IDyOM Project

h1. Running IDyOM 

{{>toc}}

h2. <code>idyom:idyom</code> 

The main workhorse function is <code>idyom:idyom</code>, which has three required arguments and a number of optional keyword arguments.

h3. Required parameters

* dataset-id: a dataset id, e.g. 1.

* target-viewpoints: a list of basic viewpoints to predict, e.g. '(:cpitch :bioi)

* source-viewpoints: a list of viewpoints to use in prediction, e.g. '((:cpintfref :cpint) :bioi)

** Passing <code>:select</code> will trigger viewpoint selection (see further options below)

See the [[List of viewpoints]] for a description of the various viewpoints available in IDyOM.  A simple call to IDyOM would be:

<pre>

CL-USER> (idyom:idyom 1 '(cpitch) '(cpitch cpint))

2.2490792

(1.9049941 2.427845 2.0234334 1.7971386 1.8213106 1.9313766 2.3758402 1.8310248

...

</pre>

This predicts the pitch values in dataset 1, based on previous pitches (cpitch) and pitch intervals (cpint).  IDyOM computes the information content for each note, and by default returns two values: the first is a mean note IC for the dataset, the second a list of mean note ICs for the individual compositions.  The first value is calculated as the mean of the second.

h3. Statistical modelling parameters

See "Pearce [2005, chapter 6]":http://webprojects.eecs.qmul.ac.uk/marcusp/papers/Pearce2005.pdf for further description and explanation of these parameters.

* @models@: the type of IDyOM model to use.  Options are:

** @:stm@ - short-term model only;

** @:ltm@ - long-term model only;

** @:ltm+@ - the long-term model trained incrementally on the test set;

** @:both@ - a combination of :stm and :ltm;

** @:both+@ -  a combination of :stm and :ltm+ (this is the default).

The LTM and STM can be configured using the @ltmo@ and @stmo@ parameters.  These accept a property list which must define ALL of the following parameters (the default values are used only when no list is supplied):

* @:order-bound@: an integer indicating the bound on the order of the model, i.e. the number of past events used by the model.  The default is @nil@, no bound.

* @:mixtures@: whether to use mixtures for the model. (Default @t@).

* @:update-exclusion@: whether to use update exclusion. (LTM default @nil@, STM default @t@.)

* @:escape@: the model's escape method.  One of @:a :b :c :d :x@.  (LTM default @:c@, STM default @:x@.)

h3. Training parameters

When using IDyOM to estimate note IC for a given dataset, the long-term models can be trained on other datasets (pretraining) and/or on the current dataset, i.e. via resampling (cross-validation).  In the latter case, the dataset is partitioned into a training set (used to train the LTMs) and a test set (for which note IC is computed).  This split is called a fold, and the modelling process can be repeated with a number of different folds in order to model the entire dataset.

* @pretraining-ids@: a list of dataset ids used to pretrain the long-term models (done before resampling).

* @k@: the number of resampling (cross-validation) folds to use.  The default value is 10.

** @1@ = no resampling, but also no training set unless the models are pretrained; 

** @:full@ = as many folds as there are compositions in the dataset

* @resampling-indices@: a list of numbers designating which resampling folds to use, i.e. a subset of @[0, 1, ..., k - 1]@.  By default, all folds are used.

h3. Viewpoint selection parameters

* dp: the number of decimal places to use when comparing information contents in viewpoint selection

** full floating point precision is used if this is <code>nil</code> (the default)

* max-links: the maximum number of links to use when creating linked viewpoints in viewpoint selection

** the default is 2

h3. Output parameters

* output-path: a string indicating a directory in which to write the output 

** output is only written to the console if this is <code>nil</code>

* detail: an integer which determines how the information content is averaged in the output: 

** 1: averaged over the entire dataset 

** 2: and also averaged over each composition 

** 3: and also for each event in each composition

h2. <code>resampling:idyom-resample</code>

The top-level function in turn passes its arguments on to a number of sub-functions which can be used independently. 

