matplotlib

On Fri, 2005-01-28 at 21:36 +0100, Sigve Tjora wrote: 
> Hi,
> I have written a basic QtAgg backend for Matplot-lib. It is only tested
> on Windows, so some more testing should be done, but it is my hope that
> it can some day be included in the matplotlib distribution. I will test
> on linux when I get the time, but all the code is python so it should
> probably work ok.
Well done. I tried it out on Linux (Fedora 3) and it works reasonably
well.

I noticed a few things:
It puts the toolbar at the top rather than the bottom of the window.

The graph is smaller and the colours are different than that produced
with GTKAgg, which is strange since they both use Agg rendering.

I think it would be better to have a 'backend_qt.py' file used as a base
class (with a FigureCanvasQtAgg.draw() method that raises
NotImplementedError), and with backend_qtagg.py overriding just those
functions/methods which are specific to drawing using Agg. This would
allow backend_qt.py code to be used to produce a QtCairo backend at a
later date.

I copied the file into cvs so other people can try it out.

Regards
Steve

>>>>> "Sigve" == Sigve Tjora <public@...> writes:

    Sigve> Hi, I have written a basic QtAgg backend for
    Sigve> Matplot-lib. It is only tested on Windows, so some more
    Sigve> testing should be done, but it is my hope that it can some
    Sigve> day be included in the matplotlib distribution. I will test
    Sigve> on linux when I get the time, but all the code is python so
    Sigve> it should probably work ok.

Indeed -- it's a testament to QT and python that the untested code
appears to run the same on linux (per Steve's report) and OS X, which
I tested today.  Good work!

    Sigve> See http://www.tjora.no/python/matplotlib/ for the code.

    Sigve> Most examples run ok, but the keypress_demo.py does not
    Sigve> work yet. Not all examples have been tested, just some
    Sigve> random double-clicking on the one that look interesting.

I think the problem here was that you had 

    def keyPressEvent(self, ev):
        self.keyPressEvent(ev.text())
    def keyReleaseEvent(self, ev): 
        self.keyReleaseEvent(ev.text())

where you meant

    def keyPressEvent(self, ev):
        self.key_press_event(ev.text())
    def keyReleaseEvent(self, ev): 
        self.key_release_event(ev.text())

I made these changes in CVS, and the keypress_demo indeed works.

    Sigve> A lot of the comments from the backend_template.py has not
    Sigve> been removed or fixed. The code is also somewhat rough
    Sigve> around the edges, but it works.

    Sigve> Is there any unittest-suite it is possible to run the
    Sigve> backend against?

The standard test suite is examples/backend_driver. You just need to
set the "backends" list to the list of backends you want to test.  Eg

backends = ['QtAgg']


    Sigve> Feedback is welcome.

I agree with most of the comments Steve made.  

  * I think factoring the class into a module that handles the qt
    widgets, and have a small module that does qtagg would be a good
    thing.  Steve would likely do a qtcairo backend if you did this...

  * The only explanation I can think of for why the colors are wrong
    is that qt has a different byte order for images than you are
    getting from agg (is could it be bgra instead of rgba, or that
    something is wrong with the endianess...)  If you need a different
    byte ordering /pixel format, I can provide the required method in
    the agg backend.  I note that the saved figures *do* have the
    right color scheme, which supports this idea.
  
  * The control key modifier is apparently not recognized -- this
    provides aspect ratio preserving zoom.  I can't test any of the
    right click drag functionality right now, actually, since I'm on a
    mac w/o an external mouse plugged in **&$$ macs!

Thanks!
JDH

John Hunter a =E9crit :
> Here is what I think would be ideal, and I wanted to sketch some of
> these ideas in hopes that you might have some ideas on how to apply
> them.  Basically, the idea is that we want one axes to be able to
> share the x or y limits with another axes in a more general way.  Your
> approach works fine as long as the two axes are overlayed.  It would
> be nice if we could do something like
>=20
>     # separate axes
>     ax1 =3D subplot(211)
>     plot([1,2,3])
>=20
>     ax2 =3D subplot(212, twinx=3Dax1)
>     plot([4,5,6])
>=20
That sounds like a good syntax.
Its not a problem the share the x/ylims. As long as you can get the lazy=20
values at Axes creation, you can set the transforms correctly. The=20
question is whether or not to share the Axis instance. If you do, you=20
will have trouble drawing, if you don't, you have trouble when you want=20
to change the attributes (see below). I got away with this only because=20
I have to draw only one "physical" axis.

> To do overlayed axes, you could do
>=20
>     # overalyed axes
>     ax1 =3D subplot(111)
>     plot([1,2,3])
>=20
>     ax2 =3D subplot(111, twinx=3Dax1, yticks=3D'right')
>     plot([4,5,6])
>=20
OK, you also have to set the frame to off, we'll see the details later.=20
Maybe we can keep just the twin function in pylab.py as a shortcut.

