Archive for March, 2010

Plot LimInclude properties

Wednesday, March 31st, 2010

Concluding my three-part mini-series on hidden and undocumented plot/axes properties, I would like to present a set of properties that I find very useful in dynamic plots: XLimInclude, YLimInclude, ZLimInclude, ALimInclude and CLimInclude. These properties, which are relevant for plot/axes objects, have an ‘on’ value by default. When set to ‘off’, they exclude their object from the automatic computation of the corresponding axes limits (XLim/YLim/ZLim/ALim/CLim).

For example, here’s a simple sine wave with a wavefront line marker. Note how the too-tall wavefront line affects the entire axes Y-limits:

cla;
t=0:.01:7.5;
plot(t,sin(t));
line('xdata',[7.5,7.5], 'ydata',[-5,5], 'color','r'); 
box off

Regular plot (YLimInclude on)

Regular plot (YLimInclude on)

This situation is quickly fixed using the YLimInclude property:

cla;
t=0:.01:7.5;
plot(t,sin(t));
line('xdata',[7.5,7.5], 'ydata',[-5,5], 'color','r', ...
     'YLimInclude','off'); 
box off

YLimInclude off

YLimInclude off

Beside the functional importance of this feature, it also has a large potential for improved application performance: I recently designed a monitor-like GUI for a medical application, where the data is constantly updated from an external sensor connected to the computer. The GUI presents the latest 10 seconds of monitored data, which bounce up and down the chart. A red wave-front line is presented and constantly updated, to indicate the current data position. Since the monitored data jumps up and down, the Y-limits of the monitor chart often changes, and with it I would need to modify the wavefront’s YData based on the updated axes YLim. This turned out to steal precious CPU time from the actual monitoring application. Came YLimInclude to the rescue, by letting me specify the wavefront line as:

hWavefront = line(..., 'YData',[-99,99], 'YLimInclude','off');

Now the wavefront line never needed to update its YData (only XData, which is much less CPU-intensive) – it always spanned the entire axes height, since [-99,99] were assured (in my particular case) to exceed the actual monitored data. This looked better (no flicker effects) and performed faster than the regular (documented) approach.

Note that although all these properties exist, to the best of my knowledge, for all Handle-Graphic plot objects, they are sometimes meaningless. For example, ZLimInclude is irrelevant for a 2D patchless plot; CLimInclude relates to the axes color limits which are irrelevant if you’re not using a colormap or something similar; ALimInclude relates to patch transparency (alpha-channel) and is irrelevant elsewhere. In these and similar cases, setting these properties, while allowed and harmless, will simply have no effect.

This concludes my mini-series of undocumented plot/axes properties. To recap, the other articles dealt with the LooseInset and LineSmoothing properties.

Have you found other similar properties or use-cases that you find useful? I will be most interested to read about them in the comments section below.

Axes LooseInset property

Wednesday, March 24th, 2010

Last week, I wrote an article about the hidden/undocumented LineSmoothing plot property. This week, I want to introduce another useful hidden/undocumented property – the plot axes’ LooseInset property. This follows on the wake of an email I received from a reader about this property, which had some new information for me (thanks Ben!).

Apparently, LooseInset, which is automatically set to a factory value of [0.13, 0.11, 0.095, 0.075], is used by Matlab axes to reserve a small empty margin around the axes, presumably to enable space for tick marks. These empty margins can be very annoying at times, especially when we have directly control on the axes contents.

figure; t=0:0.01:7; plot(t,2*sin(t));

Axes with default LooseInset values (note the excessive margins)

Axes with default LooseInset values
(note the excessive margins)

If you set Position to [0 0 1 1], the labels are cut-off; if you set Position to something like [0.05 0.05 0.9 0.9], you can get the labels to show up, but if you now resize the image the labels may be cut off… Similarly, setting TightInset also does not work.

Theoretically, the solution should be to set OuterPosition to [0 0 1 1]. This is supposed to make the axes (including labels) take up the entire figure. However, it usually over-estimates the required margins, causing wasted space. Using OuterPosition also causes unexpected behaviors with sub-plots.

