Information for contributors

We would be very happy to receive contributions!

You don’t have to be a Matlab / Octave programmer. Useful code contributions are very much appreciated, but improved documentation, ideas on how our web site can be made prettier, or other ideas are also valued highly.

If you are not a Matlab / Octave programmer but would like to contribute or suggest improvements on the documentation or examples, please contact us directly.

For programmers, the preferred way to contribute is using git and github. If you would like to contribute code in another way, also please contact us directly.

Directory locations and naming conventions

Meta - naming conventions in the documentation

In what follows we use the following naming conventions. - Path names are Unix-like-based; / is the path separation character, rather than \ used on Windows platforms. - Directories have / as the last character, and are relative to the root directory where CoSMoMVPA resides. For example, if CoSMoMVPA is located in /Users/karen/git/CoSMoMVPA, then mvpa/ refers to /Users/karen/git/CoSMoMVPA/mvpa.

Setting the Matlab path

To use CoSMoMVPA functionality, it is recommended to set the path using cosmo set path. Optionally run savepath afterwards if you want to store the new path, so that it is set the next time Matlab or Octave is started.

Alternatively the path can be set manually as follows:

  • add the mvpa/ directory to the Matlab path.

  • also add externals/ and its subdirectories to the Matlab path.

  • do not add examples/ or tests to the Matlab path.
    • To run examples, cd to examples/ and run scripts from there.

    • To run unit tests, run cosmo_run_tests (this requires the xUnit or MOxUnit testing frameworks).

Organization of files and directories

  • Core CoSMoMVPA Matlab functions are in mvpa/. File names should match the pattern cosmo_*.m.

  • Runnable Matlab example scripts are in examples/. File names should match the pattern run_*.m or demo_*.m.

  • Unit tests are in tests/. File names should match the pattern test_*.m for unit tests, and any other prefix for helper functionality.

  • External libraries are in external.

  • Documentation is in doc/source/:
    • Documentation files have the .rst extension and are formatted as reStructuredText.

    • Exercises have the prefix ex_.

    • Other documentation files, unless automatically generated (see ‘build system’ below), should not have the prefix cosmo_ or run_, as running make clean in doc/ will remove them.

    • Other file types, such as images, are stored in doc/source/_static/.

    • Generated matlab output files, using the publish functionality in mvpa/cosmo_publish_run_scripts (for developers only), are stored in doc/source/_static/publish/.

  • Example data is stored separately.

Setting up the documentation build system

The documentation is built using Sphinx, Python, sphinxcontrib-matlabdomain, sphinxcontrib-bibtex and customly written Python code. Currently only Unix-like systems are supported; we have tested it on Linux and Mac OS. Python is a required dependency; we have tested with with version 2.7. Building requires using a shell, for example bash.

Installation can be done using easy_install. To install as root:

easy_install sphinx
easy_install -U sphinxcontrib-matlabdomain
easy_install -U sphinxcontrib-bibtex

Installation as non-root requires creating a local directory in which the Python packages are stored. To install these in ~/python-lib, for example:

cd
mkdir python-lib
cd python-lib
export PYTHONPATH=${PYTHONPATH}:`pwd`
export PATH=${PATH}:`pwd`
easy_install --install-dir . sphinx
easy_install --install-dir . -U sphinxcontrib-matlabdomain
easy_install --install-dir . -U sphinxcontrib-bibtex

In this case, it is useful to add the export-commands to ~/.bash_profile so that the paths are set automatically upon login.

Alternatively pip can be used:

pip install sphinxcontrib-matlab
pip install sphinxcontrib-matlab
pip install sphinxcontrib-matlabdomain
To build the documentation:
  • (optionally) in Matlab, cd to the mvpa/ directory, then run cosmo_publish_run_scripts. This generates the matlab output of the scripts in examples/.

  • On the terminal, cd to doc/, then run make html. (To clean previously built documentation, run make clean first).

Information for contributing programmers

The remainder of this section is for those experienced in programming; it describes how others can make code contributions, explains how and where different parts of the toolbox are stored, and provides some developer guidelines. If in doubt, or if you have questions or comments, do not hesitate to contact us.

Code development

Contributing using git

The git distributed version control system is used for code development. It has several advantages over just downloading the zip archive:
  • a distributed workflow: multiple people can work on the code simultaneously, and almost always their changes can be merged without conflicts. In the unlikely event of merge conflicts (when multiple people have changed the same code), these conflicts are resolved easly.

  • keeping track of individual contributions. Through git it is possible to see every change made, by anybody. It provides functionality similar to a time-machine, but with some kind of tagging: every change is annotated (see below). This allows anyone to see what was changed, when this happended, and by who.