> I think this would be a nice consistent interface and in both cases
> ax2 would share the xlim with ax1.  As far as I can see, the only
> thing getting in the way of extending your approach to handle this
> case are the tick labels, since they would be in the wrong place for
> ax2 in the separate axes case.  For the separate axes case, you would
> typically want tick labeling on the lower axes, but it would be ideal
> to be able to control that as well.
>=20
> I've been meaning to decouple the axis line and tick locations from
> the main axes frame so that you could have offset labels and ticks.
> Perhaps this would be a good time to make both changes together.
>=20
That one is more tricky. All of matplotlib drawing model is based on the=20
premise that one objet draws at one given place. This has many=20
advantages. It allows the object to draw itself, which is good design.=20
It is also necessary for the cases where the final size matters (fonts,=20
linewidth). I don't think we want to change that.

< I stopped here, thought about it for some time, went back later >

I'm starting to wonder which properties of Axis we really want to share=20
between axes: limits, scale (linear/log), fmtdata, label, tick positions=20
(thus a common "tick factory"), tick label texts, viewLim, dataLim (we=20
need to think about update_datalim).

and which we do *not* want to share: all graphic objects, visibility of=20
labels (ticks ?), style of ticks and tick labels (you may want smaller=20
fonts in an inset), type of axis (XAxis, YAxis, ThetaAxis, Color, ...,=20
for ex. we may want to plot z(x,y) as pcolor and z(x) as line with=20
shared z lims)

Maybe we want to split an axis into 2 objects ? I need to think more=20
about this.

> The other problem in the current implementation (and in your patch)
> stems from the fact that for event handling, only one axes gets the
> event.  So in your two_scales.py example, if you pan/zoom the xlimits
> behave correctly but only one axes gets the pan/zoom event for y.  It
> would be nice to pass these events on to all the axes the user is over
> to handle the case of overlapping axes.
>=20
This is uncorrelated, and probably easier than the one above. We would=20
need to modify event.inaxes to be a list (in backend_bases), and act=20
upon all those axes in pan/zoom. When only one axes can be used (for ex.=20
the coordinates of the mouse in the status bar), we would use=20
event.inaxes[0], or maybe the most recently used axes (more difficult,=20
but maybe more consistent...we'll see at implementation time).

> Just some thoughts on what I think the proper behavior should be.  If
> you have any ideas on how to handle these, let me know. =20
>=20
well, we need to discuss that a little bit more, so we'll get a clearer=20
view of where we go :-)

Cheers,
Baptiste

>>>>> "Baptiste" == Baptiste Carvello <baptiste13@...> writes:
    >> 
    Baptiste> That sounds like a good syntax.  Its not a problem the
    Baptiste> share the x/ylims. As long as you can get the lazy
    Baptiste> values at Axes creation, you can set the transforms
    Baptiste> correctly. The question is whether or not to share the
    Baptiste> Axis instance. If you do, you will have trouble drawing,
    Baptiste> if you don't, you have trouble when you want to change
    Baptiste> the attributes (see below). I got away with this only
    Baptiste> because I have to draw only one "physical" axis.

Hi Baptiste,

After reading your email and studying your patch more, I now see how
to generalize this to the case of non-overlapping axes which share
either the x or y axis, eg ganged plots.

I have added some changes to CVS to support sharex and sharey kwargs
(I thought this was a better name than twinx and twiny).  I had to
make some minor changes to axis to support sharing tick locators and
formatters, but nothing substantial.  You can now instantiate an axes
with, eg

  subplot(212, sharex=ax1)

and ditto for sharey.  The view limits, transform function, and tick
Locator and Formatter are shared.  This allows you to pan and zoom on
one axes and have the others follow, which is very nice.  There is a
new example showing how to use this example/shared_axis_demo.py.

I was able to remove the TwinAxes class altogether and use the shared
kwargs in its stead, which is cleaner.  I preserved the "twin"
convenience function (naming it twinx) and all it does is pass the
proper kwargs and make the calls to tick_right.
examples/two_scales.py is updated


    Baptiste> I'm starting to wonder which properties of Axis we
    Baptiste> really want to share between axes: limits, scale
    Baptiste> (linear/log), fmtdata, label, tick positions (thus a
    Baptiste> common "tick factory"), tick label texts, viewLim,
    Baptiste> dataLim (we need to think about update_datalim).

    Baptiste> and which we do *not* want to share: all graphic
    Baptiste> objects, visibility of labels (ticks ?), style of ticks
    Baptiste> and tick labels (you may want smaller fonts in an
    Baptiste> inset), type of axis (XAxis, YAxis, ThetaAxis, Color,
    Baptiste> ..., for ex. we may want to plot z(x,y) as pcolor and
    Baptiste> z(x) as line with shared z lims)

    Baptiste> Maybe we want to split an axis into 2 objects ? I need
    Baptiste> to think more about this.

I don't think the datalim need to be shared because they are
responsible only for autoscaling.  The new approach allows the
different axes to have different properties on their labels, eg font
size as you mentioned.  While you cannot turn off labels selectively
on one axes with 

   set(ax2, xticklabels=[])

because this changes the Formatter which is shared, you can achieve
the same effect by setting the visibility property, which is not
shared

  set( ax2.get_xticklabels(), visible=False)

examples/ganged_plots.py, examples/two_scales.py, and
examples/shared_axis_demo.py show off the new features.

