tf.Variable support about tfgo HOT 1 OPEN

Because tfgo uses the official TensorFlow''s Go bindings under the hood (a fork I created because these bindings are not go-gettable) and these bindings do not support tf.Variable.

The Go bindings are almost abandoned, just think about that we are constrained to the static graph + session execution (without tf.Variable support, for using them you have to define the model in Python as a SavedModel and then load it in tfgo) while the C API has the support for the Eager Execution and creating variables on the fly.

For adding the support for tf.Variable the only possible way is to write in my fork ( https://github.com/galeone/tensorflow ) the tf.Variable support in C, and expose them in Go via cgo.

But it's not worth it IMO since the support for static-graph + session execution is enough for every application: just write the graph in Python (a function decorated with @tf.function) and export it as a SavedModel - the graph can contain any layer, with variables and so on. The exported graph can be loaded into tfgo and used, being the state (the value of the variables) managed by the TensorFlow runtime (the C library) and not by the bindings.

What it would be worth doing, instead, is exposing all the new eager C API to the Go bindings (my fork) - but that's a lot of work and people at Google are not doing it, hence I suppose it's something not trivial

from tfgo.