  • code sharing on multiple computers: everyone has their own copy of the code, and can merge changes made by others.

  • maintaining multiple versions: through branching one can create multiple copies of the code, each which its own new features. This is very useful for new experimental features or bug-fixing without affecting the prestine master code. Once changes are considered ready for the master repository, they can be merged easily.

The instructions below assume a unix-like platform, and the command below should be run on the terminal (for example in bash).

Initial git / github setup

To get started with git and github to allow for code contributions to CoSMoMVPA:

  • set up a working installation of git (see installing git).

  • tell git about your name and email address:

    git config --global user.name "Your full name"
    git config --global user.email "your_email@the_domain.com"
    

    By setting these options, all commits you make will have this information, so that everybody can identify who changed what.

  • make an account on github, if you have not done so.

  • on the github project page, fork the repository, and follow the instructions there.

  • get a local copy of your forked repository: run:

    git clone https://github.com/karen/CoSMoMVPA.git
    

    if karen is your github user name.

  • change to the directory just created:

    cd CoSMoMVPA
    
  • tell git about the offical release, which we call upstream:

    git remote add upstream https://github.com/CoSMoMVPA/CoSMoMVPA.git
    

Proposing a change in the code (a.k.a. submitting a Pull Request (PR)

  • to update your repository to the latest official code, first make sure you are on the master branch, then pull the current code:

    git checkout master
    git pull upstream master
    

    (This assumes that added the remote upstream as described in the previous section.)

  • to add a new feature or provide a bugfix, start a new branch:

    git checkout -b my_awesome_new_feature
    
  • make the desired changes, then commit them. See below for details.

  • push these changes to your github account:

    git push origin my_awesome_new_feature
    
  • go to your own github page, i.e. https://github.com/karen/CoSMoMVPA if your git user name is karen. Typically the github webpage already mentions that a new branch was pushed, so just click on create pull request. Otherwise click pull requests in the right-hand bar, then click New pull request. Add a description to the pull request, and then submit it.

    We’ll get back to you to review and discuss the code. Once the code is ready for the official master it will be merged. You should receive notifications by email when the code is discussed or merged.

  • if you want go back to using code from the master branch (the official code), run:

    git checkout master
    

    Keep in mind that the master branch is supposed to contain working, runnable code. Proposed changes, including experimental code and bug-fixes, should preferably be submitted in separate branches.

There are many great resources on using git on the web; a detailed explanation is beyond the scope of this documentation.

Notes on committing

  • Please review your changes before commiting them. Useful commands are git status and git diff.

  • Do not use git -a; instead manually add the (changes to) files individually. Preferably commits should be atomic, that is change just one feature. For example if you changed a file at two places by (1) improving the documentation and (2) refactoring code used internally, then preferably you should make two commits. Using the tags below these could be DOC: ... and RF: .... - To add a new file my_new_file.m, run:

    git add my_new_file.m
    
    • To commit changes to a file, run:

      git add -i
      

      then press p (for ‘patch’), indicate which files to patch, and press y or n for each meaningful ‘atomic’ change, and q to quit. (Usually followed by git commit ...).

    • To view the history of previous commits, gitk is useful.

    • Use the following tags (inspired by PyMVPA) for commits:

      • ACK: Acknowledge someone else. Acknowledgees should be placed between # characters, so that acknowledgements can be generated automatically.

      • BF: Bugfix. Preferably this comes also with a unit test (i.e., BF+TST) that checks whether the bug was indeed fixed.

      • BK: Breaks existing functionality, or the signature of functions (changes in the number, or the meaning, of input and output arguments).

      • BLD: Changes in the build system.

      • BIG: Major change. Please use together with another tag.

      • CLN: Cleanup of code or documentation. SML can be omitted.

      • DOC: Change in documentation of matlab code (in examples/, mvpa/, tests/).

      • EXC: Change in exercises. This could go together with WEB or DOC, and/or RUN.

      • FT: (Fixed during testing) Fix bug caught by test before the code was merged into the master branch. These are issues that would be considered bugs if they had been merged into master before. Introduced February 2017.

      • MSC: Miscellaneous changes, not covered by any of the other tags.

      • NF: New feature or functionality.

      • OCTV: Change in GNU Octave compatibility.

      • OPT: Optimalization. A special case of refactoring. It should be used when the new code runs faster or uses less memory.

      • RF: Refactoring. To be used for changes in code that does not affect its external behaviour.

      • RUN: Change in runnable example scripts (in examples/).

      • SML: Minor change.

      • TST: Change in test functions (functions in tests/, or documentation tests).

      • WEB: Changes affecting web site content (either documentation in .rst files, or other files such as images).

