Albert Zeyer

Hm, `available_for_inference` is a bit more controversial. So I would keep that for a separate discussion, and for a separate change.

This here is a case where we do not necessarily need a new `behavior_version` and instead could just print a deprecation warning.

Yea, ok. Feel free to create a PR. Do you have suggestions for the prefix? `root:`? > I heard it is planned to remove the from parameter from subnetworks or...

To implement this, you basically would search for all current `"base:"` prefix logic we currently have anywhere in RETURNN, and would add such `"root:"` prefix logic accordingly. There are probably...

Commit a31f683b006fac8328b1eccfa80d035930254b46 might have improved things. But not sure. With that commit, all the procs seems to reach the exit code. I see 4 times `Trainer not finalized, quitting. (pid...

Note that sharing params is probably easier with the new `name_scope` logic. But I'm not arguing that this should not be fixed.

But still it should be done like that. This just shows that the tests don't cover this. I wonder actually, I thought e.g. `loss` should always become a `tf.Tensor`.

Can you open a separate issue also about what exactly broke, with some details what broke?

> Removing the `-std=...` flag as default in #896 ... Also, the flag was not removed?

> I think one starting point would be to have certain test cases that are configured with: `runs-on: ubuntu-16.04` I think Ubuntu 18.04 is the oldest Ubuntu version which is...