is model_type restricted or just a string?

[ryandawsonuk]

I see the example where it has value 'neural network' and appears to be a sting. I'm wondering whether there are plans to restrict it to a limited range of values or have a convention to avoid differences of wording e.g. 'neural network' vs 'nn'.

I'm also wondering whether this is enough or there should be a 'model_architecture' vs 'model_type'. For instance, you might have a NN handing image data and another handling tabular data - those are quite different things. It could also be important for other components in the cluster to be able to discover whether models handle images, tabular data or text. For example, there might be a need to look up the training data for the model and this could be in a different place for images.

[zhenghuiwang]

It is intended to be free-form string type. If some organizations want to enforce a set of canonical type names, they can always define the constant strings in a lib and ask people to use them.

On the other hand, if we enforce constant string type by default, adding new constants for unknown use cases becomes difficult.

There are many use cases that current metadata fields aren't enough to help. We will add them later based on CUJs that we want to enable.

[jtfogarty]

/kind question
/area engprod
/priority p2