      Using these tags:

      • allows others to quickly see what kind of changes were made

      • the web page build system to generate summary reports on the kind of changes automatically.

      • generate statistics on types of changes over type.

      Please describe what changes you made. The tags don’t have to name which files were changed, as git takes care of that.

      Tags can be combined, as it may occur that multiple tags apply; use the +-character to concatenate them.

      Examples:

      • git commit -m 'ENH: support two-dimensional cell arrays as feature attributes'

      • git commit -m 'RF: build a lookup table mapping all voxels to those in the dataset

      • git commit -m 'BF+TST: throw an error if partitions are not balanced; added unit test'

      • git commit -m 'DOC+SML: fixed a typo'

      • git commit -m 'BF+ACK: show error message when negative radius is provided. Thanks to #John Doe# and #Jane Doe# for bringing up this use case.

Build system

The build system is used to generate documentation for the web site (or local use).

  • Matlab files do not require building.

  • Documentation is built as follows:
    • all Matlab files in mvpa/ and examples/ are converted to reStructuredText format using the Python script matlab2rst.py in source/. This script generates three reStructuredText versions of all Matlab functions and example scripts:
      • Full contents (no suffix), containing the full source code.

      • Header contents (suffix _hdr), containing only the header (i.e. the first line and every subsequent line until the first line that does not start with % (note: line continuation, explained below, is currently not supported).

      • Skeleton contents (suffix _skl), containing a skeleton of the source code. In the skeleton version, lines between a starting line % >@@> and ending line % <@@< is replaced by a text saying %%%% Your code comes here %%%%. Skeleton files are intended for exercises.

    • Both .txt files (with the raw contents preceded by an include statement) and .rst files (with a title and label) are generated; the latter contain include statements to include the former.

    • the Makefile in source/, when used through make html, uses mat2rst.py to generate reStructuredText Matlab files and then uses Sphinx to convert these files to html.

  • The build.sh script builds the documentation and datasets.

Note: building the documentation, as described in the previous points, is currently supported on Unix-like systems only, and requires additional dependencies (see download).

  • The mvpa/cosmo_publish_run_scripts.m function generates the output from all the runnable examples in examples/ as html files and puts them in doc/source/_static/publish/. This function is used to produce output for the web site.

Matlab code guidelines

The following are guidelines, intended to improve:

  • consistency in code layout across contributers, so that the final result is more consistent.

  • readability, so that less time is spent in understanding how the code works or what it does.

  • performance, so that execution time or memory usage is reduced.

Note: None of these guidelines are set in stone. Try to use common sense when considering not to follow them. Indeed, for each guideline there may be a good reason to deviate from it.

Maximum line length is 75 characters

Try to keep line lengths limited to 75 characters, so that files can be viewed in a standard terminal window, possibly with scroll bars, without line breaks. (The Matlab editor shows a vertical line after the 75-th character). Use line continuation (... at the very end of the line) followed by indentation for the continued lines.

  • An exception to this rule is including URLs in documentation, because they allow for copy-pasting the URL directly.

  • If a binary operator is used together with line continuation, put the operator before the line continuation.

  • Lines should not end with whitespace.

    bad:

    my_output_data=my_awesome_function(first_argument, second_argument, third_argument, fourth_argument, fifth_argument, sixth_argument);
    
    apply_mask=user_has_supplied_mask ...
                    && ~isempty(user_mask) ...
                    && sum(user_mask)>0
    
    my_string='Nick was right (as he usually is) in stating that although the use of long expressions may sometimes seem unavoidable, the use of line continuations hardly ever is.';
    

    good:

    my_output_data=my_awesome_function(first_argument, second_argument,...
                                       third_argument, fourth_argument,...
                                       fifth_argument, sixth_argument);
    
    apply_mask=user_has_supplied_mask && ...
                    ~isempty(user_mask) && ...
                    sum(user_mask)>0
    
    my_string=['Nick was right (as he usually is) in stating that '...
                'although the use of long expressions may sometimes '...
                'seem unavoidable, the use of line continuations '...
                'hardly ever is.'];
    

    (Yes, we break this rule occasionely.)

Indentation is 4 spaces (no tabs)

Indentation should be used for if-else-end, while and function blocks. Expressions of the form if expr, else, elseif expr, var=function(var), while expr, and end should be on a single line, except for very short statements that either set a default value for an input argument or raise an exception. For consitency please use exactly four spaces for every indent level; no more, no less.

