Skip to content

Latest commit

 

History

History
122 lines (91 loc) · 4.7 KB

orca_quick_install.md

File metadata and controls

122 lines (91 loc) · 4.7 KB
Raw

Installing ORCA

This is a quick installation guide. If you experiment any problems, please read the detailed installation guide. ORCA has been developed and tested in GNU/Linux systems and ported to Windows. It works in Mac using GNU compilers. It has been tested in MATLAB R2009a-R2017b and Octave >= 4.2.

Installation requirements

In order to use ORCA you need:

  • GNU gcc and g++
  • MATLAB/Octave (Octave >= 4.2), including mex.
    • MATLAB toolboxes: MATLAB compiler, Optimization, Statistics and Machine Learning. Optional Parallel Computing.
    • Octave packages (ORCA will install them for you): recent versions from octave forge of statistics, optim and dependencies io, struct. Depending on your GNU/Linux distribution you may also have to install liboctave-dev with your distribution package manager.

Download and install ORCA

To download ORCA you can simply clone this GitHub repository by using the following commands:

$ git clone https://github.com/ayrna/orca

All the contents of the repository can also be downloaded from the GitHub site by using the "Download ZIP" button.

To build ORCA, build_orca.m will check dependencies, install them for Octave, and run basic tests. From MATLAB/Octave:

> build_orca.m

For instance, to install Octave, download and install ORCA in Ubuntu 18.10:

$ sudo apt-get install octave liboctave-dev
$ git clone https://github.com/ayrna/orca.git
$ cd orca
$ octave-cli build_orca.m

Binaries

We provide binary files for several platforms (Debian based and CentOS GNU/Linux and Windows). The compressed files include the git files, so git pull should work. Binaries can be downloaded in the release page.

Compilation of mex files from system shell

If you prefer to build mex files from the Linux shell, you can use standard make tool at src/Algorithms directory:

  1. Set up the proper path for MATLABDIR or OCTAVEDIR variables, see examples at ../src/Algorithms/Makefile
  2. Run make for MATLAB or make octave for GNU Octave.

Installation testing

We provide a set of basic tests to for checking that all the algorithms work, both using ORCA's API and experiment scripts (see tutorial for more information).

The way to run the tests checking the API (see single test scripts) is the following (running time is ~12 seconds):

>> cd src/
>> runtests_single
...
.........................
Performing test for SVORLin
Accuracy Train 0.262222, Accuracy Test 0.266667
Test accuracy matchs reference accuracy
Processing svrTest.m...
.........................
Performing test for SVR
Accuracy Train 0.995556, Accuracy Test 0.973333
Test accuracy matchs reference accuracy
All tests ended successfully

To run the tests checking the experiment scripts (running time is ~123 seconds):

>> runtests_cv
...
Running experiment exp-svr-real1-toy-1
Processing Experiments/exp-2017-11-16-13-59-1/exp-svr-real1-toy-1
Calculating results...
Experiments/exp-2017-11-16-13-59-1/Results/toy-svr-real1/dataset
Experiments/exp-2017-11-16-13-59-1/Results/toy-svr-real1/dataset
Test passed for svr
All tests ended successfully

If any of these tests fail, please read the detailed installation guide.

Use ORCA as a toolbox

The first of three tutorials covers the basic use of ORCA as a toolbox but also as an experimental framework: 'how to' tutorial (Jupyter Notebook, MD).

For instance, to use ORCA as a toolbox:

% Create an Algorithm object
addpath('src/Algorithms/')
kdlorAlgorithm = KDLOR();
% Load dataset
load exampledata/1-holdout/toy/matlab/train_toy.0
load exampledata/1-holdout/toy/matlab/test_toy.0
train.patterns = train_toy(:,1:(size(train_toy,2)-1));
train.targets = train_toy(:,size(train_toy,2));
test.patterns = test_toy(:,1:(size(test_toy,2)-1));
test.targets = test_toy(:,size(test_toy,2));
% Fit the model and predict with test data
info = kdlorAlgorithm.fitpredict(train,test);

% You can evaluate performance with ordinal metrics:
addpath('src/Measures/')
CCR.calculateMetric(info.predictedTest,test.targets)
MAE.calculateMetric(info.predictedTest,test.targets)