[RFC] Dataset class. Cross-module download, duration, from_jams, load, to_jams, and validate

[lostanlen]

Reduce module-level code to the bare essentials:

1. README
2. official name
3. BibTex reference(s)
4. download remotes
5. cached properties
6. metadata parsing (optional)

Closes #196, closes #197, closes #210, closes #217
Offers a development path towards closing #81, #153, #176, #184, and #196

New features:

• modules no longer have a to_jams implementation. The to_jams method is shared across modules, and is implemented in the parent class.
• metadata is now a cached property of Dataset. I re-used the implementation of LargeData so that there's no loss in performance. Dataset construction is still very fast.
• BibTeX citation is not only printable via cite(), but also accessible as a string via dataset.bibtex
• The Dataset class overloads __getitem__, so it's still possible to load tracks one by one. I figured that this would be easier to use for newcomers than for them to learn about the Track constructor
• load_index() now turns the index paths into machine-specific absolute paths. As a result, there is no need to store data_home in the Track object anymore.
• a new function dataset.choice() picks a Track in the Dataset uniformly at random and loads it.
• dataset.download now has a verbose flag and has an opt-in for remotes. I called the kwarg download_items, in accordance with Download() refactor #216, but the kwarg seems a bit long in my opinion. Perhaps we can find a shorter name? I'll defer to @magdalenafuentes for this decision.
• validate is now possible both at the Track level and the Dataset level. Ties down with Check metadata consistency when running validator() #213
• new utility function from_jams which converts a JAMS Annotation into mirdata namedtuples: BeatData, ChordData, KeyData, SectionData.
• "fail-safe" mode allows to load some metadata of a Dataset before anything is downloaded (not event a dataset-wide CSV annotation file). This is done via a static method named parse_track_id. I am using this in Guitarset. This ties down with mirdata doesn't work cleanly for datasets not on disk #128 and resolves Throw warning when annotation path does not exist #196 quite nicely. (we can error / warn / pass if metadata is not found).

Demo:

import mirdata.dataset
gset = mirdata.dataset.Dataset(mirdata.guitarset)

# Fail-safe load. This flag will be called differently eventually. I'll let you decide
gset.load(flag196="pass")

# We haven't downloaded anything yet, but we already have some metadata, obtained by parsing track_ids from the JSON index.
print(gset.choice())

# Now download the annotation.
gset.download(download_items=["annotation"])

# Now we have beats, keys, and chords. This is done with the `from_jams` utility
print(gset.choice())

Worth noting that these new features come alongside a ~4x reduction in line count.
(although i haven't removed anything from the previous implementation because i want unit tests to keep working)

The new class is called track2.Track2 for the time being

Covered loaders:

• beatles
• gtzan_genre
• guitarset
• ikala
• medley_solos_db
• medleydb_melody
• medleydb_pitch
• orchset
• rwc_classical
• rwc_jazz
• rwc_popular
• salami
• tinysol

(i'm not counting DALI because DALI is broken at the moment)

• ensure backwards compatibility
• ensure that test coverage increases
• ensure that the LOC net count is negative
• bikeshed the name of the "failsafe" flag (issue Throw warning when annotation path does not exist #196). I called it flag196 temporarily

Let me know your thoughts!

[codecov]

Codecov Report

Merging #219 into master will not change coverage by %.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master     #219   +/-   ##
=======================================
  Coverage   79.19%   79.19%           
=======================================
  Files          21       21           
  Lines        2442     2442           
=======================================
  Hits         1934     1934           
  Misses        508      508