One of the biggest causes of controversy in mpl, and of difficulty in
teaching and learning mpl, is the divide between pyplot and the rest of
the library. There are at least two aspects:
1) plt.title() versus ax.set_title(), etc; that is, all the clunky
getters and setters on the OO side. JDH used to note that they were a
side-effect of his C++ heritage, and that he probably wouldn't have used
them if he had had more Python experience when he started mpl.
2) For interactive use, such as in the ipython console, one really wants
pyplot's draw_if_interactive() functionality; one doesn't want to have
to type explicit plt.draw() commands. Worse, plt.draw() operates on the
current figure, which might not be the figure that one just updated with
"ax2.set_title('the other plot')".
I think that both of these speed bumps can be removed fairly easily, in
an entirely backwards-compatible way.
The first is just a matter of propagating some shorter-form pyplot
function names back to Axes and Figure. This idea is mentioned at the
end of MEP 13, as an alternative to properties.
The second requires accepting some behavior in the Axes and Figure
classes that is conditional on the backend and the interactive state. I
think it would look *roughly* like this:
Add a method to Figure:
def draw_if_interactive():
if not is_interactive:
return
if not isinstance(self.canvas, interactive_canvases):
return
self.canvas.draw()
Append this method to suitable Figure methods such as suptitle().
Add a method to Axes:
def draw_if_interactive():
self.figure.draw_if_interactive()
Append this method to suitable Axes methods--all those that execute
changes, or at least those with corresponding pyplot functions.
Some additional logic (either a kwarg, or temporary manipulation of the
"interactive" flag, or of an Axes instance flag) would be needed to
block the drawing at intermediate stages--e.g., when boxplot is drawing
all its bits and pieces.
After these changes, the pyplot functions could be simplified; they
would not need their own draw_if_interactive calls.
Am I missing some major impediment? If not, I think this set of changes
would make it much easier to use and teach the OO interface, with pyplot
still being used where it is most helpful, such as in the subplots() call.
Eric
I am against pushing the pyplot style title/xlabel/.. function down
into the OO layer, I really do not like the different behaviour and
returns depending on the arguments. That has always struck me as a
MATLAB-ism that should be dropped, but we are stuck with to maintain
back-compatibility.
I have been thinking about going a very different route and pulling
almost all of the plotting function _off_ of the axes objects and just
having functions with signatures like
def plotter_function(ax, data1, data2, style1, style2,...)
art = create_artists(...)
ax.add_artists(art)
return art_list
And then almost all of pyplot can be replaced with a wrapper function:
def wrap_for_pyplot(func):
def inner(*args, **kwargs)
ax = plt.gca()
art_list = func(ax, *args, **kwargs)
if plt.is_interactive():
ax.figure.canvas.draw()
inner.__name__ = func.__name__
inner.__doc__ = strip_ax_arg(func.__doc__)
return inner
for f in funcs_to_wrap:
pyplot.setattr(f.__name__, wrap_for_pyplot(f))
Which pushes all of the interactive/global state related stuff up to
one place and removes the need for keywords to suppress re-drawing/the
need to manage that. This will make embedding a lot easier as well.
I have also been thinking quite a bit about the semantic
artist/manager layer of objects which I think would also go a long way
making the library easier to use, but that is a different story.
Tom
On Sat, Sep 27, 2014 at 7:40 PM, Eric Firing <ef...@ha...> wrote:
> One of the biggest causes of controversy in mpl, and of difficulty in
> teaching and learning mpl, is the divide between pyplot and the rest of
> the library. There are at least two aspects:
>
> 1) plt.title() versus ax.set_title(), etc; that is, all the clunky
> getters and setters on the OO side. JDH used to note that they were a
> side-effect of his C++ heritage, and that he probably wouldn't have used
> them if he had had more Python experience when he started mpl.
>
> 2) For interactive use, such as in the ipython console, one really wants
> pyplot's draw_if_interactive() functionality; one doesn't want to have
> to type explicit plt.draw() commands. Worse, plt.draw() operates on the
> current figure, which might not be the figure that one just updated with
> "ax2.set_title('the other plot')".
>
> I think that both of these speed bumps can be removed fairly easily, in
> an entirely backwards-compatible way.
>
> The first is just a matter of propagating some shorter-form pyplot
> function names back to Axes and Figure. This idea is mentioned at the
> end of MEP 13, as an alternative to properties.
>
> The second requires accepting some behavior in the Axes and Figure
> classes that is conditional on the backend and the interactive state. I
> think it would look *roughly* like this:
>
> Add a method to Figure:
>
> def draw_if_interactive():
> if not is_interactive:
> return
> if not isinstance(self.canvas, interactive_canvases):
> return
> self.canvas.draw()
>
> Append this method to suitable Figure methods such as suptitle().
>
> Add a method to Axes:
>
> def draw_if_interactive():
> self.figure.draw_if_interactive()
>
> Append this method to suitable Axes methods--all those that execute
> changes, or at least those with corresponding pyplot functions.