If this guideline and the previous one do not give you enough room to express yourself, then most likely you are overcomplicating things; consider rewriting the code and/or use subfunctions.

bad:

if a>0
if b>0, c=1;
        else
    d=1; end
end
function r=my_min3(x,y,z)
  if x<y
    if z<x
      r=z;
    else
      r=x;
    end
  else
    if z<y
      r=z;
    else
      r=y;
    end
 end
if my_awesome_function(some_input, more_input)>100, the_array_value(:)=42; else other_array_value(:)=31; end

acceptable:

if nargin<2 || isempty(more_input), more_input=42; end
if nsamples~=ntrain, error('Size mismatch: %d ~= %d', nsamples, ntrain); end

good:

if my_awesome_function(some_input, more_input)>100
    the_array_value(:)=42;
else
    other_array_value(:)=31;
end
if a>0
    if b>0
        c=1;
    else
        d=1;
    end
end
function r=my_min3(x,y,z)
    if x<y
        if z<x
            r=z;
        else
            r=x;
        end
    else
        if z<y
            r=z;
        else
            r=y;
        end
    end

Use lower-case letters for variable names

Use underscores (_) to separate words.

bad:

myVar=0;
MyVar=0;
My_Var=0;

good:

my_var=0;

Throw an (informative) error early

Throw an error as soon as something seems out of order. When doing so, try to provide an informative error message.

bad:

error('What do you mean?');

This is bad because the user has no idea why an error was thrown.

if ntemplate~=nsamples
    % this is bad because the friggofrag analysis is invalid.
    % Telling the user that they provided wrong input could harm their
    % self-esteem however, so let's just make up some data that,
    % although completely meaningless, will ensure that the script
    % does not crash.
    samples=randn(ntemplate);
end

This is very bad because instead of reporting that data was of incorrect shape, the code generates new (random) data, which the user most likely neither expects or desires.

good:

error('targets have size %d x %d, expected %d % d', ...
             target_size, expected_target_size);
if strcmp(caught_exception.identifier,...
                    'stats:svmtrain:NoConvergence');
    error(['SVM training did not converge. Your options are:\n'...
           ' 1) increase ''boxconstraint''\n'...
           ' 2) increase ''tolkkt''\n'...
           ' 3) set ''kktviolationlevel'' to a positive value\n'...
           ' 4) use a different classifier\n'...
           'If you do not have a strong preference for '...
           'either option, you are advised to try option (4) '...
           'using cosmo_classify_lda'],'');
else
    rethrow(caught_exception);
end

Do not repeat yourself

If the same expression is evaluated multiple times, evaluate it once and assign its result to a variable.

bad:

if nfeatures>nsamples
    delta=nfeatures-nsamples
else
    delta=0;
end

aggregate_size=[nfeatures,nsamples,delta];

if nsamples<=nfeatures
    cosmo_warning('%d samples < %d features', nsamples, nfeatures);
end

good

has_more_features_than_samples=nfeatures>nsamples

if has_more_features_than_samples
    delta=nfeatures-nsamples
else
    delta=0;
end

aggregate_size=[nfeatures,nsamples,delta];

if has_more_features_than_samples
    cosmo_warning('%d samples < %d features', nsamples, nfeatures);
end

Write in normal, understandable english

Avoid using capital letters in the documentation, unless you want others to PERCEIVE YOUR MESSAGE AS SHOUTING, normal spelling dictates this (start of a sentence, proper names), tag code, or to refer to variable names. Avoid capital letters for variable names. If possible, give informative error messages.

bad:

% NOW RUN THE CROSS-VALIDATION
error('YOU SERIOUSLY MESSED UP THE INPUT - ARE YOU CRAZY???');
fprintf('STARTING ANALYSIS... PLEASE BE PATIENT!!!\n');
error('What do you mean?');

acceptable:

% Note: this function is EXPERIMENTAL.

good:

% Note: this function is *** experimental ***.
% The function Y=abs(X), where X is NxQ, returns an array Y of size NxQ
% so that, assuming that all elements in X are real, Y==X.*sign(X);
% NNO Sep 2013
% TODO: support more than two different chunks.
% Now run the cross-validation
if show_progress
    fprintf('Starting analysis - please be patient...\n');
end

Document functions

When writing function definitions:
  • start with a short sentence (one line) describing its purpose.

