README.md

In pmcharrison/hvr: Harmonic Viewpoint Regression

hvr

The hvr package implements the harmony viewpoint regression model from Peter M. C. Harrison’s PhD thesis. This model generates predictions for chord progressions on the basis of an assortment of continuous and categorical perceptual features.

Installation

From GitHub:

if (!requireNamespace("devtools")) install.packages("devtools")
devtools::install_github("pmcharrison/hvr")

Example usage

x <- model_dataset(corpus_test = hcorp::popular_1[1],
                   corpus_pretrain = hcorp::popular_1[2:10],
                   output_dir = "tmp", 
                   viewpoints = hvr_viewpoints[c("root_pc", 
                                                 "hutch_78_roughness",
                                                 "spec_sim_3")],
                   poly_degree = 1)
#> 1/5: Computing viewpoints...
#> Computing training viewpoints...
#> Analysing test sequence 1 out of 1...
#> Computing spectral similarities (half life = 3)...
#> 
#> 2/5: Computing PPM analyses...
#> Computing PPM analyses for fold 1 out of 1...
#>   Pretraining long-term models...
#>   Predicting test sequence 1 out of 1...
#> 
#> 3/5: Constructing model matrix...
#> Getting model matrix for continuous viewpoints...
#> Computing moments...
#> Getting model matrix for discrete viewpoints...
#> Saving outputs...
#> 
#> 4/5: Fitting viewpoint regression model...
#> Loading data...
#>   done.
#> Fitting conditional logit model with optimiser L-BFGS-B...
#> i = 1, cost = 14.5849051613387
#> i = 2, cost = 10.6736467958291
#> i = 3, cost = 7.47732891193488
#> i = 4, cost = 6.21768515070835
#> i = 5, cost = 5.72067609993417
#> i = 6, cost = 5.59570082968013
#> i = 7, cost = 5.60478721301665
#> i = 8, cost = 5.57861243746749
#> i = 9, cost = 5.56841067151148
#> i = 10, cost = 5.56797160957621
#> i = 11, cost = 5.56798982125514
#> i = 12, cost = 5.56796715247232
#> i = 13, cost = 5.56796616667616
#> i = 14, cost = 5.56796616601682
#> Computing permutation feature importances...
#> 
#> 5/5: Generating final predictions...

plot(x$pred$information_content,
     xlab = "Chord position",
     ylab = "Surprisal (bits)",
     type = "l")

ggplot2::theme_set(ggplot2::theme_classic())
plot_perm_int(x$mod, error_bars = TRUE)

Usage notes

API

The top-level function is model_dataset, which automates the standard modelling pipeline for harmonic viewpoint regression. This function wraps several lower-level subroutines (e.g. viewpoint computation, PPM analyses) which can be run individually if desired. See ?model_dataset for more details.

Training set sizes

Optimising feature weights is memory-intensive, and best done with datasets of \< 1,000 chords. Larger datasets can be modelled using downsampling. Training categorical viewpoint models is relatively efficient; the implementation can cope well with datasets of c. 100,000 chords.

Future development

We have exciting ideas about how to improve the computational tractability, robustness, and interpretability of the model. We expect to be able to share an updated version by 2021, but feel free to contact Peter Harrison (pmc.harrison [at] gmail.com) before then for a progress report and potentially a draft implementation.

pmcharrison/hvr documentation built on April 14, 2020, 2:47 a.m.