<code>RESAMPLING:DATASET-PREDICTION</code> accepts the following arguments (all but the first three are optional, keyword arguments): 

* dataset-id: a dataset id 

** e.g., 1

* basic-attributes: a list of basic attributes to predict 

** e.g., '(cpitch bioi)

* attributes: a list of attributes to use in prediction

** e.g., '((cpintfref cpint) bioi)

* pretraining-ids: a list of dataset-ids to pretrain the long-term models 

** e.g., '(0 1 7)

* k: an integer designating the number of cross-validation folds to use 

** 1 = no cross-validation, but also no training set unless the models are pretrained; 

** :full = as many folds as there are compositions in the dataset

** default = 10 

* resampling-indices: you can limit the modelling to a particular set of resampling folds

* models: whether to use the short-term or long-term models or both

** :stm - short-term model only 

** :ltm - long-term model only 

** :ltm+ - the long-term model trained incrementally on the test set 

** :both - :stm + :ltm 

** :both+ - :stm + :ltm+ (this is the default)

* ltm-order-bound: the order bound for the long-term model (the default <code>nil</code> means no order bound, otherwise an integer indicates the bound in number of events)

* ltm-mixtures: whether to use mixtures for the LTM (default <code>t</code>)

* ltm-update-exclusion: whether to use update exclusion for the LTM (default <code>nil</code>)

* ltm-escape: the escape method to use for the LTM (<code>:a :b :c :d :x</code> - default <code>:c</code>)

* stm-order-bound: the order bound to use for the short-term model (default <code>nil</code>)

* stm-mixtures: whether to use mixtures for the STM (default <code>t</code>)

* stm-update-exclusion: whether to use update exclusion for the STM (default <code>t</code>)

* stm-escape: the escape method for the STM (default <code>:x</code>)

<code>RESAMPLING:OUTPUT-INFORMATION-CONTENT</code> takes the output of <code>RESAMPLING:DATASET-PREDICTION</code> and returns the average information content. It takes the following arguments:

* predictions: the output of <code>RESAMPLING:DATASET-PREDICTION</code>

* detail: an integer which determines how the information content is averaged (these are returned as multiple values): 

** 1: averaged over the entire dataset 

** 2: and also averaged over each composition 

** 3: and also for each event in each composition

h2. <code>resampling:format-information-content</code>

<code>RESAMPLING:FORMAT-INFORMATION-CONTENT</code> takes the output of <code>RESAMPLING:DATASET-PREDICTION</code> and writes it to file. It takes the following arguments:

* predictions: the output of <code>RESAMPLING:DATASET-PREDICTION</code>

* file: a string denoting a file

* dataset-id: an integer reflecting the dataset-id

* detail: an integer which determines how the information content is averaged (these are returned as multiple values): 

** 1: averaged over the entire dataset 

** 2: and also averaged over each composition 

** 3: and also for each event in each composition

h2. Examples

h3. Mean melody IC

To get mean information contents for each melody of dataset 0 in a list 

<pre>

CL-USER> (resampling:output-information-content 

          (resampling:dataset-prediction 0 '(cpitch) '(cpintfref cpint))

          2)

