search

NVIDIA / framework-reproducibility

Providing reproducibility in deep learning frameworks
Apache License 2.0
423 stars 40 forks source link

Early versions of TensorFlow (e.g. 1.9) do not have a version attribute #37

Open SageAgastya opened 3 years ago

SageAgastya commented 3 years ago

I am getting AttributeError for tf 1.9, saying that tensorflow doesn't have "version" as attribute. Can you help?

duncanriach commented 3 years ago

Hi @SageAgastya,

I assume that you're attempting to apply the patch function from the PyPI distribution. We have paused development in the middle of a major update and will not be able to make changes to the PyPI distribution anytime soon. However, even if version 1.9 of TensorFlow had a version attribute, and therefore didn't exit with a TypeError for that reason, the patch code would proceed to exit with another TypeError and a message saying, "No patch available for version 1.9 of TensorFlow".

The patch function is not intended to be executed on any NGC TensorFlow containers nor on versions of stock TensorFlow earlier than 1.14 and is designed to throw a TypeError if that is attempted.

I hope this helps. I'm going to close this issue, but I'm happy to continue this discussion, if you wish.

Duncan

duncanriach commented 3 years ago

Actually, I'm going to leave this open so that it serves as a marker to possibly address this issue in a future release. We could check if there is a version attribute and if there is none then we could assume that the TensorFlow version is earlier than 1.14.

duncanriach commented 3 years ago

Thank you, @SageAgastya, for bringing this to my attention.