  • describe the signature of the function (input and output arguments)

  • describe the input parameters

  • describe the output parameters

  • whenever appropriate, add examples, notes, and references.

bad:

function [winners,classes]=cosmo_winner_indices(pred)
% uses a pretty cool hack using bsxfun to decide about the winners!

good:

function [winners,classes]=cosmo_winner_indices(pred)
% Given multiple predictions, get indices that were predicted most often.
%
% [winners,classes]=cosmo_winner_indices(pred)
%
% Input:
%   pred              PxQ prediction values for Q features and P
%                     predictions per feature. Values of NaN are ignored,
%                     i.e. can never be a winner.
%
% Output:
%   winners           Px1 indices of classes that occur most often.
%                     winners(k)==w means that no value in
%                     classes(pred(k,:)) occurs more often than classes(w).
%   classes           The sorted list of unique predicted values, across
%                     all non-ignored (non-NaN) values in pred.
%
% Examples:
%     % a single prediction, with the third one missing
%     pred=[4; 4; NaN; 5];
%     [p, c]=cosmo_winner_indices(pred);
%     p'
%     %|| [1 1 NaN 2]
%     c'
%     %|| [4, 5]
%
%     % one prediction per fold (e.g. using cosmo_nfold_partitioner)
%     pred=[4 NaN NaN; 6 NaN NaN; NaN 3 NaN; NaN NaN NaN; NaN NaN 3];
%     [p, c]=cosmo_winner_indices(pred);
%     p'
%     %|| [2, 3, 1, NaN, 1]
%     c'
%     %|| [3 4 6]
%
%     % given up to three predictions each for eight samples, compute
%     % which predictions occur most often. NaNs are ignored.
%     pred=[4 4 4;4 5 6;6 5 4;5 6 4;4 5 6; NaN NaN NaN; 6 0 0;0 0 NaN];
%     [p, c]=cosmo_winner_indices(pred);
%     p'
%     %|| [2, 3, 4, 2, 3, NaN, 1, 1]
%     c'
%     %|| [0, 4, 5, 6]
%
% Notes:
% - The typical use case is combining results from multiple classification
%   predictions, such as in binary support vector machines (SVMs) and
%   cosmo_crossvalidate
% - The current implementation selects a winner pseudo-randomly (but
%   deterministically) and (presumably) unbiased in case of a tie between
%   multiple winners. That is, using the present implementation, repeatedly
%   calling this function with identical input yields identical output,
%   but unbiased with respect to which class is the 'winner' sample-wise.
% - Samples with no predictions are assigned a value of NaN.
%
% See also: cosmo_classify_matlabsvm, cosmo_crossvalidate
%
% #   For CoSMoMVPA's copyright information and license terms,   #
% #   see the COPYING file distributed with CoSMoMVPA.           #

Pre-allocate space for data

  • Allocate space for output or intermediate results beforehand, rather than let arrays grow in a for or while loop.
    • This can greatly improve performance. Growing an array requires reallocating memory, which slows down code execution.

    • It also indicates what the size of the output is, which can help in understanding what code does.

    • This guideline is especially important when large arrays of data are used.

    bad:

    ndata=size(data,1);
    accs=[]; % start with empty array, then let it grow
    for k=1:ndata;
        acc=a_func(data(k,:)) % compute accuracy
        accs=[acss acc];
    end
    

    good:

    ndata=size(data,1);
    accs=zeros(1,ndata); % allocate space for output
    for k=1:ndata
        acc=a_func(data(k,:)) % compute accuracy
        accs(k)=acc;
    end
    

Use vectorization

When possible use vectorization rather than a for or while loop.
  • Many functions support vectorized functions, where the same function is applied to elements in arrays.

  • Vectorization reduces the number of lines of code.

  • Vectorization typically reduces execution time.

really bad: (see previous guideline)

[nrows,ncols]=size(data);

abs_data=[]; % start with empty array, then let it grow
for k=1:nrows
    row_abs_data=[]; % absolute data for the k-th row
    for j=1:ncols
        row_abs_data=[row_abs_data, abs(data(k,j))];
    end
    abs_data=[abs_data; row_abs_data]; % add this row to the output
end

bad:

[nrows,ncols]=size(data);

abs_data=zeros(nrows,ncols); % allocate space for output

% compute absolute value for each value in data
for k=1:nrows
    for j=1:ncols
        abs_data(k,j)=abs(data(k,j));
    end
end

good:

abs_data=abs(data);

Use clear variable names

The aim is to find a good balance between length and readability. Short variable names are fine if their use is clear (e.g., i, j, k for loop variables; n for number of elements, f for a function). It is recommended to document what a statement does if this cannot be deduced easily from the variable/function names.