Let me know if this design covers the use cases you can think of.

    Baptiste> This is uncorrelated, and probably easier than the one
    Baptiste> above. We would need to modify event.inaxes to be a list
    Baptiste> (in backend_bases), and act upon all those axes in
    Baptiste> pan/zoom. When only one axes can be used (for ex. the
    Baptiste> coordinates of the mouse in the status bar), we would
    Baptiste> use event.inaxes[0], or maybe the most recently used
    Baptiste> axes (more difficult, but maybe more consistent...we'll
    Baptiste> see at implementation time).

Yes, this needs to be fixed and it is not very hard to do.  There are
three event variables affected, inaxes, xdata and ydata, the latter
two give the data coords of the point in the axes the mouse is over.
As you know, the problem is that the current implementation only
registers the first axes and then breaks out of the loop.  It would be
easy to fix this but I'm worried about two things: backward
compatibility (not a huge problem since only power users are using
this feature) and ease of use.  It is really a corner case to be over
multiple axes, and I am hesitant to force the newbie to deal with this
in the typical case when there are no overlapping axes.

One possibility is to leave inaxes, xdata and ydata alone which
satisfies both problems above.  And then to add a new attribute, axseq
which is a list of (ax, xdata, ydata ) tuples to handle the
overlapping axes case.  Internally, we could use axseq so that pan/zoom
will be handled properly in the two_scales case.  The draw back here
is that having more than one obvious way to do it may also confuse
people down the road.  For clarity, the two interfaces I'm discussing are

def callback1(event):  # current impl.
    if event.inaxes is not None:
        do_something with event.inaxes, event.xdata, event.ydata

def callback2(event):  # candidate impl.
    for ax, xdata, ydata in event.axseq:
        do_something with ax, xdata, ydata



I'm weakly inclined to break backward compatibility and go with the
cleaner, comprehensive design in callback2, which on the face of it
doesn't look too hard to use.  

Suggestions?

JDH

John Hunter a =E9crit :
> I have added some changes to CVS to support sharex and sharey kwargs
> (I thought this was a better name than twinx and twiny).  I had to
> make some minor changes to axis to support sharing tick locators and
> formatters, but nothing substantial.  You can now instantiate an axes
> with, eg
>=20
>   subplot(212, sharex=3Dax1)
>=20
> and ditto for sharey.  The view limits, transform function, and tick
> Locator and Formatter are shared.  This allows you to pan and zoom on
> one axes and have the others follow, which is very nice.  There is a
> new example showing how to use this example/shared_axis_demo.py.
>=20
Hi John,

I like this implementation a lot. Cool stuff !

> I was able to remove the TwinAxes class altogether and use the shared
> kwargs in its stead, which is cleaner.  I preserved the "twin"
> convenience function (naming it twinx) and all it does is pass the
> proper kwargs and make the calls to tick_right.
> examples/two_scales.py is updated
>=20
you also need to call tick_left on the original axis, but that is a=20
minor correction.

> I don't think the datalim need to be shared because they are
> responsible only for autoscaling.
>
There is still a problem because autoscaling is disabled on the second=20
axes. I have an idea on how to solve this, I'll try it and send a patch=20
later in the week.

> One possibility is to leave inaxes, xdata and ydata alone which
> satisfies both problems above.  And then to add a new attribute, axseq
> which is a list of (ax, xdata, ydata ) tuples to handle the
> overlapping axes case.  Internally, we could use axseq so that pan/zoom
> will be handled properly in the two_scales case.  The draw back here
> is that having more than one obvious way to do it may also confuse
> people down the road.
>=20
This sounds reasonable to me. I don't think it is confusing. If people=20
know what to do with multiple axes, they'll go with axseq, if not=20
inaxes, xdata and ydata will provide the most reasonable choice for them.
I'll play a little bit with this, and see if it breaks anything I use.

Cheers,
BC

John Hunter wrote:
>>>>>>"Baptiste" == Baptiste Carvello <baptiste13@...> writes:
> 
> 
>     Baptiste> I know the patch is intrusive, though maybe not as much
>     Baptiste> as it looks (everything except protect_mutex.py is
>     Baptiste> pretty trivial). I didn't expected it to be accepted
>     Baptiste> right away, I see it more as a contribution to the whole
>     Baptiste> gui thread discution.
> 
>     Baptiste> As for gui_thread, I looked at it last year, I just
>     Baptiste> remember I didn't like that code that much. Maybe I
>     Baptiste> should get another look...
> 
> Most of what I've heard of gui thread (caveat, the situation may be
> different now but this was the last I heard) was that the code was
> hairy, required scipy distutils to build, was wx specific and had
> problems that required significant work to fix.  In that context, a
> patch, even an intrusive one, that allowed gtk* and wx* to work from
> an arbitrary python shell would be most welcome.  While I think
> ipython is the one true shell, there are many other shells in
> existence and that won't change.  If for example, this patch allowed
> gtk* to be used from within IDLE or pycrust in addition to the
> standard python shell, that would be a big win.

