Integration with TensorVision

[MarvinTeichmann]

Right now the Repository has its own stand alone code for using the model (training, evaluation). This Code is basically a copy of TensorVision Code. As of now, MediSeg is 100% compatible to TensorVision, so the model can be trained ether by using tv-train or by running python train.py. The question is, whether it should remain like this. It has been slightly touched in #3.

What we can do basically is: 1) Remove the control code and ship only ship the core model. Feature Versions can be only trained using TensorVision. 2) Keep a copy of controlling code inside the repository, making it possible to use MediSeg as stand alone.

A mixture might also be considered. One might remove training code, forcing to use TV for training, but keep evaluation or prediction code inside the repository.

I am going to keep compatibility between TV and MediSeg, so the question is mainly about whether we should provide a stand-alone Version by hand-copy code.

[MartinThoma]

I definitely want (1). We can probably add a git submodule to MediSeg, but it doesn't make much sense to me not to use TensorVision if we build it anyway.

[MarvinTeichmann]

Sure, I do prefer (1) as well. I did interpret your comment #3, as if you prefer some kind of stand-alone code for the MediSeg model.