Solution: simply set LooseInset to [0 0 0 0]:

set(gca, 'LooseInset', [0,0,0,0]);

Axes with empty LooseInset values

Axes with empty LooseInset values

To modify all future axes in the same way (i.e., have an empty LooseInset):

set(0,'DefaultAxesLooseInset',[0,0,0,0])

Clearing the LooseInset margins has a drawback: if the axes is zoomed or modified in such a way that the labels change, then the active axes plot region needs to shrink accordingly. For example:

Axes with empty LooseInset values, wide tick labels (note the changed plot region size)

Axes with empty LooseInset values, wide tick labels
(note the changed plot region size)

When determining the size of the axes, it seems that Matlab takes into account larger of the documented TightInset and the undocumented LooseInset. So, perhaps a better generic solution would be the one suggested by another blog reader:

set(gca,'LooseInset',get(gca,'TightInset'))

Note that the LooseInset property was first reported on CSSM back in 2007 (also here). The LooseInset property has remained hidden and undocumented to this day (Matlab 7.10, R2010a), although it has even featured in an official MathWorks Technical Solution to a reported problem about unexpected axes sizes last year.

p.s. – another undocumented property of Matlab axes, ContentsVisible, was described by Matt Whittaker in a comment on my original article that introduced undocumented properties.

Plot LineSmoothing property

Wednesday, March 17th, 2010

I have already written about several undocumented hidden properties in the past. Today, I would like to introduce one of my favorites: the plot-line’s LineSmoothing property. Compare the following two outputs:

% Standard (non-smoothed) plot-line
plot(1:5,2:6,'o-', 'LineWidth',1);

Standard plot line (not anti-aliased)

Standard plot line (not anti-aliased)

% Smoothed (anti-aliased) plot line
plot(1:5,2:6,'o-', 'LineWidth',1, 'LineSmoothing','on');

Anti-aliased (smoothed) plot line

Anti-aliased (smoothed) plot line

Line smoothing (aka anti-aliasing) works by inserting semi-transparent pixels at the edges of the actual plot line, thereby giving an optical illusion of a smooth line without pixelization effects. In Matlab, antialiasing is done automatically for fonts, but unfortunately not for plot lines that have non-default line-widths.

Line-smoothing has been around for a long time. It was even mentioned in a user comment on the official MathWorks Pick-of-the-Week article that introduced the excellent Myaa utility (Myaa uses applicative Matlab code to create anti-aliased plot effects).

However, to this day (R2010a), the LineSmoothing property remains hidden, undocumented and not officially supported.

Perhaps the reason for this is the following bug: text objects that cross a smoothed line are obscured by it. Depending on the text size and the line width, the text might be completely hidden, although its handle indicates that it is visible and despite it being created after the plot line!

plot(1:5,2:6,'.-', 'LineWidth',5, 'LineSmoothing','on');
text(2.2,3.5, 'abcd','Color','r');

Smoothed line obscuring text

Smoothed line obscuring text

Note that this does not happen for standard (non-smoothed) lines:

plot(1:5,2:6,'.-', 'LineWidth',5);
text(2.2,3.5, 'abcd','Color','r');

Non-smoothed line does NOT obscure text

Non-smoothed line does NOT obscure text

Luckily, there’s a very simple workaround for this, that allows both line-smoothing and non-obstruction: simply set the text’s z-position to a higher value than the plot line’s. In our example, we used a simple 2D plot line (i.e., z-position = 0), so let’s set z-position=1 for our text label:

plot(1:5,2:6,'.-', 'LineWidth',5, 'LineSmoothing','on');
text(2.2,3.5,1, 'abcd','Color','r');

Smoothed line not obscuring text

Smoothed line not obscuring text

One final note: the LineSmoothing property also exists for line, patch, surf, mesh and other similar objects.

Do you use have any other favorite undocumented/hidden property? If so, please share it in a comment below.

Matlab and the Event Dispatch Thread (EDT)

Wednesday, March 10th, 2010

Once again I welcome guest blogger Matt Whitaker, with the long awaited EDT article.

Java Swing’s Event Dispatch Thread (EDT)
or: why does my GUI foul up?

Matlab for the most part is a single threaded environment. That is, all commands are executed sequentially along a single execution thread. The main exception to this are the Handle Graphics (GUI) components whose operations execute on the Java Event Dispatch Thread (EDT). EDT effects are reflected even in mundane Matlab GUI operations.