I certainly have never been a fan of digging too deep into gui_thread myself, 
I admit that much :)  Last time I looked at the code, I got a bit scared.  On 
the other hand, Prabhu recently did a bunch of work on it, so it may be 
cleaner/better, I just don't know.

And as I said, I really was not speaking from the ipython side of things at 
all, I don't know why Baptiste took it that way (I didn't even mention ipython 
in my post).  My opinion was strictly based on the generic approach offered by 
gui_thread, and the potential benefit of doing things that way.

> The problem we need to solve is substantially easier than the one
> gui_thread wants to solve, if I understand the issues correctly.  gui
> thread enables you to interact with general wx apps from a general
> python shell.  We want gtk* and wx* backends to work from a general
> python shell.  If this patch addresses that, it's worth a close look.

One thing that I like about gui_thread is that it works in a generic manner. 
What happens with this patch if someone uses matplotlib but adds new widgets? 
Or if they embed a matplotlib window into something else? Are such usage cases 
covered automatically, or do they require special handling by third-party code?

There certainly are benefits to a matplotlib-based self-contained solution, 
and gui_thread comes with its own baggage, so I'll let you guys judge where 
the balance rests.

Cheers,

f

Hi John and Fernando,

John, here is the quick answer to your questions:

* the patch does almost nothing (a few calls to a dummy function) except 
when the gui_protect rcParam is enabled

* when enabled the overhead is acceptable (on my Pentium III 500MHz), 
but import time of pylab is a little longer, because most of the work 
happens there. Also, only the functions and methods defined by 
matplotlib are protected. The idea is that if you call functions or 
methods from the gtk module, you ought to know what you're doing anyway 
(for that matter, I am not as ambitious as gui_thread).

* the patch currently doesn't work with wx because I've not finished the 
work, but it is doable. The GTK or WX mainloop is launched at the time 
the backend is imported, and only the one for the backend you are using 
will be launched.

* if you use interactive.py, or the code in IPython, you'd better not 
enable gui_protect. If you do, 2 GTK mainloops will be launched. This is 
allowed in GTK, but doesn't make sense.

* the GTK mainloop (probably same for WX) can run concurrently with tk. 
This means no problem plotting GTK from IDLE. You can even switch 
backends at runtime by reloading matplotlib.backends and matplotlib.pylab

* the GTK and WX mainloops cannot run concurrently. This is because, at 
least on Linux, WX itself launches a GTK mainloop (it's called WxGTK for 
a reason :). It might be possible to use this very mainloop to also 
display GTK figures, but I've had no luck with this until now.

Fernando, I've been looking again at gui_thread. I now understand the 
beauty and power of it, but it is also more brittle, which means a lot 
of work and testing. It might prove the best solution in the end, though.
I took no offense at all about IPython, so neither should you :-)
I just mentioned it because it is the advised solution in matplotlib doc.

Have a nice day,
BC

PS: You don't have to CC me, I'm subscribed to the list.

Regarding Baptiste's patch
- I'm not too keen on the idea of wrapping gtk.gdk.threads_enter() /
gtk.gdk.threads_leave() around every GDK function call.
- I share Fernando's concern with possible problems when adding new
widgets or embedding a matplotlib window into something else.

My understanding is that wrapping gtk.gdk.threads_enter() /
gtk.gdk.threads_leave() around GDK functions is not necessary if all the
GDK calls occur in the same thread. So I think a simpler solution is to
arrange for all gui calls to go to a single thread and do away with
enter()/leave().

I created a module that allows interactive matplotlib access from a
python shell, its a modified version of
http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/65109

This method seems to work for me (on Linux) and is not very intrusive to
matplotlib code - it just requires that 'gui_cmd()' is used to execute
all gui commands when in interactive mode.

Any comments?

Regards
Steve

Steve Chaplin wrote:

> I created a module that allows interactive matplotlib access from a
> python shell, its a modified version of
> http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/65109
> 
> This method seems to work for me (on Linux) and is not very intrusive to
> matplotlib code - it just requires that 'gui_cmd()' is used to execute
> all gui commands when in interactive mode.
> 
> Any comments?

In case anyone is interested, here's a lightweight, standalone (non-ipython) 
GTK threaded interactive interpreter with matplotlib support.  This code is 
also based on the same ASPN recipe, it's basically the prototype I used to 
understand that recipe when I was adding the matplotlib support to IPython.

The threading tricks are the gold in that recipe, but the rest of the code is 
rather messy.  I pretty much rewrote it when studying it (while keeping the 
threading core).  Perhaps this may be of use to others.  The ASPN recipe 
duplicates a lot of the functionality in the code.py module from the standard 
library, which is completely unnecessary (it's guaranteed to be there in any 
python installation).  By using code.py, the whole thing can be rewritten much 
more cleanly (and you benefit from a good implementation in the stdlib for all 
the interpreter emulation parts).

