Make it a package

[ocefpaf]

This is not ready to be merged yet!

I'm not sure if the classification/ and examples/ should be part of the package or not.
Please let me know what you think and if this is the right path to make this Software an installable package. I plan to send a quick tarball and install tests in the next commits.

[rsignell-usgs]

@ocefpaf , thanks so much for taking a stab at this!
Being able to install this from conda-forge would be awesome!

[kvos]

@ocefpaf that's great, much appreciated! I'm a newbie in terms of packaging so I'm happy to follow your advice.
By making it a package, what would the workflow be for new users?
The classifiers (in the /classification/models subfolder) are necessary to run the toolbox. Also I think it's good to have the example data so users can run the tutorial and see how each input file needs to be formatted.

[ocefpaf]

@ocefpaf that's great, much appreciated! I'm a newbie in terms of packaging so I'm happy to follow your advice.
By making it a package, what would the workflow be for new users?

There are a few options like shipping with the data and examples or adding functionality to download them on demand.

The classifiers (in the /classification/models subfolder) are necessary to run the toolbox. Also I think it's good to have the example data so users can run the tutorial and see how each input file needs to be formatted.

I need to check the pickles you have there b/c they may contain Python version specific code. Ideally we should move away from pickles. I'm not familiar with the data in there but if there is another format that makes sense for you library?

[ocefpaf]

It turns out this is not easy! See https://uwekorn.com/2021/04/26/implications-of-pickling-ml-models.html

I'll keep researching to see what we can do ship the trained models.

[kvos]

I think you can also save the mlp models as .h5, is that a better format?

[ocefpaf]

I think you can also save the mlp models as .h5, is that a better format?

I'm not experience enough to say but, as long as it does not tie you to a specific environment and the versions of the packages used there, it would work.