Re: Outerposition Patch

[Ben Abbott]

On Feb 18, 2011, at 6:35 PM, logari81 wrote:

> On Fri, 2011-02-18 at 22:40 +0100, Konstantinos Poulios wrote:
>> On Fri, Feb 18, 2011 at 6:07 PM, bpabbott <address@hidden> wrote:
>>> On Feb 18, 2011, at 12:02 PM, bpabbott <address@hidden> wrote:
>>> 
>>> On Feb 18, 2011, at 11:04 AM, Konstantinos Poulios <address@hidden>
>>> wrote:
>>> 
>>> On Fri, Feb 18, 2011 at 2:36 PM, Ben Abbott <address@hidden> wrote:
>>>> On Feb 18, 2011, at 2:37 AM, logari81 wrote:
>>>> 
>>>>> On Thu, 2011-02-17 at 21:09 -0500, Ben Abbott wrote:
>>>>> 
>>>>>> On Feb 17, 2011, at 4:32 PM, logari81 wrote:
>>>>>> 
>>>>>>> On Wed, 2011-02-16 at 19:37 -0500, Ben Abbott wrote:
>>>>>>> 
>>>>>>>> On Feb 13, 2011, at 12:05 PM, logari81 wrote:
>>>>>>>> 
>>>>>>>>> On Thu, 2011-02-10 at 20:00 +0100, Konstantinos Poulios wrote:
>>>>>>>>> 
>>>>>>>>>> On Thu, Feb 10, 2011 at 1:59 PM, Konstantinos Poulios wrote:
>>>>>>>>>> 
>>>>>>>>>>> On Thu, Feb 10, 2011 at 9:25 AM, David Bateman <address@hidden>
>>>>>>>>>>> wrote:
>>>>>>>>>>> 
>>>>>>>>>>>> Le 10 févr. 2011 à 00:25, logari81 <address@hidden> a
>>>>>>>>>>>> écrit :
>>>>>>>>>>>> 
>>>>>>>>>>>>> thank you for this information.
>>>>>>>>>>>>> 
>>>>>>>>>>>>> It seems that the previously attached patch causes problems only
>>>>>>>>>>>>> with
>>>>>>>>>>>>> legends. However, in order to treat legends correctly I need to
>>>>>>>>>>>>> understand their logic. How do legends exploit the outerposition
>>>>>>>>>>>>> and
>>>>>>>>>>>>> position properties? Is anyone familiar with legend.m to give me
>>>>>>>>>>>>> a short
>>>>>>>>>>>>> introduction?
>>>>>>>>>>>>> 
>>>>>>>>>>>>> Best regards
>>>>>>>>>>>> 
>>>>>>>>>>>> You shouldn't try to understand the logic of legend's use of the
>>>>>>>>>>>> position and outerposition properties. It's just a hack that 
>>>>>>>>>>>> worked with the
>>>>>>>>>>>> existing behavior. If your patch doesn't work well with legend it 
>>>>>>>>>>>> is
>>>>>>>>>>>> probably legend that needs to be fixed
>>>>>>>>>>>> 
>>>>>>>>>>>> D.
>>>>>>>>>>> 
>>>>>>>>>>> The attached patch replaces the previous one and implements the
>>>>>>>>>>> calculation of both position and outerposition depending on the
>>>>>>>>>>> value
>>>>>>>>>>> of activepositionproperty.
>>>>>>>>>>> 
>>>>>>>>>>> It is not very well tested yet, so there will probably be some
>>>>>>>>>>> issues,
>>>>>>>>>>> e.g. legends will not work, but it brings a feature that maybe is
>>>>>>>>>>> awesome. It is something that Matlab cannot do and maybe you like
>>>>>>>>>>> it.
>>>>>>>>>>> See the following video:
>>>>>>>>>>> 
>>>>>>>>>>> http://ubuntuone.com/p/cYM/
>>>>>>>>>>> 
>>>>>>>>>>> The position property is calculated dynamically while you rotate
>>>>>>>>>>> the
>>>>>>>>>>> view, so that all labels fit in outerposition. I think it works
>>>>>>>>>>> quite
>>>>>>>>>>> well in order to keep it. What do you think?
>>>>>>>>>>> 
>>>>>>>>>>> As this operation involves certain computational overhead, it would
>>>>>>>>>>> be
>>>>>>>>>>> interesting to get some tests on older machines. Unfortunately all
>>>>>>>>>>> pc's that I have access to, are too fast.