I think I sent this to John a while back so he could replace the 
interactive.py shipped with matplotlib.  I fully acknowledge in the header the 
ASPN recipe as the origin of the threading solution.

The pylab support in current ipython is basically this same idea, but 
generalized to work with either GTK or WX, and blended into the IPython code 
structure (that wasn't too bad, since ipython itself is based on code.py).

Cheers,

f

On Wed, 2005-01-26 at 12:17 +0800, Steve Chaplin wrote:
> I created a module that allows interactive matplotlib access from a
> python shell, its a modified version of
> http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/65109
> 
> This method seems to work for me (on Linux) and is not very intrusive to
> matplotlib code - it just requires that 'gui_cmd()' is used to execute
> all gui commands when in interactive mode.

I may not have clearly explained my idea. I'm not proposing another
Python shell/interpreter to run matplotlib. The mt.py module is to test
an idea without needing to modify any matplotlib code. 

To use it as intended would involve moving it into matplotlib and
getting matplotlib to call gui_cmd() to execute interactive gtk/gdk
calls in their own thread.
This is the part I'm not sure about - what is the minimal set of
functions to wrap with gui_cmd() to ensure that it encloses all gtk/gdk
functions?
gui_cmd() could be called from draw_if_interactive() and maybe
new_figure_manager(), I'm not sure where else it might be needed.
Also it would need to be fed import commands.

Steve

Steve Chaplin a =E9crit :
> Regarding Baptiste's patch
> - I'm not too keen on the idea of wrapping gtk.gdk.threads_enter() /
> gtk.gdk.threads_leave() around every GDK function call.
> - I share Fernando's concern with possible problems when adding new
> widgets or embedding a matplotlib window into something else.
>=20
> My understanding is that wrapping gtk.gdk.threads_enter() /
> gtk.gdk.threads_leave() around GDK functions is not necessary if all th=
e
> GDK calls occur in the same thread. So I think a simpler solution is to
> arrange for all gui calls to go to a single thread and do away with
> enter()/leave().
>=20
This is mostly a question of taste.
A bug in the handling of threads_enter()/threads_leave() leads to a=20
deadlock, while a bug with a gui thread yields "Xlib async reply".=20
Choose which ugly death you prefer !

> I created a module that allows interactive matplotlib access from a
> python shell, its a modified version of
> http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/65109
>=20
In fact your module fits well into the framework of my proposal: all you=20
have to do is rename your gui_cmd to gtk_protect, and change a few=20
things like passing the function directly, passing the arguments, and=20
returning the value. Then you will have a drop-in replacement for=20
protect_mutex.py.

> This method seems to work for me (on Linux) and is not very intrusive t=
o
> matplotlib code - it just requires that 'gui_cmd()' is used to execute
> all gui commands when in interactive mode.
>=20
This is where the real beef is.
There are several strategies:
* wrap everything: this is the examples/interactive.py strategy, simple,=20
but the gui freezes on long calculations
* wrap all gui calls: this is the gui_thread strategy: elegant, but=20
difficult
* wrap only pylab functions (draw_if_interactive etc), as you seem to=20
suggest in your second email: imho this is not acceptable because it=20
breaks method calls of matplotlib objects
* draw the line somewhere in the backend: this is what I proposed.

In the end, the backends developpers have the final say, because they=20
will have to do the day-to-day maintenance.

> Any comments?
>=20
> Regards
> Steve
>=20
Cheers, BC

PS: tell me if I need to CC you or not

On Tue, 2005-01-25 at 22:25 -0700, Fernando Perez wrote:
> In case anyone is interested, here's a lightweight, standalone (non-ipython) 
> GTK threaded interactive interpreter with matplotlib support.  This code is 
> also based on the same ASPN recipe, it's basically the prototype I used to 
> understand that recipe when I was adding the matplotlib support to IPython.
Thanks, its a good example of code.InteractiveConsole, when I saw codeop in 
the ASPN recipe I wondered why code.py was not used.

I noticed a few small points:
- GTKInterpreter says "Run a gtk mainloop() in a separate thread.", yet
gtk.main() runs in the main thread, and the interpreter/console runs in
a separate thread.
- gtk.timeout_add() is now deprecated in favour of gobject.timeout_add()

- runsource() and runcode() are synchronised using wait()/notify() so
they deal with one code object at a time. If self.code_to_run is changed
to a code queue (using list.append() and list.pop(0)) then the wait
()/notify() could be removed and the interpreter thread would return
immediately it adds the code to self.code_to_run and not when the main
thread has executed the code.

> I think I sent this to John a while back so he could replace the 
> interactive.py shipped with matplotlib.  I fully acknowledge in the header the 
> ASPN recipe as the origin of the threading solution.

John, this looks like an much improved interactive.py, is it OK to use
it to replace interactive.py in cvs, and if so should the '-pylab'
option be left as it is or removed to be compatible with the current
interactive.py?

