NN models for Unity ML-agents

[danibix95]

Discuss how to deal with the problem that any trained model should be converted into Barracuda format.
In addition this open potential future works where new images are developed (scikit-learn & Pytorch), as long as the training algorithm can save their model into TensorFlow or ONNX format, which they can be converted into Barracuda one.

[danibix95]

It might be possible that in this manner the TensorFlow version constraint does not matter anymore.

In fact, while TensorFlow 1.7.1 is needed for running ml-agents trainers, if a custom model is built and then converted into the proper Barracuda format, then it does not influence TensorFlow version.

[danibix95]

It seems that to run the training, ml-agents library is not required and that it can run on TensorFlow 1.14.0