If we execute the code below we will probably see nothing until the loop completes and the figure appears with the text label showing ‘10000′:

h = figure;
txt = uicontrol('Parent',h, 'Style','text', 'String','1');
for n = 1:10000
    set(txt,'String',int2str(n))
end %for

By adding a couple of drawnow commands we get the figure and text label to render and then we see the count progress to 10000.

h = figure;
txt = uicontrol('Parent',h, 'Style','text', 'String','1');
drawnow;
for n = 1:10000
    set(txt,'String',int2str(n));
    drawnow;
end %for

The drawnow function allows the EDT queue to be flushed and the pending graphics operations to be evaluated. This will also happen with pause and several other commands.

If we want to use Swing (or AWT) components in our user interfaces we need to take this multi-threaded environment into account. The Swing toolkit designers decided to make all the Swing components thread un-safe in order to decrease their complexity. As a consequence, all access to Swing components should be done from the event dispatch thread (EDT), to ensure the operations are executed sequentially, at the exact order in which they were dispatched. Any action on a Swing component done on another thread (Matlab’s main processing thread in our case) risks a race-condition or deadlock with the EDT, which could (and often does) result in weird, non-deterministic and non-repetitive behavior – all of which should be avoided in any application which should behave in a precisely deterministic manner.

In Java, the usual pattern to accomplish EDT dispatching is to create a Runnable object, encapsulate the GUI code in the run method of the Runnable object, then pass the Runnable object to the static EventQueue.invokeLater (or EventQueue.invokeAndWait if we need to block operations to get a return value) method.

Runnable runnable = new Runnable()
{
    public void run()
    {
        //GUI Code here
    }
}
 
EventQueue.invokeLater(runnable);

There are several functions in Matlab that implement this programming pattern for us: javaObjectEDT, javaMethodEDT, awtinvoke, awtcreate and javacomponent. JavaMethodEDT and javaObjectEDT were introduced in version R2008b (7.7) and are minimally and only partially documented although they have reasonably complete help comments. The other three are semi-documented (meaning they are unsupported but if you edit or type their m-file you’ll see a fairly detailed help section), and although there is some overlap in their functionality they are still available.

javaObjectEDT and javaMethodEDT

javaObjectEDT is the the preferred method since R2008b of creating swing components to be used on the EDT. An object created with javaObjectEDT will have all of its subsequent method calls run on the EDT. This is termed Auto Delegation. Auto-delegation greatly simplifies and increases the readability of code. Note that objects created as a result of method calls may not be implemented on the EDT.

If you have an existing Java object, you can pass it to javaObjectEDT at any time – all its subsequent calls will then onward run on the EDT. Note that this useful functionality is an under-documented javaObjectEDT feature: it is not mentioned in the main help section but only implied from the example.

% Create a button on the EDT
btn = javaObjectEDT('javax.swing.JButton');
% this will run on EDT since btn was javaObjectEDT-created
btn.setText('Button');
 
% Create a button NOT on the EDT
btn2 = javax.swing.JButton;
% Dangerous! call will run on main Matlab thread
btn2.setText('Button2');
% modify btn2 so its methods will start running on the EDT
javaObjectEDT(btn2);
btn2.setText('Button2');

The following example shows the use of javaObjectEDT and javaMethodEDT in a more complex situation using a JTable:

function tableExample
hFig = figure;
drawnow; %need to get figure rendered
 
%use Yair's createTable to add a javax.swing.JTable
%http://www.mathworks.com/matlabcentral/fileexchange/14225-java-based-data-table
%wrap ceateTable in javaObjectEDT to put the ensuing method calls on the EDT
f = java.awt.Font(java.lang.String('Dialog'),java.awt.Font.PLAIN,14);
headers = {'Selected','File','Analysis Routine','Task Status'};
tbl = javaObjectEDT(createTable(hFig,headers,[],false,'Font',f));
 