Note: i and j are used in Matlab to indicate the imagery unit (for which it holds that i^2==-1), but for functions that do not use complex numbers (currently all of them) their use as a loop variable is acceptable.

bad:

msxs = find(sm)
hkrw8ingmuch = max([v,vv,vvv])
my_very_long_variable_name_that_describes_something_i_forgot = ...
    apply_function_work(with_a_very_long_argument_name,...
                        and_another_long_argument_name);

borderline acceptable:

% get the indices of the sample mask
msxs = find(sm)
[ns, nf]=size(ds.samples);

good:

mask_indices=find(sample_mask);
max_dimen=max([x_dim, y_dim, z_dim]);
sliced_ds=cosmo_dataset(ds, mask_indices);
n=size(data,1); % number of samples
for k=1:n
    data_result(k)=f(data(k,:));
end
[nsamples, nfeatures]=size(ds.samples);

Avoid side effects

Generally try to avoid side effects, and if that is not possible, indicate such effects clearly in the function name.

very bad:

function init_my_toolbox()

    restoredefaultpath();
    addpath('my_functions');

The above is bad because:

  • The function name init_my_toolbox does something one would not expect based on its name, namely it resets the Matlab path.

  • Functions that were accessible before are not longer in the Matlab path. In particular, any other external toolboxes or code not part of the Matlab installation becomes unavailable.

It is acceptable to add something to the Matlab path, if the function name clearly indicates that it does so:

acceptable:

function my_toolbox_set_path()

    addpath('my_functions');

Use onCleanup if an earlier state needs to be reset

In general, functions should not change the current working directory, the path, or the warning state. Sometimes this cannot be avoided, but in that case these changes should be undone when leaving the function using onCleanup. try - catch constructions should be avoided.

bad:

function do_computation()

    addpath('my_functions');
    original_dir=pwd();
    cd('other_functions/private');
    warning('off');

    do_stuff();

    rmpath('my_functions');
    cd(original_dir);
    warning('on');

The above is bad, because:

  • the user may have added my_functions to the path themselves; after calling this function, it is removed from the path.

  • the user may have set the warning state themselves to off; this is undone after calling this funciton

  • the current working directory and the path are not restored when execution is interrupted because of an error or a user interrupt (ctrl+C).

bad:

function do_computation()

    original_path=path();
    original_dir=pwd();

    try
        addpath('my_functions');
        cd('other_functions/private');

        do_stuff();

    catch
        e=lasterror();

        cd(original_path);
        path(original_path);

        rethrow(e);
    end

    cd(original_path);
    path(original_path);

Although the path and pwd are restored, this is bad because:

  • it introduces code duplication.

  • it does not protect against the user pressing ctrl+c.

good:

function do_computation()

   original_path=path();
   original_working_dir=pwd();
   original_warning_state=warning();

   path_resetter=onCleanup(@()cd(original_path));
   working_dir_resetter=onCleanup(@()path(original_working_dir));
   warning_state_resetter=onCleanup(@()warning(original_warning_state));

   addpath('my_functions');
   warning('off');

   cd('other_functions/private');

   do_stuff();

   % (the path, working directory, and warning state are reset when
   %  execution of the code in the function body is completed or
   %  interrupted.)

Tests should not require user interaction

When implementing unit tests (in the tests) directory, functions should run automatically without any user interaction. If a test were to require user interaction, one of the main advantages of the test suite (fully automated testing) is lost.

Do not use global variables

Global variables can have nasty and unpredictable side effects. Therefore, CoSMoMVPA does not use global variables. In almost all cases it is preferable that output of a function should depend on the input only; there are some exceptions, such as cosmo warning which by default shows each warning only once. If necessary (e.g. for caching), use persistent variables.

Avoid long and complicated expressions

  • Avoid long expressions with many nested parentheses; rather use multiple lines in which variables (with informative names) are assigned in succession. Although this carries a minor speed penalty in Matlab, it improves readability.

    borderline acceptable:

    for j=1:npartitions
        test_indices{j}=find(chunk_idx2count(combis(j,:)));
    end
    

    good:

    for j=1:npartitions
        combi=combis(j,:);
        sample_count=chunk_idx2count(combi);
        test_indices{j}=find(sample_count);
    end
    

Use sprintf or fprint when formatting strings

  • When formatting strings use sprintf or fprintf, rather than num2str and string concatenation. Avoid using disp when printing strings; use fprintf instead.

    bad:

    disp(['Accuracy for ' label ' is ' num2str(mean_value) ' +/-' ...
                num2str(std_value)]);
    

    good:

    fprintf('Accuracy for %s is %.3f +/- %.3f\n', ...
                        label, mean_value, std_value);
    

    Note: newer Matlab versions provide strjoin, but for compatibility reasons with older versions, an alternative implementation is provided as cosmo_strjoin.