>>>>>>>>>>> 
>>>>>>>>>>> This patch also fixes http://savannah.gnu.org/bugs/?31610 for the
>>>>>>>>>>> fltk toolkit.
>>>>>>>>>>> 
>>>>>>>>>>> Finally, if we adopt the attached patch we have to adapt legend.m
>>>>>>>>>>> also.
>>>>>>>>>>> 
>>>>>>>>>>> BR
>>>>>>>>>>> 
>>>>>>>>>>> Kostas
>>>>>>>>>> 
>>>>>>>>>> After some more testing and fixes I think the patch is quite mature
>>>>>>>>>> in
>>>>>>>>>> the form you find in the attachment. I think it could be checked in.
>>>>>> 
>>>>>>>>> 
>>>>>>>>> I have just checked in this changeset along with some further
>>>>>>>>> fixes/improvements. Now, I would like to provide some additional
>>>>>>>>> information and ask for some help with regard to the open issues that
>>>>>>>>> I
>>>>>>>>> had listed.
>>>>>>>>> 
>>>>>>>>>> There are still some general issues with fltk that I will try to sum
>>>>>>>>>> up:
>>>>>>>>>> 
>>>>>>>>>> 1. In some demo plots axes labels seem to be too close to the axes
>>>>>>>>>> (e.g. demo legend 9). Probably in some of the previous changes there
>>>>>>>>>> is something that I have overseen.
>>>>>>>>> 
>>>>>>>>> Actually after testing older revisions of octave I realized that the
>>>>>>>>> problem is not new. The reason that I hadn't noticed it before is
>>>>>>>>> because the problem appears only in the print output and not in the
>>>>>>>>> plot
>>>>>>>>> window. It seems that gl-render and gl2ps position strings
>>>>>>>>> differently
>>>>>>>>> considering either the bottom line or the baseline of the string
>>>>>>>>> respectively. It is not difficult to fix, we just have to decide
>>>>>>>>> which
>>>>>>>>> of gl-render and gl2ps are we going to fix in order to make both
>>>>>>>>> consistent.
>>>>>>>>> 
>>>>>>>>>> 2. Legends for barplots don't show colors (this is an old problem).
>>>>>>>>>> 
>>>>>>>>>> 3. Some small y axes interference for plotyy (also not new).
>>>>>>>>>> 
>>>>>>>>>> 4. Now there is no labels-titles interference in demo subplot 1, so
>>>>>>>>>> there is no need for extra space between the subplots, we can reduce
>>>>>>>>>> a
>>>>>>>>>> bit the padding (someone which is familiar with subplot.m I
>>>>>>>>>> suppose).
>>>>>>>>> 
>>>>>>>>> Waiting for someone familiar with subplot.m
>>>>>>>> 
>>>>>>>> I"ve just pushed a changeset that improves the layout of the subplots.
>>>>>>>> 
>>>>>>>>   http://hg.savannah.gnu.org/hgweb/octave/rev/7b67bbf9dbbb
>>>>>>>> 
>>>>>>>> I'm also attaching a test script that runs under Octave and Matlab.
>>>>>>>> Results for both are attached.
>>>>>>> 
>>>>>>> This script is cool, I was thinking of doing something like that but I
>>>>>>> didn't realize that it can be done so easily.
>>>>>>> 
>>>>>>>> The test script places dashed blue lines around the position of each
>>>>>>>> axis, and dashed red around the outerposition.
>>>>>>> 
>>>>>>> You mean blue lines around the original axes position before adding
>>>>>>> labels and titles. The version of the script that I have attached in
>>>>>>> this email visualizes the updated positions which correctly coincide
>>>>>>> with the axes.
>>>>>> 
>>>>>> Ok. I see your point. I'll have to do some experimenting with the
>>>>>> corrected version.
>>>>>> 
>>>>>>>> When subplot (3,3,1:3) is used to replace the first row of subplots, a
>>>>>>>> green dashed box is used to encompass the new position, and a dashed 
>>>>>>>> magenta
>>>>>>>> for the outerposition.
>>>>>>>> 
>>>>>>>> The problems I see are ...
>>>>>>>> 
>>>>>>>> 1) The activeposition property is still "outerposition".