%set column 1 to use check boxes and set up a change callback
tbl.setCheckBoxEditor(1);
jtable = javaObjectEDT(tbl.getTable); %get the underlying Java Table. IMPORTANT: we need to put jtable on the EDT
columnModel = javaObjectEDT(jtable.getColumnModel); %now we can now do direct calls safely on jtable
selectColumn = javaObjectEDT(columnModel.getColumn(0));
selectColumnCellEditor = selectColumn.getCellEditor;
chk = javaMethodEDT('getComponent',selectColumnCellEditor);
set(chk,'ItemStateChangedCallback',@chkChange_Callback);
 
%make column three a combo drop down
analysisTable = {'Analysis1';'Analysis2';'Analysis3'};
cb = javaObjectEDT('com.mathworks.mwswing.MJComboBox',analysisTable);
cb.setEditable(false);
cb.setFont(f);
set(cb,'ItemStateChangedCallback',@cbChange_Callback);
editor = javaObjectEDT('javax.swing.DefaultCellEditor',cb);
analysisColumn = javaObjectEDT(columnModel.getColumn(2));
analysisColumn.setCellEditor(editor);
 
%set some column with restrictions
selectColumn.setMaxWidth(100);
analysisColumn.setPreferredWidth(300);
 
%set the data
SELECTED = java.awt.event.ItemEvent.SELECTED;
tbl.setData({false,'file1','Analysis2','Analysis2';...
             true,'file2','Analysis3','Analysis3'});
drawnow;
 
    function cbChange_Callback(src,ev) %#ok
        jRow = jtable.getSelectedRow;
        stateChange = javaMethodEDT('getStateChange',ev);
        if stateChange == SELECTED
            newData = javaMethodEDT('getItem',ev);
            model = jtable.getModel;
            javaMethodEDT('setValueAt',model,newData,jRow,3);
        end %if
    end %cbChange
 
    function chkChange_Callback(src,ev) %#ok
        chkBox = javaMethodEDT('getItem',ev);
        if logical(javaMethodEDT('isSelected',chkBox))
            beep; %put useful code here
        else
            beep;
            pause(0.1)
            beep; %put useful code here
        end %if
    end %chkChange_Callback
 
end %tableExample

If you are running Matlab R2008a or later, javacomponent uses the javaObjectEDT function to create the returned objects so you do not have to do anything further to these objects to have their calls dispatched on the EDT. Users need to take care that objects added directly to the components created by javacomponent are on the EDT as well as specialized sub-components (e.g. CellRenderers and CellEditors). The overhead of calling javaMethodEDT is fairly small so if in doubt, use it.

javaObjectEDT and its kin first appeared in R2008a, although they only became supported in R2008b. Unfortunately, using them on R2008a sometimes causes hangs and all sorts of other mis-behaviors. This problem was fixed in the R2008b release, when javaObjectEDT became a fully-supported function. The problem with using javaObjectEDT in our application is that if it ever runs on an R2008a platform it might hang! (on Matlab release R2007b and earlier we will get an informative message saying that the javaObjectEDT function does not exist)

For this reason, I am using the following method in my projects:

function result = javaObjEDT(varargin)
%Placeholder of Matlab's buggy javaObjectEDT function on R2008a
 
% Programmed by Yair M. Altman: altmany(at)gmail.com
% $Revision: 1.2 $  $Date: 2009/01/25 11:31:08 $
 
  try
      try
          result = varargin{1};
      catch
          result = [];
      end
      v = version;
      if str2double(v(1:3)) > 7.6
          result = builtin('javaObjectEDT',varargin{:});
      end
  catch
      % never mind
  end
end

Note that javaMethodEDT has the method name as its first input argument, and the object name or reference as its second arg. This is inconsistent with many other Matlab/Java functions, which normally accept the target object as the first argument (compare: invoke, awtinvoke, notify etc.). It also means that we cannot use the familiar obj.javaMethodEDT(methodName) format.

One final note: when javaObjectEDT and javaMethodEDT first appeared in R2008a, they were complemented by the javaObjectMT and javaMethodMT functions, which create and delegate Java objects on the main Matlab computational thread. Their internal documentation says that there are cases when execution must occur on the MT rather than EDT, although I am personally not aware of any such case.

awtcreate and awtinvoke

For users with versions prior to R2008b the user must use the awtcreate function to create objects on the EDT. One huge disadvantage of this older function is that if you have to pass java objects in the parameter list you must use the very cumbersome JNI style notation. For example, for the simple task of setting a button label, one has to use:

btn = awtcreate('javax.swing.JButton');
awtinvoke(btn,'setText(Ljava/lang/String;)','click me')

The other disadvantage is that creating the object using awtcreate does not ensure that its subsequent method calls will be executed on the EDT. The awtinvoke function must be used for each call.