Steve

Steve Chaplin wrote:
> On Tue, 2005-01-25 at 22:25 -0700, Fernando Perez wrote:
> 
>>In case anyone is interested, here's a lightweight, standalone (non-ipython) 
>>GTK threaded interactive interpreter with matplotlib support.  This code is 
>>also based on the same ASPN recipe, it's basically the prototype I used to 
>>understand that recipe when I was adding the matplotlib support to IPython.
> 
> Thanks, its a good example of code.InteractiveConsole, when I saw codeop in 
> the ASPN recipe I wondered why code.py was not used.
> 
> I noticed a few small points:
> - GTKInterpreter says "Run a gtk mainloop() in a separate thread.", yet
> gtk.main() runs in the main thread, and the interpreter/console runs in
> a separate thread.
> - gtk.timeout_add() is now deprecated in favour of gobject.timeout_add()
> 
> - runsource() and runcode() are synchronised using wait()/notify() so
> they deal with one code object at a time. If self.code_to_run is changed
> to a code queue (using list.append() and list.pop(0)) then the wait
> ()/notify() could be removed and the interpreter thread would return
> immediately it adds the code to self.code_to_run and not when the main
> thread has executed the code.

I am _really_ ignorant of thread issues, and so I have avoided touching the 
threading part of this code at all, though I've always suspected it could be 
cleaner.  Would you mind applying these fixes and sending it back my way 
(including the GTK stuff)?  I could then port your improvements over to the 
code in ipython proper which implements this same functionality.

Cheers,

f

>>>>> "Steve" == Steve Chaplin <stevech1097@...> writes:
    Steve> John, this looks like an much improved interactive.py, is
    Steve> it OK to use it to replace interactive.py in cvs, and if so
    Steve> should the '-pylab' option be left as it is or removed to
    Steve> be compatible with the current interactive.py?

Hey Steve,

I probably just forgot to update interactive.py last time Fernando
sent it to me.  Feel free to do update interactive.py and its
interface and command line args (eg, pylab or not) as you see fit.
Just update the changelog so I can include them in the next release
notes.  

After you're done, you may want to take a look at interactive.html
and/or \section{Interactive} of the user's guide, and make updates as
you see fit.

Thanks!
JDH

Steve Chaplin wrote:
> On Tue, 2005-01-25 at 22:25 -0700, Fernando Perez wrote:
> 
>>In case anyone is interested, here's a lightweight, standalone (non-ipython) 
>>GTK threaded interactive interpreter with matplotlib support.  This code is 
>>also based on the same ASPN recipe, it's basically the prototype I used to 
>>understand that recipe when I was adding the matplotlib support to IPython.
> 
> Thanks, its a good example of code.InteractiveConsole, when I saw codeop in 
> the ASPN recipe I wondered why code.py was not used.
> 
> I noticed a few small points:
> - GTKInterpreter says "Run a gtk mainloop() in a separate thread.", yet
> gtk.main() runs in the main thread, and the interpreter/console runs in
> a separate thread.

You are correct.

> - gtk.timeout_add() is now deprecated in favour of gobject.timeout_add()

Feel free to fix it, I didn't know better.

> - runsource() and runcode() are synchronised using wait()/notify() so
> they deal with one code object at a time. If self.code_to_run is changed
> to a code queue (using list.append() and list.pop(0)) then the wait
> ()/notify() could be removed and the interpreter thread would return
> immediately it adds the code to self.code_to_run and not when the main
> thread has executed the code.

On second thought, I don't think this is a good idea in general.  The problem 
I see is the serialization of output.  Imagine you start a long running job 
and you get the prompt back right away.  If your job prints to stdout along 
the way (as is very common with scientific codes), your input line will get 
clobbered by this output.

If you could run in a gui window where you could associate the stdout for each 
input 'cell' with its own output one, it would be different.  This is exactly 
how Mathematica notebooks work, they stack a queue of cells for running 
sequentially, you get your typing prompt back immediately, but the output of 
each command is collected in an Out[] cell next to the corresponding In[] one. 
  In a plain terminal, I don't see how we can make this work (in fact, in a 
terminal Mathematica also serializes output for this very reason).

I have been toying with the idea of adding a way to 'background' processes 
automatically to ipython, by creating a list of worker threads for long 
running jobs.  But the problem of handling stdout has me stumped.

I'm open to comments, though, or perhaps I'm missing something obvious here.

Best,

f

On Thu, 2005-01-27 at 11:43 -0600, John Hunter wrote:
> I probably just forgot to update interactive.py last time Fernando
> sent it to me.  Feel free to do update interactive.py and its
> interface and command line args (eg, pylab or not) as you see fit.
> Just update the changelog so I can include them in the next release
> notes.  
> 
> After you're done, you may want to take a look at interactive.html
> and/or \section{Interactive} of the user's guide, and make updates as
> you see fit.
I replaced examples/interactive.py with Fernando's updated version
(adding a comment that gtk.main runs in the main thread and replacing
gtk.timeout_add with gobject.timeout_add).

I looked at interactive.html.template and user_guide.tex and saw a
reference to ipython, but no mention of interactive.py, so it does not
look like they need updating. I seem to recall interactive.py being
mentioned in the docs in the past, but they must have been changed since
then.

