Some primitives could be more aligned with best practices

[mitar]

For example, why has AudioDatasetLoaderPrimitive to be from Dataset to DatFrame? Why not just use denormalize + then do DataFrame to DataFrame?

Ideally, the boundaries of the primitive would be very similar to that in common primitives, but maybe with additional features (it looks like you support sampling). Then TA2s can interchange the primitives and see which one works better in a suitable combination.

Also, type is:

AudioDatasetLoaderPrimitive(transformer.TransformerPrimitiveBase[container.Dataset, container.List, Hyperparams])

But you are really outputting a DataFrame, not List?

I suggest you use a variable at the beginning of the file and then use that variable everywhere, to assure things stay in sync. Or run mypy to validate you code. :-)

[mitar]

Also, instead of operating on networkx objects (which will go away after summer), the current direction is to normalize graphs first into edge-list format and go from there.

[cdbethune]

We definitely need to take another pass over the primitive set - there are some obvious issues like those you mention above, but beyond that, we didn't make as much progress as we wanted to on the exposure of the various hyperparameters, especially for the different learners. We're also not well aligned with some of the more general control hyperparams that are used across common primitives. I'm hoping to touch base with TA2 teams at the workshop to figure out what we should focus on now that we've got a first pass of everything submitted and validating.

[mitar]

I mean, the first thing would be to assure that you can express all your existing baselines exline pipelines now. So all exline scores used for evaluation comparison should have a corresponding pipeline. Focus on that.

I agree that the rest can be done later.