I upvoted but didn't justify the vote, so here goes: signature information is readily available and the PL structure seems to be standardised so it would appear to be moderately low effort approach that would significantly increase what can be done in PL at a stroke - optional/kwarg arguments could either have UI elements for the Settings area generated automatically (nicer!), or the user could simply specify them in custom code (easier)

One caveat however: there are other TF packages such as TensorFlow Probability, which I would very much like to use. However, simply adding TFP layers such as densevariational does require additional effort since the output value (prediction) is no longer a single value but needs to be accessed via .mean(), std_dev() [or similar, IDK exactly)

That said, I think TFP statistical info will also contribute greatly to the value of exploitable models (other that e.g. classifications with softmax outputs where label probabilities are standard) where some idea of the confidence in the output is desirable.