2.493305

(2.1368716 2.8534691 2.6938546 2.6491673 2.4993074 2.6098127 2.7728052 2.772861

 2.5921957 2.905856 2.3591626 2.957503 2.4042292 2.7562473 2.3996017 2.8073587

 2.114944 1.7434102 2.2310295 2.6374347 2.361792 1.9476132 2.501488 2.5472867

 2.1056154 2.8225484 2.134257 2.9162033 3.0715692 2.9012227 2.7291088 2.866882

 2.8795822 2.4571223 2.9277062 2.7861307 2.6623116 2.3304622 2.4217033

 2.0556943 2.4048684 2.914848 2.7182267 3.0894585 2.873869 1.8821808 2.640174

 2.8165438 2.5423129 2.3011856 3.1477294 2.655349 2.5216308 2.0667994 3.2579045

 2.573013 2.6035044 2.202191 2.622113 2.2621205 2.3617425 2.7526956 2.3281655

 2.9357266 2.3372407 3.1848125 2.67367 2.1906006 2.7835917 2.6332111 3.206142

 2.1426969 2.194259 2.415167 1.9769101 2.0870917 2.7844474 2.2373738 2.772138

 2.9702199 1.724408 2.473073 2.2464263 2.2452457 2.688889 2.6299863 2.2223835

 2.8082614 2.673671 2.7693706 2.3369458 2.5016947 2.3837066 2.3682225 2.795649

 2.9063463 2.5880773 2.0457468 1.8635312 2.4522712 1.5877498 2.8802161

 2.7988417 2.3125513 1.7245895 2.2404804 2.1694546 2.365556 1.5905867 1.3827317

 2.2706041 3.023884 2.2864542 2.1259797 2.713626 2.1967313 2.5721254 2.5812547

 2.8233812 2.3134546 2.6203637 2.945946 2.601433 2.1920888 2.3732007 2.440137

 2.4291563 2.3676903 2.734724 3.0283954 2.8076048 2.7796154 2.326931 2.1779459

 2.2570527 2.2688026 1.3976555 2.030298 2.640235 2.568248 2.6338177 2.157162

 2.3915367 2.7873137 2.3088667 2.2176988 2.4402564 2.8062992 2.784044 2.4296925

 2.3520193 2.6146257)

</pre>

h3. Write note IC to file

To write the information contents for each note of each melody in dataset 0 to a file 

<pre>

CL-USER> (resampling:format-information-content 

          (resampling:dataset-prediction 0 '(cpitch) '(cpintfref cpint))

          "/tmp/foo.dat"

          0

          3)

</pre>

h3. Conklin & Witten (1995)

To simulate the experiments of Conklin & Witten (1995) 

<pre>

CL-USER> (resampling:conkwit95)

Simulation of the experiments of Conklin & Witten (1995, Table 4).

System 1; Mean Information Content: 2.33 

System 2; Mean Information Content: 2.36 

System 3; Mean Information Content: 2.09 

System 4; Mean Information Content: 2.01 

System 5; Mean Information Content: 2.08 

System 6; Mean Information Content: 1.90 

System 7; Mean Information Content: 1.88 

System 8; Mean Information Content: 1.86 

NIL

</pre>

Compare with "Conklin & Witten [1995, JNMR, table 4]":http://www.sc.ehu.es/ccwbayes/members/conklin/papers/jnmr95.pdf

h2. Viewpoint Selection 

Two functions are supplied for searching a space of viewpoints: <code>run-hill-climber</code> and <code>run-best-first</code>, which take 4 arguments:

* a list of viewpoints: the algorithm searches through the space of combinations of these viewpoints

* a start state (usually nil, the empty viewpoint system)

* an evaluation function returning a numeric performance metric: e.g., the mean information content of the dataset returned by <code>dataset-prediction</code>

* a symbol describing which way to optimise the metric: <code>:desc</code> mean lower values are better <code>:asc</code> mean greater values are better

Here is an example:

<pre>

CL-USER> (viewpoint-selection:run-hill-climber 

          '(:cpitch :cpintfref :cpint :contour)

          nil

          #'(lambda (viewpoints)

              (utils:round-to-nearest-decimal-place 

               (resampling:output-information-content 

                (resampling:dataset-prediction 0 '(cpitch) viewpoints :k 10 :models :both+) 

                1)

               2))

          :desc)

 =============================================================================

   System                                                Score

 -----------------------------------------------------------------------------

   NIL                                                   NIL

   (CPITCH)                                              2.52

   (CPINT CPITCH)                                        2.43

   (CPINTFREF CPINT CPITCH)                              2.38

 =============================================================================

#S(VIEWPOINT-SELECTION::RECORD :STATE (:CPINTFREF :CPINT :CPITCH) :WEIGHT 2.38)

</pre>

Since this can be quite a time consuming process, there are also functions for caching the results.

<pre>

(initialise-vs-cache)

(load-vs-cache filename package)

(store-vs-cache filename package)

</pre>