4

Since yesterday or the day before, I get a

ValueError: Attempted relative import in non-package

for an import in my main trainer file like

from . import mobilenet_v1 as mobilenet

when running the exact same trainer code with the exact same parameters on cloud ML using the exact same training job. I'm bundling my trainer using the gcloud tool. I tried rolling my own setup.py instead without luck. Any pointers as to what this could be caused by?

Catalin Voss
  • 91
  • 6
  • This might be a bug being rolled out. We will investigate. – rhaertel80 Oct 18 '17 at 02:28
  • Can you verify the date of your last successful run? – rhaertel80 Oct 18 '17 at 02:30
  • Oct 14, 2017, 8:42:27 AM (pacific time, I believe) – Catalin Voss Oct 18 '17 at 02:40
  • Thanks for taking a look. For what it's worth, I had set up my stuff to follow the style guidelines [here](https://www.tensorflow.org/community/style_guide) with absolute import. As such my import statements have to look like the above. It seems like you're not doing that in the samples, e.g. [here](https://github.com/GoogleCloudPlatform/cloudml-samples/tree/master/census). – Catalin Voss Oct 18 '17 at 02:41

1 Answers1

1

Looks like this was actually a Cloud ML bug. It has been fixed! Thank you for the super fast turnaround.

Catalin Voss
  • 91
  • 6
  • Hi Catalin, we are investigating the root cause of the issue. Can you please provide your module structure and the module name that's passed to gcloud? It would be also helpful if you can send us the project number and job id via cloudml-feedback@google.com. Thank you for you help! – Guoqing Xu Oct 18 '17 at 21:37
  • Sure! Will do. Thank you! – Catalin Voss Oct 18 '17 at 23:44