>>>>>>> 
>>>>>>> why is this a problem? Maybe we prefer this, maybe not, see my comment
>>>>>>> on (2). ML sets it to "position" but we do not have necessarily to do
>>>>>>> the same.
>>>>>> 
>>>>>> We may decide to deviate from compatibility with Matlab, but before
>>>>>> doing so we should discuss it on the list. The list has already discussed
>>>>>> and agreed to Matlab compatibility (before my time here), it would be
>>>>>> improper to deviate from that agreed upon approach without discussion 
>>>>>> first.
>>>>>> 
>>>>>> Can we abide by Matlab's example for now and discuss changes later. If
>>>>>> nothing else, that would make it easier (for me) to review the state of
>>>>>> graphics for Octave (via dump_demos and such).
>>>>>> 
>>>>>>>> 2) The width of subplot (3,3,1:3) has been improperly modified on the
>>>>>>>> c++ side.
>>>>>>> 
>>>>>>> Actually this is not really "improperly". It is doing what it was
>>>>>>> expected to do. What we programmed in c++ is a minimum left margin of
>>>>>>> 13% of outerposition(3). For the upper subplot the total width is 3
>>>>>>> times the width of the other subplots so the minimum left margin is
>>>>>>> also
>>>>>>> 3 times higher. It is ugly.
>>>>>>> This would be a reason for switching to
>>>>>>> activepositionproperty=position.
>>>>>>> This way, we wouldn't let sync_position do its job but we would do it
>>>>>>> manually in the frontend. Now we are able to, before we couldn't
>>>>>>> because
>>>>>>> we couldn't get any tightinset values.
>>>>>> 
>>>>>> If consistent with Matlab, the subplot(3,3,1:3) would produce an axes
>>>>>> with a position property that encompasses the original 3 axes (Matlab
>>>>>> documents this, but I've noticed some minor "bugs" on their part).
>>>>>> 
>>>>>> For now, can the position/outerposition synchronization be implemented
>>>>>> in the manner that is consistent with Matlab's documentation?
>>>>>> 
>>>>>> Meaning that when outerposition is active …
>>>>>> 
>>>>>> I) position(1) is adjusted to the right (never to the left), to ensure
>>>>>> no object extends to the left of outerposition(1).
>>>>>> 
>>>>> Right now, no objects should extend left of outerposition(1). If there
>>>>> is a test case not respecting this rule, please let me know (only
>>>>> exception is if you make the plot window tiny).
>>>> 
>>>> I don't see any cases of that. What I do see is that you're requiring a
>>>> minimum of space between the outerposition and position boxes. So you're
>>>> adding features, correct?
>>>> 
>>>>>> II) position(2) is adjusted upward (never downward), to ensure no object
>>>>>> extends below the outerposition(2)
>>>>>> 
>>>>> 
>>>>> likewise
>>>> 
>>>> Your approach is not consistent with the "never downward" part, Correct?
>>>> 
>>>>>> III) position(3) is decreased (never increased) to ensure no object
>>>>>> extends beyond the outerposition(1)+outerposition(3).
>>>>>> 
>>>>> 
>>>>> likewise
>>>> 
>>>> Same.
>>>> 
>>>>>> IV) position(4) is decreased (never increased) to ensure no object
>>>>>> extends beyond the outerposition(2)+outerposition(4).
>>>>>> 
>>>>> 
>>>>> likewise
>>>> 
>>>> Same again.
>>>> 
>>>>>> When the position property is active the relationship is reversed. Its
>>>>>> been a couple of years since I looked that this in detail. Is my
>>>>>> understanding of how Matlab works consistent with yours?
>>>>>> 
>>>>>>>> 3) The positions have been shifted to the left relative to what was
>>>>>>>> specified by subplot.m. Originally, their left edges were very close 
>>>>>>>> to the
>>>>>>>> left edge of the outerpositon.
>>>>>>>> 
>>>>>>> 
>>>>>>> What do you mean by "left edges" I don't get this point.
>>>>>> 
>>>>>> Opps ... not "left", but "right"!
>>>>>> 
>>>>>> My observation was that the right edge of the "position" has been
>>>>>> shifted to the left even though no object impinged upon the right edge of
>>>>>> the outerposition.