Avoid using eval

Statements with eval can obfuscate the code considerably, and also make refactoring (such as changing variable names) more difficult. In almost all cases code can rewritten that avoids eval. If necessary use function handles.

very bad:

% for even samples apply f_even, for odd ones f_odd
results=[];
for k=1:nsamples
    if mod(k,2)==0
        eval(['results=[results; f_even(data(' num2str(k) '))];']);
    else
        eval(['results=[results; f_odd(data(' num2str(k) '))];']);
    end
end

This is bad because it uses ``eval`` and does not pre-allocate space for data.

bad:

% for even samples apply f_even, for odd ones f_odd
results=zeros(nsamples,1);
f_names={'f_odd','f_even'};
for k=1:nsamples
    f_index=mod(k+1,2)+1);
    f_name=f_names{f_index};
    eval(sprintf('results(%d)=%s(data(%d));', k, f_name, k));
end

This is bad because it uses ``eval``.

good:

% for even samples apply f_even, for odd ones f_odd
results=zeros(nsamples,1);
f_handles={@f_odd, @f_even};
for k=1:nsamples
    f_index=mod(k+1,2)+1;
    f_handle=f_handles{f_index};
    f_data=f_handle(data(k));
end

Minimize using try and catch

The use try and catch statements is generally avoided; we aim to throw an exception when the input to a function is wrong. Consider that code for use in a Mars rover should never crash even in unexcepted circumstances, whereas in CoSMoMVPA the code is aimed at analysis of neuroscience data, where getting correct results is very important (and knowing that something is wrong is important too). Some current exceptions are:

  • cosmo publish run scripts, that builds the Matlab output from the scripts in examples/. We don’t want that function to crash if any of the scripts it is publishing crashes.

  • cosmo classify libsvm and cosmo classify matlabsvm, that check whether the required externals are present if they fail, as that is a likely scenarion. In that case, even though the error is caught initially, always a subsequent error is thrown.

  • cosmo searchlight, which if an error is thrown by the measure function handle, prefixes the error message with the feature id that caused the error, and then throws a new error.

No private functions

Do not use private functions. If functions are considered useful enough for one particular function, it may also be useful for other functions or for the users.

No file duplication

Do not duplicate files in multiple locations. Doing so would add additional maintenance burden in keeping all files up to date.

Check input arguments

Generally it is good to check the input arguments, although there is a subjective component in deciding how much should be checked, or when an error should be thrown. Checking more means less concise code and longer execution times, but can also prevent the user from making mistakes that would otherwise go undetected.

CoSMoMVPA-specific guidelines

Writing exercises

To indicate that a code block is an exercise, place a line containing % >@@> before the block and one containing % <@@< after the block. When using the build system (see above), this will replace the corresponding block by a message saying %%%% Your code comes here %%%% in the online documentation.

example:

% set the training and test indices for each chunk
for k=1:nchunks
    % >@@>
    test_msk=unq(k)==chunks;
    train_indices{k}=find(~test_msk)';
    test_indices{k}=find(test_msk)';
    % <@@<
end

Documentation tests

When providing examples it is a good idea to write them in the shape of examples, so that running cosmo run tests will actually test whether the code runs as advertised. This uses the MOdox framework. Many modules have such doctests; you can spot them in the Examples: section of the help info, where the expected output is preceded by %||. For example:

function [split, nsplit]=cosmo_strsplit(string, delim, varargin)
% splits a string based on another delimeter string
%
% [split,n]=cosmo_strsplit(string[,delim,][pos1, delim2, pos2, delim2,...])
%
% Inputs:
%   string          input to be split
%   delim           delimiter string. delim can contain backslash-escaped
%                   characters that are interpreted by sprintf; for
%                   example '\t', '\n' and '\\' represent a tab, newline
%                   and backslash character, respectively.
%                   If omitted or equal to [], then the string is split
%                   based on whitespaces occuring in string
%
%   pos             (optional) a single index indicating which split part
%                   should be returned. If string is split in N elements,
%                   then a negative value for pos is equivalent to pos+1+N
%                   (similar to Python). For example, pos=-1 means that the
%                   last element is returned, and pos=-2 means that the
%                   element before the last element is returned.
%                   If omitted a cell with all parts are returned.
%   delim*          (optional) subsequent delimeters applied after applying
%                   pos. It requires that the preceding pos has a single
%                   value.
%
% Output:
%   split           when the last argument is non-positional, split is a
%                   cell with the string split by delim. When there are N
%                   non-overlapping occurences of delim in string, then
%                   split has N+1 elements, and the string
%                   [delim split{1} delim split{2} ... split{N} delim]
%                   is equal to string.
%                   If the last argument is positional, then split is
%                   a string with value split_{pos} where split_ is the
%                   result if pos where not the last arugment.
%   n               the number of elements in split, if split is a cell;
%                   0 otherwise
%
% Examples:
%   % split by '*'
%   cosmo_strsplit('A*AbbAbA*AbA*A*Ab','*')
%   %||      { 'A'    'AbbAbA'    'AbA'    'A'    'Ab' }
%
%   % split by 'A*A'
%   cosmo_strsplit('A*AbbAbA*AbA*A*Ab','A*A')
%   %||  { ''    'bbAb'    'b'    '*Ab' }
%
%   % take second element after split
%   cosmo_strsplit('A*AbbAbA*AbA*A*Ab','A*A',2)
%   %||  'bbAb'
%
%   % get last element after split
%   cosmo_strsplit('A*AbbAbA*AbA*A*Ab','A*A',-1)
%   %||  '*Ab'
%
%   % split twice, first on 'A*A', take second element, then on 'A'
%   cosmo_strsplit('A*AbbAbA*AbA*A*Ab','A*A',2,'A')
%   %||    { 'bb'    'b' }
%
%   % take first element after second split
%   cosmo_strsplit('A*AbbAbA*AbA*A*Ab','A*A',2,'A',1)
%   %||  'bb'
%
%   % illustrate effect of not using a delimiter string
%   % (which causes the string to be split by whitespace) and using
%   % a space as delimiter
%   cosmo_strsplit(' CoSMoMVPA makes live...  easy!')
%   %||   { 'CoSMoMVPA'    'makes'    'live...'    'easy!' }
%   cosmo_strsplit(' CoSMoMVPA makes live...  easy!',' ')
%   %||   {  ''    'CoSMoMVPA'    'makes'    'live...'    ''    'easy!' }
%
% #   For CoSMoMVPA's copyright information and license terms,   #
% #   see the COPYING file distributed with CoSMoMVPA.           #

