Dataset.records (.tables)?

[rabitt]

For top-level metadata, sql tables, and other "cross-track" data, we've talked about supporting a Dataset.records object as a compliment to Dataset.tracks and Dataset.multitracks

Some things to decide/discuss:

• What should the API look like?
• How should we support cross record queries (e.g. sql joins)
• Can we abstract existing "metadata" into records?

[nkundiushuti]

I had two meetings with Andres Ferraro & Sergio Oramas to talk about this implementation:

• What file formats do we take as data input: csv, tsv, json
• Implementation:
  a. iterator (return records-tuples)
  b. return a data-frame (load the table with pandas and return it)
• what problems we may face: data is too large to be fit into memory; solution: generator
• what can we do to ease access: Helper functions - examples of data loading : split on time (train/test)
• MTG datasets we want to implement
  a. https://www.upf.edu/web/mtg/mumu
  b. https://mtg.github.io/melon-music-dataset/
  c. https://www.upf.edu/web/mtg/lastfm360k