|
| 1 | +# Adding a new method to ORCA |
| 2 | + |
| 3 | +ORCA is designed to ease the process of adding new methods. You only need to add the new algorithm's class file to folder `src/Algorithms`. After that, the method will be available to the framework and configuration files can be used to automate experiments. |
| 4 | + |
| 5 | +## Method template |
| 6 | + |
| 7 | +The code has follow the following template, which basically satisfies the API defined in the `Algorithm` abstract class: |
| 8 | + |
| 9 | +```MATLAB |
| 10 | +classdef NEWMETHOD < Algorithm |
| 11 | + properties |
| 12 | + description = 'my NEWMETHOD method description'; |
| 13 | + % Parameters to optimize and default value |
| 14 | + parameters = struct('k', 5); |
| 15 | + end |
| 16 | +
|
| 17 | + methods |
| 18 | + function obj = NEWMETHOD(varargin) |
| 19 | + % Process key-values pairs of parameters |
| 20 | + obj.parseArgs(varargin); |
| 21 | + end |
| 22 | +
|
| 23 | + function [projectedTrain, predictedTrain]= privfit( obj, train, param) |
| 24 | + % fit the model and return prediction of train set. It is called by |
| 25 | + % super class Algorithm.fit() method. |
| 26 | + ... |
| 27 | + % Save the model |
| 28 | + obj.model = model; |
| 29 | + end |
| 30 | +
|
| 31 | + function [projected, predicted] = privpredict(obj, testPatterns) |
| 32 | + % predict unseen patterns with 'obj.model' and return prediction and |
| 33 | + % projection of patterns (for threshold models) |
| 34 | + % It is called by super class Algorithm.predict() method. |
| 35 | + end |
| 36 | + end |
| 37 | +end |
| 38 | +``` |
| 39 | + |
| 40 | +Where `train` is a structure with `train.patterns` being a matrix of patterns and `train.targets` being a vector with the corresponding labels. `model` class property stores the model built with the train data. |
| 41 | + |
| 42 | +## Example: adding KNN to ORCA |
| 43 | + |
| 44 | +To illustrate the one-step process of adding a new method, we will add the KNN classifier to ORCA. Just copy the file [KNN.m](KNN.m) to folder `src/Algorithms`: |
| 45 | + |
| 46 | +```MATLAB |
| 47 | +classdef KNN < Algorithm |
| 48 | + %KNN Basic k-nearest neighbors algorithm based on Euclidean distance |
| 49 | +
|
| 50 | + properties |
| 51 | + description = 'k-nearest neighbors algorithm'; |
| 52 | + % Parameters to optimize and default value |
| 53 | + parameters = struct('k', 5); |
| 54 | + end |
| 55 | +
|
| 56 | + methods |
| 57 | + function obj = KNN(varargin) |
| 58 | + %KNN constructs an object of the class KNN. Default k is 5 |
| 59 | + % |
| 60 | + % OBJ = KNN('k', neighbours) |
| 61 | + % builds KNN with NEIGHBOURS as number of neighbours to consider |
| 62 | + % to label new patterns. |
| 63 | + obj.parseArgs(varargin); |
| 64 | + end |
| 65 | +
|
| 66 | + function [projectedTrain, predictedTrain]= privfit( obj, train, param) |
| 67 | + if(nargin == 3) |
| 68 | + obj.parameters.k = param.k; |
| 69 | + end |
| 70 | +
|
| 71 | + % save train data in the model structure |
| 72 | + obj.model.train = train; |
| 73 | + obj.model.parameters = obj.parameters; |
| 74 | + % Predict train labels |
| 75 | + [projectedTrain, predictedTrain] = predict(obj, train.patterns); |
| 76 | + end |
| 77 | +
|
| 78 | + function [projected, predicted] = privpredict(obj, testPatterns) |
| 79 | + % Variables aliases |
| 80 | + x = obj.model.train.patterns; |
| 81 | + xlabel = obj.model.train.targets; |
| 82 | + k = obj.model.parameters.k; |
| 83 | +
|
| 84 | + dist = pdist2(testPatterns,x); |
| 85 | + % indicies of nearest neighbors |
| 86 | + [~,nearest] = sort(dist,2); |
| 87 | + % k nearest |
| 88 | + nearest = nearest(:,1:k); |
| 89 | + % mode of k nearest |
| 90 | + val = xlabel(nearest); |
| 91 | + predicted = mode(val,2); |
| 92 | +
|
| 93 | + % dummy value for projections |
| 94 | + projected = -1.*ones(length(testPatterns),1); |
| 95 | + end |
| 96 | + end |
| 97 | +end |
| 98 | +``` |
| 99 | + |
| 100 | +Then, you can define a configuration file such as [knntoy.ini](knntoy.ini) to describe experiments using KNN: |
| 101 | + |
| 102 | +```INI |
| 103 | +;Experiment ID |
| 104 | +[knn-mae-toy] |
| 105 | +{general-conf} |
| 106 | +;Datasets path |
| 107 | +basedir = ../exampledata/30-holdout |
| 108 | +;Datasets to process (comma separated list) |
| 109 | +datasets = toy |
| 110 | +;Activate data standardization |
| 111 | +standarize = true |
| 112 | +;Number of folds for the parameters optimization |
| 113 | +num_folds = 5 |
| 114 | +;Crossvalidation metric |
| 115 | +cvmetric = mae |
| 116 | + |
| 117 | +;Method: algorithm and parameter |
| 118 | +{algorithm-parameters} |
| 119 | +algorithm = KNN |
| 120 | + |
| 121 | +;Method's hyper-parameter values to optimize |
| 122 | +{algorithm-hyper-parameters-to-cv} |
| 123 | +k = 3,5,7 |
| 124 | +``` |
| 125 | + |
| 126 | +To run experiments described in that file, from `src` folder type: |
| 127 | + |
| 128 | +```MATLAB |
| 129 | +Utilities.runExperiments('../doc/addmethod/knntoy.ini') |
| 130 | +``` |
0 commit comments