Compatibility notes

CoSMoMVPA aims to be compatible with GNU Octave 3.8 and later, and with Matlab versions from at least 2010b onwards.

Test suite

CoSMoMVPA uses a test suite, which can automatically test most of the code. This helps in maintaining or improving the quality of the code, and to check whether refactoring code does not introduce undesired changes in behaviour (bugs). Tests are located in tests/ and use the MOxUnit framework. To run them, either:

  • run cosmo run tests: when using xUnit.

  • run moxunit_run_tests in the tests directory: when using MOxUnit. Supported on the Matlab and Octave platform. Documentation tests are not supported yet.

Currently we use Travis-ci for continuous integration testing and coveralls.io with MOcov for testing coverage. If you have a github account and a CoSMoMVPA fork, you can also use it to test new branches. To do so:

  • Make an account on Travis-ci.

  • Link it to your github account.

  • Now, after every ‘push’ to github, the test suite is run automatically using moxunit_run_testa on Octave.

  • If any tests fails, or passes if it failed before, you will be notified by email.

When proposing new functionality through a pull request, please include tests that test this functionality.

Good tests

Some general features that make tests good:

  • expected function mapping: whether output is expected for particular input. It is even better if input is generated randomly and properties of the expected output are tested.

  • code coverage: check whether different input arguments (for example different switches or options) results in different expected outputs. Code coverage estimates from coveralls.io or MOcov can be useful in guiding which code needs more testing. Although code coverage is not a goal in itself, as a rule of thumb we like 80% coverage and are very happy with coverage above 90%.

  • raise exceptions when expcted: test whether wrong or illegal inputs raise an exception (through assertExceptionThrown).

  • modular: test one particular function (for unit tests), or whether multiple functions work together as expected (integration test)

  • no data dependencies: tests that use data as input should generate this data themselves. The cosmo synthetic dataset function can generate data in various modalities and parameters. We aim to not include binary files for testing. For example, some of our input and output tests actually generate files in AFNI, NIFTI, GIFTI and BrainVoyager format.

We also like fast tests, because it allows running all tests, or a significant subset, in a short amount of time. This gives quick feedback to the developer whether code works or not, allowing for data-driven development techniques. It also scales well: imagine running 500 tests and each test takes 5 minutes: the suite would take more than a day and a half to run! But even half a minute per test would make take 4 hours, which can be used to drink lots of cups of coffee between implementing a change and getting feedback whether the suite passes.

What is fast then? Running on the Matlab platform (which typically runs a bit faster than Octave), here are some rule of thumb for the runtime t of a test:

  • t < 100 ms: excellent.

  • 100 ms < t < 1 s: ok.

  • 1 s < t < 5 s: borderline acceptable but not preferred.

  • 5 s < t < 60 s: acceptable only in exceptional cases.

  • t > 60 s: generally not acceptable.

One important principle in writing fast tests is using small data, as this generally reduces computation time for most functions. To estimate the runtime for tests, you can use moxunit_runtests with the verbose option.

For examples of tests, see the test directory.