Also, both awtcreate and awtinvoke have some limitations due to bugs in the private parseJavaSignature function (for example, invoking methods which accept a java.lang.Object) which forces one to use the direct call to the method, using the main Matlab thread. This can result in the undesired effects described above. In this situation the best workaround is to call pause(0.01) to allow the event queue to clear.

Versions of javacomponent earlier than R2008a use awtcreate and objects created by these versions must have their subsequent methods called by awtinvoke to be used on the EDT.

A very rare CSSM thread discusses the usage of awtcreate and awtinvoke with some very interesting remarks by MathWorks personnel.

There is an interesting option in awtinvoke that was not carried over into the newer javaMethodEDT. This option allows the user to pass a function handle in the argument list along with its parameters. This option creates an undocumented com.mathworks.jmi.Callback object that has a delayed callback. The delayed callback is dispatched on the EDT so that it will be called once the java method used in awtinvoke is finished. Note that the actual function will still execute on the main Matlab thread the delayed callback will just control when it is called. However this may be useful at times. It is possible to put this functionality into a separate function we can call to delay execution until the event queue is cleared.

%CALLBACKONEDTQUEUE will place a callback on the EDT to asynchronously
%run a function.
%CALLBBACKONEDTQUEUE(FCN) will run function handle FCN once all previous
%methods dispatched to the EDT have completed.
%CALLBBACKONEDTQUEUE(FCN,ARG1,ARG2,...) ill run function handle FCN with
%arguments ARG1,ARG2...once all previous methods dispatched to the EDT
%have completed.
%Note that the function is still executing on the main Matlab thread. This
%function just delays when it will be called.
function callbackOnEDTQueue(varargin)
    validateattributes(varargin{1},{'function_handle'},{});
    callbackObj = handle(com.mathworks.jmi.Callback,'callbackProperties');
    set(callbackObj,'delayedCallback',{@cbEval,varargin(:)});
    callbackObj.postCallback;
 
    function cbEval(src,evt,args) %#ok
        feval(args{:});
    end %cbEval
end %callbackOnEDTQueue

Setting desktop tab completions

Wednesday, March 3rd, 2010

This site has lately focused on quite detailed Java-related topics. Next week I will present the promised EDT article, which will dive into even deeper Java territory. So I thought to take a short break and present an entirely pure-Matlab non-Java undocumented feature, which is simple and yet quite useful.

A few months ago, a CSSM reader asked whether it is possible to customize Matlab tab-completion for user-defined functions (see related). A similar question on StackOverflow provided the necessary solution lead:

Apparently, Matlab has a file called TC.xml in its [matlabroot '/toolbox/local/'] folder that contains the definitions of the tab-completable functions and their arguments. In order for a user-defined function’s arguments to support tab-completion, a new entry needs to be added to this XML file.

TC.xml & TC.xsd

The full syntax of the TC.xml file can be found in the TC.xsd file, which is located in the same folder as TC.xml. Here are some sample definitions from my TC.xml file (which might vary across Matlab releases):

<binding name="addpath" ctype="DIR"/>
<binding name="help"    ctype="FUN SUBFUN"/>
<binding name="clear"   ctype="FUN VAR"/>
 
<binding name="whos"    ctype="VAR">
  <arg previous="-file" ctype= "MATFILE"/>
</binding>
 
<binding name="open">
  <arg argn="1" ctype="VAR MATFILE FIGFILE MFILE MDLFILE FILE"/>
</binding>
 
<binding name="openfig">
  <arg argn="1" ctype="FIGFILE"/>
  <arg argn="2" ctype="VAR" value="new visible invisible reuse"/>
</binding>
 
<binding name="mlint"   ctype="FUN">
  <arg argn="2:10" ctype="VAR" value="-struct -string -id"/>
</binding>

The first example defines that an unlimited number of addpath arguments are all of type DIR. Therefore, when completing any argument of this function in the Command-Window, Matlab will present only relevant DIR (=folder) elements in the pop-up window (lexically sorted):

Tab-completion of type DIR

Tab-completion of type DIR

Similarly, help defines all its arguments to be a function or sub-function type, so the popup-up will only be populated with the function names currently visible in the desktop:

Tab-completion of types FUN & SUBFUN

Tab-completion of types FUN & SUBFUN

Similarly, clear defines all its arguments as function names or variables. Note that the list of available functions and variables may change depending on the current execution stack position. The full list of supported types is defined in the TC.xsd file. It is: VAR, FUN, SUBFUN, DIR, FILE, MFILE, MATFILE, FIGFILE and MDLFILE.

