTEM_VERSION_COMPAT=0

Posted by
cfoxzor
Up vote reply of cfoxzor Down vote reply of cfoxzor
Add a Comment