Test loading/inference on all supported Python and TF versions

[mikegerber]

Learning from qurator-spk/eynollah#87/qurator-spk/eynollah#87 I think we should test loading/test inference on all supported Python and TF versions. The worst that can happen is to learn what isn't working.

I'd do this after #61.

[mikegerber]

The best that can happen: Have an understanding of how to save a model so it can be loaded under all (supported) circumstances.

[bertsky]

Indeed. And to make it work better, as a precaution, we should try to get Calamari to support SavedModel format. In fact, the problem has already surfaced.

[mikegerber]

@bertsky I agree (tentatively, as I have not yet made myself familiar with it, but it sure looks like the best solution)

[mikegerber]

We already do this (the testing of Python 3.6 to 3.10), I just forgot about it :-)

• Python 3.11 (Test on Python 3.11 #77) needs to be reviewed
• [ ] Why is the CircleCI result not accessible/private? (→ Moved to Review tests #72)
• [ ] Consider scheduled tests (TensorFlow dependency tends to break, and I only notice when something new is pushed) (Already in Review tests #72)

[mikegerber]

Python 3.11 does not work currently and I don't think I'll investigate further until we can update to Calamari 2. (→ #77)

[mikegerber]

(Moved comment to #72)

[mikegerber]

Closing as this is already done and for the other stuff we have #72.