Hyperparameter optimization

[Luke-Poeppel]

Some Sunday TODOs...
Fixes #195 (umbrella).

• Re-run CF3D on composition set (try <0.8, 0.1, 0.1>). Does it still converge to 100%? (Yes 😄)
• Change default cost function to 3DCF
• Re-add SLSQP script.
• Clean up plotting scripts. Decide on plots for Transcription vs. Composition accuracy (together or separate).

Sub-Issues:

• Add test for Ex106 #190
• Weird result #179
• NC-Ex114 Missing Spondee?  #145

[Luke-Poeppel]

Accuracy up to 76.63% for the transcriptions (200/261)!

[Luke-Poeppel]

Using <0.8, 0.1, 0.1>, we still get 100% on the encoded works. Note: subdivision has to be enabled for the compositions. There's a lot of repeated code, though, which I'd like to simplify eventually.

sept_haikai_0
[21, 21]
sept_haikai_1
[21, 21]
liturgie
[29, 29]
livre_dorgue_0
[50, 50]
livre_dorgue_1
[15, 15]
(136, 136, 1.0)

[Luke-Poeppel]

Finer grid search shows:

[0.81, 0.1, 0.09]
100%|████████████████████████████████████████████| 103/103 [00:15<00:00,  6.62it/s]
100%|████████████████████████████████████████████████| 5/5 [00:21<00:00,  4.29s/it]

([0.81, 0.1, 0.09], '->', [334, 397])
[0.81, 0.11, 0.08]
100%|████████████████████████████████████████████| 103/103 [00:15<00:00,  6.65it/s]
100%|████████████████████████████████████████████████| 5/5 [00:21<00:00,  4.27s/it]
([0.81, 0.11, 0.08], '->', [337, 397])

84.89% across datasets. Using [0.81, 0.1, 0.09].

[Luke-Poeppel]

Remaining issues:

• There's still an onset range bug with split_extractions. See the series of Proceleusmatic in SF1.
• I added a condition to ensure any remaining sinks that aren't in the path (but still can be) are chosen. But SF1 ignores a viable trochee at the end.
• It seems like Tribrachs are getting consistently ignored... Probably good idea to add Ditribrach for consitency. Might solve some of this.