build inference script

[alanngnet]

Build a script to use the trained model to generate the top N closest matches to a given CQT feature array.

[alanngnet]

so far I have confirmed that original CoverHunter provided almost nothing to support true inference beyond generating (useful) embeddings for each training sample. So this issue probably needs chunking into a series of enhancement issues. 2 possible approaches: build a way to generate class centroid embeddings (one embedding per distinct song_id) or modify the training model to add a layer that has class inference built into the final layer, so the self-contained model could be used to directly output inferred song_ids. Probably leaning towards the former, so that we can then do ranking of nearest neighbors and use that in various creative ways for real-world applications.

Other implementation ideas, anyone?

[alanngnet]

building an MVP using just direct comparison of query embedding with the entire set of reference embeddings, adapting from the distance-matrix computation approach taken in eval_testset.py.

[alanngnet]

Calling this MVP good enough for now, especially since we have higher priorities for a while.