Split training and features from application

[hagenw]

As correctly stated in the documentation, feature extraction and training is not needed if you just want to run the model.
So it might be worth in the long run to split this package as it installs way too many dependencies for just using the model. This makes it also very likely that a user will get some conflicts if she tries to install it together with other models.
One solution for the application would be to use a more generic framework like ONNX, which would remove all the tensorflow related dependencies.

For all the others I haven't checked yet, which are only needed for feature extraction and training, but I hope there will be a few more packages.

This is nothing you have to do for the JOSS review, but just a hint for future versions.

[BreezeWhite]

Hi, thanks for the suggestion. I'll do some survey on how to do it.

[keunwoochoi]

(Commenting to link this to openjournals/joss-reviews#3391)