>
> Some additional logic (either a kwarg, or temporary manipulation of the
> "interactive" flag, or of an Axes instance flag) would be needed to
> block the drawing at intermediate stages--e.g., when boxplot is drawing
> all its bits and pieces.
>
> After these changes, the pyplot functions could be simplified; they
> would not need their own draw_if_interactive calls.
>
> Am I missing some major impediment? If not, I think this set of changes
> would make it much easier to use and teach the OO interface, with pyplot
> still being used where it is most helpful, such as in the subplots() call.
>
> Eric
>
> ------------------------------------------------------------------------------
> Meet PCI DSS 3.0 Compliance Requirements with EventLog Analyzer
> Achieve PCI DSS 3.0 Compliant Status with Out-of-the-box PCI DSS Reports
> Are you Audit-Ready for PCI DSS 3.0 Compliance? Download White paper
> Comply to PCI DSS 3.0 Requirement 10 and 11.5 with EventLog Analyzer
> http://pubads.g.doubleclick.net/gampad/clk?id=154622311&iu=/4140/ostg.clktrk
> _______________________________________________
> Matplotlib-devel mailing list
> Mat...@li...
> https://lists.sourceforge.net/lists/listinfo/matplotlib-devel
--
Thomas Caswell
tca...@gm...
On 2014年09月28日, 6:20 AM, Thomas Caswell wrote:
> I am against pushing the pyplot style title/xlabel/.. function down
> into the OO layer, I really do not like the different behaviour and
> returns depending on the arguments. That has always struck me as a
> MATLAB-ism that should be dropped, but we are stuck with to maintain
> back-compatibility.
I don't understand your objection. In which cases do the returns depend
on the arguments? "title" returns the title text object, regardless of
whether there was an argument setting it to a new value. Same for
"xlim". I haven't checked the whole list, but I expect this is the
general pattern.
As for different behavior depending on arguments, what specifically do
you object to, for example? I think we have all that nicely and
pythonically implemented via kwargs, so you can use "xlim(xmin=3)", for
example. Yes, that is different behavior than "xlim(3, 6)", but in a
good way.
Maybe you are referring to the variety of signatures, such as for
contour. I don't really like that either; but it is already in the OO
layer, not just the pyplot layer. (The python builtin "slice()" has
this characteristic, and I don't like it there, either.)
>
> I have been thinking about going a very different route and pulling
> almost all of the plotting function _off_ of the axes objects and just
> having functions with signatures like
>
> def plotter_function(ax, data1, data2, style1, style2,...)
> art = create_artists(...)
> ax.add_artists(art)
> return art_list
This has occurred to me also--I have never particularly liked having
such an enormous number of Axes methods.
There is one major difference in using methods instead of plotter
functions, though: it allows subclassing. Whether this is ever used in
practice, I don't know.
>
> And then almost all of pyplot can be replaced with a wrapper function:
>
> def wrap_for_pyplot(func):
> def inner(*args, **kwargs)
> ax = plt.gca()
> art_list = func(ax, *args, **kwargs)
> if plt.is_interactive():
> ax.figure.canvas.draw()
>
> inner.__name__ = func.__name__
> inner.__doc__ = strip_ax_arg(func.__doc__)
> return inner
>
> for f in funcs_to_wrap:
> pyplot.setattr(f.__name__, wrap_for_pyplot(f))
>
> Which pushes all of the interactive/global state related stuff up to
> one place and removes the need for keywords to suppress re-drawing/the
> need to manage that. This will make embedding a lot easier as well.
But it does *not* take care of one of the two *big* problems I am
talking about: the lack of automatic interactive plot updating when one
wants to explicitly specify axes and figures in the plot call,
regardless of whether this is done via plotting functions or methods.
>
> I have also been thinking quite a bit about the semantic
> artist/manager layer of objects which I think would also go a long way
> making the library easier to use, but that is a different story.
I still don't really understand it, but perhaps it is orthogonal to the
issues I am raising here. As far as I can see, your proposals above do
not address either of the issues I raised, based on experience both in
teaching matplotlib to non-programmers, and in using it day-to-day.
Regarding Matlab: it is justly popular for many reasons. It is
relatively easy to learn both by design and because of its consistent
high-quality documentation. Matplotlib's success has resulted in large
measure from its pyplot layer, which can shield learners and users from
mpl's complexity, which allows learners to build on their Matlab
knowledge, and which is particularly well suited to quick interactive
data exploration. The problem with the Matlab/pyplot approach is that
it doesn't scale well, so we see a chorus of advice along the lines of
"don't use pyplot except for subplots() and show(); use the nice,
explicit OO interface for everything else". But at present, this
doesn't work well, because the OO approach is not interactive enough,
and using the getters and setters is clumsy when typing at the
console--in demonstrating, teaching, learning, and exploring
interactively, every keystroke counts!
Eric
>
> Tom
>
>
> On Sat, Sep 27, 2014 at 7:40 PM, Eric Firing <ef...@ha...> wrote:
>> One of the biggest causes of controversy in mpl, and of difficulty in
>> teaching and learning mpl, is the divide between pyplot and the rest of
>> the library. There are at least two aspects:
>>
>> 1) plt.title() versus ax.set_title(), etc; that is, all the clunky
>> getters and setters on the OO side. JDH used to note that they were a
>> side-effect of his C++ heritage, and that he probably wouldn't have used
>> them if he had had more Python experience when he started mpl.
