Should scivision models be saved in an interoperable format

[edwardchalstrey1]

One issue is that if a model is trained with an old version of a package (e.g. tensorflow) it then cannot be used in later versions of Python (e.g. that only support newer tensorflow versions).

Should we make it a requirement that all scivision models in the catalog are saved in an interoperable format (perhaps ONNX, or HDF5?) - the model code in the GitHub repo of a model entry to the catalog could therefore just be a package that loads this interoperable model file (and has the predict func etc).

This could be another way of solving the issue of clashing dependencies for models (and data plugins) that have differing requirements in scivision.

[ots22]

At least as things are currently, a model is a "just" a python package, and its dependencies and package data are entirely up to that package - the details of how it works are in some sense none of our business! If there is some behaviour we want of the package (portability, for example) it might be better to require that directly, rather than trying to anticipate file types (or whatever it happens to be).

Realistically though, there probably are one or two libraries that get used all the time (tensorflow, torch) and some particular guidance does sound useful for these, and perhaps some of this applies more generally too. There seems to be something to add to a growing list of "recommendations for model authors" (or "recommendations for data providers") - including this somewhere prominent sounds very useful indeed, and especially if we can link to more community-specific advice.

What scivision doesn't (can't?) handle well with the current architecture, is ensuring a dependency on a particular python version is met (since it uses pip) - that ends up being pushed onto the user. Ensuring a model at least declares its python version requirements sounds like something we can insist on.

[ots22]

Opened #343 to collect some of these things