Isolated environments for machine learning examples

[ncguilbeault]

Continuing the discussion from here: bonsai-rx/bonsai#2096 (comment)

I agree with @glopesdev about keeping each example self-contained, at least for the machine learning examples. Early on, I remember we were discussing ways to distribute large files (i.e. datasets, model weights, etc), and one idea was to distribute these via NuGet packages and rely on the bonsai bootstrapper to take care of this. I think it would be useful to keep this door open for the future, since moving to a single bootstrapped environment would mean adding a lot of overhead. Also, we took a lot of inspiration early on from the ML.NET project, and the ML.NET examples repo does something similar with each example having a unique csproj file that only pulls in the packages needed.

But I agree with your point @PathogenDavid, having to update the examples with every new package release is a total pain and it's very easy to overlook. It would be great if we could automate this step or make it easier to know when things have become outdated. It would be great if we could do something similar to what ML.NET does and be able to specify the package version once to get applied to all of the examples.

[PathogenDavid]

The ML.NET strategy is actually somewhat dated. The modern strategy is to use Central Package Management.

One thing this doesn't help with though is overriding a package with a locally built one. Ideally the examples in the documentation should always be rendered using the same version of Bonsai.ML that the rest of the documentation is built from, but this currently isn't super practical (especially not without making it so the examples repo can't be used independently of this one.)

I have more concrete ideas on solving both issues as a part of a wider Bonsai proposal that I've been noodling on but it needs more time in the oven.