The whos function defines all its arguments as VAR, except the single MATFILE argument that follows a ‘-file’ argument (look at whos’s help page to understand why).

The open function defines tab completion only for its first argument (with plenty of possible types…). Likewise, openfig defines its first argument as a FIGFILE, and its second as VAR with a few extra special-purpose strings that are added to the popup-up menu.

Finally, the mlint example shows that multiple arguments can be defined using a single XML definition element. In this case, args #2-10 are defined as VAR (with three extra special-purpose strings), while args #1 and #11+ are defined as FUN.

The careful user can edit the TC.xml file using any text editor (I strongly suggest saving a backup first):

edit(fullfile(matlabroot,'toolbox/local/TC.xml'))

User-defined functions can easily be added to TC.xml, and we can even add/modify the built-in Matlab functions that are already defined. Note that changes to TC.xml only take effect after a Matlab restart. From then on, all future Matlab sessions will use the modification, so a really simple one-time edit can improve our workflow for a long time – at least until we upgrade Matlab, when we’ll need to redo our edits…

TabComplete utility

In order to facilitate TC.xml editing, I have created a utility called TabComplete, which is now available on the Matlab File Exchange. The use of this utility is very simple. For example:

tabcomplete test file 'DIR +data -data no_data' VAR

defines a user-defined function test that accepts a FILE argument, followed by a DIR argument with three special-purpose strings, followed by any number of VAR arguments. If I wished to define specific argument types without any default type, I would use:

tabcomplete test file 'DIR +data -data no_data' ''

Using TabComplete for user-defined functions

Using TabComplete for user-defined functions

TabComplete can also be used to retrieve the current list of tab-completion definitions:

>> definitions = tabcomplete;
>> definitions(1)
ans = 
    functionName: 'addpath'
     defaultType: 'DIR'
     extraValues: ''
        platform: ''
    functionArgs: []
 
>> definitions(54)
ans = 
    functionName: 'openfig'
     defaultType: ''
     extraValues: ''
        platform: ''
    functionArgs: [1x2 struct]
>> definitions(54).functionArgs(1)
ans = 
    previousArg: ''
        argType: 'FIGFILE'
    extraValues: ''
>> definitions(54).functionArgs(2)
ans = 
    previousArg: ''
        argType: 'VAR'
    extraValues: 'new visible invisible reuse'

TabComplete has a few limitations: it does not support the -previous option described above (you can do this by manually editing TC.xml). There are also some inherent limitations in Matlab’s TC functionality: changes take effect only after a Matlab restart (there might be a way to reload the definitions in the current Matlab session, but I do not know of any); the list of standard types cannot be modified; and the default type does not support extra special-purpose strings as do the numbered arguments.

There is another very annoying limitation: by default, TC.xml only supports lowercase function names. This is stupid, since Matlab has many function names with UPPERCASE characters, and certainly user-defined function names also do. Luckily, this last limitation can easily be overcome by editing the TC.xsd file (note that this is the TC.XSD file, not the TC.XML file). Instead of:

<xsd:simpleType name="tcBindingNameType">
  <xsd:restriction base="xsd:token">
    <xsd:pattern value='[A-Za-z_0-9]+(/[a-z_0-9]+)?'/>
  </xsd:restriction>
</xsd:simpleType>

Change the xsd:pattern definition element to:

    <!-- Yair 21/2/2010: added A-Z -->
    <xsd:pattern value='[A-Za-z_0-9]+(/[A-Za-z_0-9]+)?'/>

(note the way comments can be added to the XSD/XML files)

P.S. an entirely different customization, for user-defined class members, was presented by Michal Kutil.