tensorflow-metal on M1: runs for 16 minutes, then hangs

Question

Level 1

21 points

tensorflow-metal on M1: runs for 16 minutes, then hangs

Yesterday I seemed to succeed installing components to run TensorFlow/Keras on my M1 MacBook Pro. I started with another recipe, but it was this one that seemed to work: “Getting Started with tensorflow-metal PluggableDevice” (Tensorflow Plugin - Metal - Apple Developer).

I was able to run a simple model training run (on MINST) which took about 2 minutes, and seemed to work fine.

This morning I modified a Jupyter notebook from my own project and launched a training run. I expected it to run about 17 hours. After 16 minutes and 45 seconds it just hung. I tried it again and got the same result. Exactly the same as near as I could tell. It was on epoch 2/100 and batch (sub-batch?) 2493/4000 of the training run. This evening I rebooted the laptop and tried again. Same result.

Most of this is new to me so it is easy to believe I made some mistake during the installation. But I don’t have much to go on. Any hints/suggestions would be appreciated.

MacBook Pro Apple Silicon

Posted on Jul 18, 2022 7:24 PM

Reply

Answer 1

Top-ranking reply

cwr Author

Level 1

21 points

Jul 21, 2022 12:29 PM in response to cwr

Ah, never mind. I confuse Developer Forums - Apple Community with Apple Developer Forums. The Tensorflow Plugin - Metal - Apple Developer doc points to a tensorflow-metal forum at the latter. There I found this unresolved bug report, of which mine seems to be a duplicate:

The new tensorflow-macos and tenso… | Apple Developer Forums

Reply

Answer 2

cwr Author

Level 1

21 points

Jul 21, 2022 11:05 AM in response to cwr

Pinging this again just in hopes of getting some kind of a hint.

I moved aside all my Conda environments and again carefully followed Tensorflow Plugin - Metal - Apple Developer.

And I still see exactly the same hang (epoch 2/100, batch 2493/4000). Worth noting that this same Jupyter notebook (except for from source/destination pathname settings) had worked fine using Google Colab.

(I'm running macOS Monterey 12.4 on a M1 Max MacBook Pro (16-inch, 2021))

Reply

tensorflow-metal on M1: runs for 16 minutes, then hangs

Similar questions