Steve

On Thu, 2005-01-27 at 17:14 -0700, Fernando Perez wrote: 
> > - runsource() and runcode() are synchronised using wait()/notify() so
> > they deal with one code object at a time. If self.code_to_run is changed
> > to a code queue (using list.append() and list.pop(0)) then the wait
> > ()/notify() could be removed and the interpreter thread would return
> > immediately it adds the code to self.code_to_run and not when the main
> > thread has executed the code.
> 
> On second thought, I don't think this is a good idea in general.  The problem 
> I see is the serialization of output.  Imagine you start a long running job 
> and you get the prompt back right away.  If your job prints to stdout along 
> the way (as is very common with scientific codes), your input line will get 
> clobbered by this output.
> 
> If you could run in a gui window where you could associate the stdout for each 
> input 'cell' with its own output one, it would be different.  This is exactly 
> how Mathematica notebooks work, they stack a queue of cells for running 
> sequentially, you get your typing prompt back immediately, but the output of 
> each command is collected in an Out[] cell next to the corresponding In[] one. 
>   In a plain terminal, I don't see how we can make this work (in fact, in a 
> terminal Mathematica also serializes output for this very reason).
> 
> I have been toying with the idea of adding a way to 'background' processes 
> automatically to ipython, by creating a list of worker threads for long 
> running jobs.  But the problem of handling stdout has me stumped.
> 
> I'm open to comments, though, or perhaps I'm missing something obvious here.

So removing wait()/notify() would not be such a good idea, it might be
useful if you are only running gui code, but not when the command prints
to stdout.

To run a background process as a thread you would need to redirect its
stdout somewhere (StringIO for example). The problem with using threads
for this is that they share the stdout with the main thread, so if you
redirect the background threads stdout you will redirect the python
interpreters stdout too!
You could run a background process as an actual process and not a thread
(and Python 2.4 has a new subprocess module which unifies the previous
solutions of os.system, os.spawn*, os.popen*, popen2.*, commands.*). It
lets you redirect stdout, but is used for running executables, not
calling python functions.

It has me stumped too, how can you call a python function and redirect
its stdout independent of the python interpreters stdout?

Steve

On Wed, 2005-01-26 at 22:18 +0100, Baptiste Carvello wrote:
I've finally got round installing the interactive gui thread patch to
have a look at it. It worked quite well for a few simple plots I tried.

> A bug in the handling of threads_enter()/threads_leave() leads to a 
> deadlock, while a bug with a gui thread yields "Xlib async reply". 
> Choose which ugly death you prefer !
What situation would yield an "Xlib async reply"?
 
> There are several strategies:
...
> * wrap only pylab functions (draw_if_interactive etc), as you seem to 
> suggest in your second email: imho this is not acceptable because it 
> breaks method calls of matplotlib objects
How does it break method calls of matplotlib objects? In interactive
mode the gtk backend does not work (by itself) so there's nothing to
break, and in non-interactive mode the gui_cmd() function would not be
used.

When thinking about using threads for interactive mode I had been
assuming people using the interactive mode would be using the pylab
interface, does anyone use the matplotlib class interface in interactive
mode?

I have a few questions about patch itself:
In import_gtk() there is time.sleep(1) - what is that for?

Shouldn't widget creation be protected as well by wrapping things like
new_figure_manager() and FigureCanvasGTK(). I know __init__ is wrapped
but objects may have __new__ methods which call gdk/gtk functions before
__init__ is called.

How about if a widgets standard methods such as show(), present() etc
are used - they are not protected (unless called from a protected
method). I think this just applies to using the class interface in
interactive mode, so it may be rarely encountered but might be a
problem.

Why use gtk_protect(), exec() and GTK_TEMPLATE, when you could use:

def gtk_protect (fun):
    def decorator (*args, **kwargs):
	if IN_GTK or threading.currentThread() is Thread:
            return fun(*args, **kwargs)
        IN_GTK.append (fun)
        gtk.threads_enter()
        try:
            res = fun(*args, **kwargs)
        finally:
            gtk.threads_leave()
            IN_GTK.pop()
            return res   
    return decorator

Steve

Steve Chaplin wrote:

> To run a background process as a thread you would need to redirect its
> stdout somewhere (StringIO for example). The problem with using threads
> for this is that they share the stdout with the main thread, so if you
> redirect the background threads stdout you will redirect the python
> interpreters stdout too!
> You could run a background process as an actual process and not a thread
> (and Python 2.4 has a new subprocess module which unifies the previous
> solutions of os.system, os.spawn*, os.popen*, popen2.*, commands.*). It
> lets you redirect stdout, but is used for running executables, not
> calling python functions.
> 
> It has me stumped too, how can you call a python function and redirect
> its stdout independent of the python interpreters stdout?

