@TRon,
Thanks, that was it. I tried @s, while it should have been pointer(s).
@Thaddy,
Thanks, Yes, I try to solve something that is needed due to the complex design of TensorFlow, far beyond me. Probably I will have to pass an array of pointers and another array with the type of the actual variables in the first array.
This is btw. very much in line with the way, how arguments are also built in three arrays, one array of string with the argument name, one with the argument type (array of integer, with types having their code, just like in tVarRec) and a third one with the pointers pointing to the values.
@alpine
Thanks, that wrapper looks interesting, but as I can only have a very limited number of types (two for inputs and if I need to mix in then two more for outputs) a second array with the type (or with a boolean if only two need to be separated) seems easier at this point.