>>>>>> 
>>>>> 
>>>>> This is 9.5%. The problem is that we consider these minimum margins 13%
>>>>> to the left, 9.5% to the right, 11% to the bottom, 7.5 %to the top. This
>>>>> is compatible with ML for normal plots, but for subplots ML reduces this
>>>>> limits. Actually subplot in ML is quite a hack. We have different
>>>>> possibilities of achieving the same behavior. I make a proposal at the
>>>>> end.
>>>> 
>>>> Actually Matlab does not have "minimum margins". Those margins are set by
>>>> the "defaultaxisposition" and "defaultaxisouterposition" properties 
>>>> (present
>>>> in the root, figure and axes objects). Thus, they are controlled on by
>>>> m-file side by the user.  So I think the current synchronization isn't
>>>> compatible with Matlab, even for normal plots.
>>>> 
>>>>>>>> 4) The xticklabels and yticklabels should be tigher to the axes.
>>>>>> 
>>>>>>> 
>>>>>>> This is adjustable I think. Maybe it makes sense to calculate the
>>>>>>> distance of ticklabels from axes as percentage of axes sizes.
>>>>>> 
>>>>>> This isn't a documented by MathWorks. However, I did some experimenting
>>>>>> and found that ...
>>>>>> 
>>>>>> a) xlabel baseline is (2*fontsize + 7) points below the axis position
>>>>>> 
>>>>>> b) x-axis ticklabels are (fontsize + 1.5) points below the axis position
>>>>>> 
>>>>> 
>>>>> the bigger the font, the higher the distance from the axis
>>>> 
>>>> Yes.
>>>> 
>>>>>> c)  the right extent of the y-axis ticklabels is (20/fontsize + 1)
>>>>>> points to the left of the axis position.
>>>>>> 
>>>>> 
>>>>> so, the bigger the font, the closer to the axis? Interesting.
>>>> 
>>>> Keep in mind that the extent property is designed for type-settting, so
>>>> there is some white space included. Thus, the visual result does not give
>>>> the impression the characters are closer to the axis box. Matlab and
>>>> Octave's extents aren't yet consistent, so there is good reason not to
>>>> blindly copy this feature. However, I do think the spacing should rely upon
>>>> the font and not the axis.
>>>> 
>>>>> Do you believe these 3 approximations a,b and c are fixed or they could
>>>>> change proportionally to the axes width/height.
>>>> 
>>>> No they do not. This is easily seen my resized the figure with the mouse.
>>>> 
>>>>>> As this isn't documented by MathWorks, they could change it. So there is
>>>>>> no compelling reason to copy the specifics.
>>>>>> 
>>>>>> However, if there are multiple axes in the same figure, I think the
>>>>>> spacing between axes, ticklabels, and labels  should be consistent 
>>>>>> (assuming
>>>>>> the fontsize is consistent). Does that make sense? Other thoughts /
>>>>>> concerns?
>>>>>> 
>>>>>> Ben
>>>>> 
>>>>> SUGGESTION:
>>>>> 
>>>>> 1st step: Add a new property (hidden?) to the axes object:
>>>>> minmargins = [l b rt]
>>>>> with default value derived from defaultaxesposition:
>>>>> l=defaultaxesposition(0)
>>>>> b=defaultaxesposition(1)
>>>>> r=1-defaultaxesposition(0)-defaultaxesposition(2)
>>>>> t=1-defaultaxesposition(1)-defaultaxesposition(3)
>>>> 
>>>> I rather not see this done. The margins are currently defined by the user
>>>> on the m-file side by changing the position/outerposition of the axes. This
>>>> just looks more complicated to me with no added capability.
>>>> 
>>>>> 2nd step: Modify sync_positions so that it takes into account minmargins
>>>>> instead of defaultaxesposition. This would mean no change for all other
>>>>> plots, but for subplots it gives as the possibility to reduce the
>>>>> minimum margins from the frontend (e.g. reduce the ugly 9.5% to the
>>>>> right).
>>>> 
>>>> I'd prefer that the synchronization limit itself to the compatible
>>>> behavior. For activepositionproperty = "outerposition"
>>>> 
>>>> I) position(1) is adjusted to the right (never to the left), to ensure no
>>>> object extends to the left of outerposition(1).
>>>> 
>>>> II) position(2) is adjusted upward (never downward), to ensure no object
>>>> extends below the outerposition(2).
>>>> 
>>>> III) position(3) is decreased (never increased) to ensure no object
>>>> extends beyond the outerposition(1)+outerposition(3).
