Recover Pipeline Architecture Type

[epeters3]

Since we are generating pipelines of multiple kinds of architectures, we need a way to be able to easily tell, given a created pipeline document, what it's pipeline architecture is. Perhaps if a basic "arch_type" field was added to each pipeline JSON document before being written to the Mongo DB, that would be enough. Possible values for the field could be one of ["straight-<k>", "ensemble-<k>", "random", "stacked-<k>"], where <k> is a relevant value e.g. "straight-5 would mean it's a straight pipeline of length 5, "ensemble-3 would mean its an ensemble of three pipelines, etc.

[epeters3]

I have a branch up for this (track-arch-type). I'm just waiting on some more info from our D3M friends before I make some final changes and open a PR.

[bjschoenfeld]

Is this information based on how the pipeline was generated or what the pipeline looks like? For example, a randomly generated pipeline could look be straight, an ensemble, etc.

[epeters3]

For simplicity's sake, I have implemented it to track data just about how it was generated. For an ensemble, I'm tracking this data:

{
    "pipeline_type": "ensemble",
    "attributes": {
        "width": 3,
        "subpipeline_length": 2
    }
}

For a randomly generated pipeline (which isn't implemented yet and won't be a part of the PR addressing this issue), we could track data like this:

{
    "pipeline_type": "random",
    "attributes": {
        "depth": 4,
        "max_width": 3,
        "max_num_inputs": 3
    }
}

[epeters3]

Since these fields will be persisted in the D3M MtL DB, it would be good to make the field names more appropriate for that scope. For example maybe "pipeline_type" should be "pipeline_generation_method" or something like that, and "attributes" should be "generation_parameters" maybe.