Not as far as I know.  That's why I think the only solution for the whole 
'backgrounding' problem with stdout is to put everything into a gui window (a 
la mathematica notebooks).  I've been thinking about it, and the proper way to 
write it is probably with _two_ python processes.  One would be the gui 
environment, and the other would be just a 'kernel' executing commands.  The 
stdout of the kernel would then be connected to an object capable of putting 
the results of each command next to the input cell which generated them.  I 
bet that's pretty much how Mathematica is organized, simply because it looks 
like the only reasonable way to deal with the issue.

Cheers,

f

Hi,

Steve Chaplin a =E9crit :
> What situation would yield an "Xlib async reply"?
> =20
using your mt module:
   >>> import mt
   >>> mt.gui_cmd('print "OK"')
   OK
   >>> import matplotlib
   >>> matplotlib.interactive(1)
   >>> import pylab
   Xlib: unexpected async reply (sequence 0x4b)!

This is on linux, I don't know what you get on other platforms, but I=20
guess it's just as bad.

> How does it break method calls of matplotlib objects? In interactive
> mode the gtk backend does not work (by itself) so there's nothing to
> break
Well, it depends on how ambitious you are :-)

> When thinking about using threads for interactive mode I had been
> assuming people using the interactive mode would be using the pylab
> interface, does anyone use the matplotlib class interface in interactiv=
e
> mode?
>=20
I sometimes do. However, I've painted the situation worse than it is. I=20
almost only use the setter/getter methods, which are not dangerous.

One annoying thing though: error_msg will hang the gui:
   >>> import mt
   >>> mt.gui_cmd('from pylab import *')
   >>> mt.gui_cmd('plot([1,2])')
   >>> mt.gui_cmd('plot([1,2],"0r")')
   *** gui hangs here, you get the prompt back, but no gui commands will=20

       work ***

> In import_gtk() there is time.sleep(1) - what is that for?
>=20
This is an attempt to avoid a race condition were a proxy function would=20
be called before the GTK thread is initialized. I don't know if this=20
would lead to a real problem. 1s wait not that bad anyway.

> Shouldn't widget creation be protected as well by wrapping things like
> new_figure_manager() and FigureCanvasGTK(). I know __init__ is wrapped
> but objects may have __new__ methods which call gdk/gtk functions befor=
e
> __init__ is called.
>=20
I didn't think of that. Most GTK object are not dangerous until they are=20
shown.

> How about if a widgets standard methods such as show(), present() etc
> are used - they are not protected (unless called from a protected
> method). I think this just applies to using the class interface in
> interactive mode, so it may be rarely encountered but might be a
> problem.
>=20
As I stated, I didn't try to protect methods inherited from GTK objects=20
(ie not defined in matplotlib). Which means people using them are on=20
their own. You have to draw the boundary somewhere, unless you want to=20
implement a full gui_thread.

> Why use gtk_protect(), exec() and GTK_TEMPLATE, when you could use:
>=20
> def gtk_protect (fun):
>     def decorator (*args, **kwargs):
>
Well this was a try to set the proxy function name as the real function=20
name, in order to avoid weirdeness when reading object help. But I could=20
not set the argument spec correctly, so I ended up adding the 'proxy=20
from' line to the docstring. So I think all the GTK_TEMPLATE machinery=20
serves no purpose now.

As a final note, I have to say that the latest version of interactive.py=20
is good enough for my needs, so I'm no longer very motivated for pushing=20
this patch.

If you just need the pylab interface, you can also do it without=20
patching matplotlib. Just create a protected_pylab.py module, where you=20
would start the gui thread, define the gtk_protect function, import=20
pylab and define command=3Dgtk_protect(pylab.command) for every command i=
n=20
pylab. Then doing 'from protected_pylab import *' would give you a=20
working pylab interface, with just the error_msg problem.

Cheers,
Baptiste

Baptiste Carvello wrote:
> Hi,
>=20
> Steve Chaplin a =E9crit :

>>When thinking about using threads for interactive mode I had been
>>assuming people using the interactive mode would be using the pylab
>>interface, does anyone use the matplotlib class interface in interactiv=
e
>>mode?
>>
>=20
> I sometimes do.

Agreed.

One should also be able to execute interactively large scripts which may =
have=20
  been written to the OO interface.  I know that I routinely test 10k LOC=
=20
modules from within ipython (though not matplotlib based).  It would be I=
MHO=20
poor design to put all sorts of special-case warnings, where certain feat=
ures=20
of matplotlib would only be available in certain ways of using it.  Speci=
al=20
cases lead to madness...

Note that I haven't followed the details close enough to vouch for any=20
particular approach.  I just would like to see any solution implemented t=
o be=20
as clean and general as possible.  As far as users are concerned, these=20
details should be, where possible, totally irrelevant (internal implement=
ation=20
only).  As a reference, ipython tries mightily hard to make sure that _an=
y_=20
code which can be executed at a system command line as 'python foo.py'=20
produces identical behavior at an ipython prompt via 'run foo'.  There ar=
e=20
special options to 'run' for modifying this, but the default attempts to=20
emulate python very faitfully, just with better tracebacks, debug support=
,=20
etc.  And I consider deviations from this bevavior as bugs.  It's not 100=
%=20
perfect, but it really works fairly well in most cases.

Regards,

f