>>>> 
>>>> IV) position(4) is decreased (never increased) to ensure no object extends
>>>> beyond the outerposition(2)+outerposition(4).
>>>> 
>>>> In short the position property never expands, but retracts to keep itself
>>>> and its children inside the outerposition.
>>>> 
>>>> Conversely, when the activepositionproperty == "position", the
>>>> outerposition never contracts, but expands so as to encompass the axis and
>>>> its children.
>>>> 
>>>> One of the difficulties I'm having with subplot is that the synchonization
>>>> second guesses the specified position. In addition, the current solution
>>>> will be difficult to document.
>>>> 
>>>>> 3rd step: Optimize subplot.m making use of the new property minmargins
>>>>> 
>>>>> Only by setting minmargins to zero would eliminate most problems that we
>>>>> observe now with subplot. More sophisticated use of minmargins would
>>>>> even allow us to synchronize the insets in rows and columns of the
>>>>> subplot grid (AFAIK is what ML does, can you confirm this?).
>>>>> 
>>>>> What do you think? Should I add the a property minmargins or something
>>>>> similar?
>>>> 
>>>> 
>>>> Ok, Please propose a changeset with the default for  minmargins set to
>>>> zero so that we'll have a compatible solution.
>>>> 
>>>> Ben
>>>> 
>>> 
>>> Hmm, I have a suggestion. Since I thought that the implementation of
>>> sync_position for single plots (not subplots) is compatible with ML,
>>> and you are saying that it isn't, this should be the first issue to
>>> fix. Could you provide me with an example of a single plot that
>>> demonstrates the difference between ML and Octave?
>>> 
>>> As soon as I fix this we can come back to subplot again and continue
>>> our discussion.
>>> 
>>> BR
>>> 
>>> Kostas
>>> 
>>> First example is for activeposition == "position"
>>> figure (1)
>>> clf
>>> set (gca, 'position', [0 0 1 1], 'activepositionproperty', 'outerposition')
>>> plot (0:1,0:1)
>>> axis ([0 1 0 1])
>>> outerposition = get (gca, 'outerposition')
>>> I've attached the result from Matlab.  The outerposition from Matlab is
>>> outerposition =   -0.1677   -0.1350    1.2903    1.2270
>>> Octave's result does not grow the outerposition.
>>> outerposition =   0   0   1   1
>> 
>> ouuuuups!!! I introduced this bug in 98772e4e8a2a. It used to work
>> correctly before. I have just pushed the fix, so it should be ok
>> again.
>> 
>>> If this example so that activeposition == "outerposition" ...
>>> figure (1)
>>> clf
>>> set (gca, 'position', [0 0 1 1], 'activepositionproperty', 'outerposition')
>>> set (gca, 'outerposition', [0 0 1 1])
>>> plot (0:1,0:1)
>>> axis ([0 1 0 1])
>>> … then I see that the default axis position is restored. This does behave in
>>> the manner you're suggesting, but it is not described by the documentation.
>>> http://www.mathworks.com/help/techdoc/creating_plots/f1-32495.html
>>> This behavior is new to me (wasn't there when I examined this a few years
>>> back). So it appears I owe you an apology for the back-n-forth.
>>> I did a quick google, and found that someone else named "Ben" had figured
>>> out what is happening.
>>> http://undocumentedmatlab.com/blog/tag/outerposition/
>>> Rather than minmargins, may I suggest you use "looseinset" as Matlab does?
>>> For the subplots, the looseinset may be set to some reasonable value by the
>>> subplot.m function.
>>> Ben
>>> 
>>> 
>>> Same article, but this time a direct link
>>> http://undocumentedmatlab.com/blog/axes-looseinset-property/
>>> Ben
>> 
>> I am working on looseinset now. It shouldn't take long to implement.
>> 
>> BR
>> 
>> Kostas
> 
> Hi Ben,
> 
> in the attached changeset you can find a first implementation of
> looseinset. The sync_position function relies now on looseinset instead
> of default_axes_position.
> 
> Known limitations: looseinset remains always in normalized units (since
> it is a hidden property I see no need to support other kinds of units)
> 
> Please test this patch and send me any comments.
> 
> BR
> 
> Kostas
> <looseinset-1a.changeset>

For consistency, how difficult is it to implement the units conversion? ... or 
maybe a more proper question is; Is there a reason that units conversion is 
prohibitive?

In the meantime, I'll do some testing with your changeset.

Ben