>>
>> 2) For interactive use, such as in the ipython console, one really wants
>> pyplot's draw_if_interactive() functionality; one doesn't want to have
>> to type explicit plt.draw() commands. Worse, plt.draw() operates on the
>> current figure, which might not be the figure that one just updated with
>> "ax2.set_title('the other plot')".
>>
>> I think that both of these speed bumps can be removed fairly easily, in
>> an entirely backwards-compatible way.
>>
>> The first is just a matter of propagating some shorter-form pyplot
>> function names back to Axes and Figure. This idea is mentioned at the
>> end of MEP 13, as an alternative to properties.
>>
>> The second requires accepting some behavior in the Axes and Figure
>> classes that is conditional on the backend and the interactive state. I
>> think it would look *roughly* like this:
>>
>> Add a method to Figure:
>>
>> def draw_if_interactive():
>> if not is_interactive:
>> return
>> if not isinstance(self.canvas, interactive_canvases):
>> return
>> self.canvas.draw()
>>
>> Append this method to suitable Figure methods such as suptitle().
>>
>> Add a method to Axes:
>>
>> def draw_if_interactive():
>> self.figure.draw_if_interactive()
>>
>> Append this method to suitable Axes methods--all those that execute
>> changes, or at least those with corresponding pyplot functions.
>>
>> Some additional logic (either a kwarg, or temporary manipulation of the
>> "interactive" flag, or of an Axes instance flag) would be needed to
>> block the drawing at intermediate stages--e.g., when boxplot is drawing
>> all its bits and pieces.
>>
>> After these changes, the pyplot functions could be simplified; they
>> would not need their own draw_if_interactive calls.
>>
>> Am I missing some major impediment? If not, I think this set of changes
>> would make it much easier to use and teach the OO interface, with pyplot
>> still being used where it is most helpful, such as in the subplots() call.
>>
>> Eric
>>
>> ------------------------------------------------------------------------------
>> Meet PCI DSS 3.0 Compliance Requirements with EventLog Analyzer
>> Achieve PCI DSS 3.0 Compliant Status with Out-of-the-box PCI DSS Reports
>> Are you Audit-Ready for PCI DSS 3.0 Compliance? Download White paper
>> Comply to PCI DSS 3.0 Requirement 10 and 11.5 with EventLog Analyzer
>> http://pubads.g.doubleclick.net/gampad/clk?id=154622311&iu=/4140/ostg.clktrk
>> _______________________________________________
>> Matplotlib-devel mailing list
>> Mat...@li...
>> https://lists.sourceforge.net/lists/listinfo/matplotlib-devel
>
>
>
On 2014年09月28日, 8:22 AM, Thomas Caswell wrote: > On Sun, Sep 28, 2014 at 1:52 PM, Eric Firing <ef...@ha...> wrote: >> >> I don't understand your objection. In which cases do the returns depend on >> the arguments? "title" returns the title text object, regardless of whether >> there was an argument setting it to a new value. Same for "xlim". I >> haven't checked the whole list, but I expect this is the general pattern. > > > Sorry, I rarely use that part of the API, I apparently remember it > wrong. I thought they returned `None` in the case where you were > using them as setters. > >> But it does *not* take care of one of the two *big* problems I am talking about: >> the lack of automatic interactive plot updating when one wants to explicitly >> specify axes and figures in the plot call, regardless of whether this is done >> via plotting functions or methods. > > I would suggest using a function identical to the pyplot wrapper, but > just preserve the ax argument: > > def interactive_wrapper(func): > @functools.wraps(func) > def inner(ax, *args, **kwargs): > ret_list = func(ax, *args, **kwargs) > ax.figure.canvas.draw() > return ret_list > > return inner Yes, this could be done. But it would require: 1) the switch from methods to plot functions, which is a huge change; 2) having not 1, not 2, but 3 sets of plot functions: the first with the ax argument and no interactivity; the second with the ax argument and interactivity; and the third being the pyplot version. > > I think it is very important to maintain a way to use mpl without > having to involve pyplot and it's global state in anyway which is > useful for embedding. This needs a closer look. I don't see that my suggestions involve pyplot's global state any more than at present, or that they would in any way affect embedding. > >> There is one major difference in using methods instead of plotter functions, though: >> it allows subclassing. Whether this is ever used in practice, I don't know > > I think this only matters if you want to over-ride the behavior of the > plotting methods which I am going back and forth on if preventing this > is a good thing or not and if monkey-patching/clever importing is the > way around it. > I tend to think that monkey-patching is something to avoid, and that the goal should be simplicity and straightforwardness rather than cleverness. Eric
Tom, Thanks for https://github.com/matplotlib/matplotlib/pull/3587, which provides a concise draft implementation of your proposal, in concept if not in detail. Although this provides a way to get interactive drawing while specifying axes and figures explicitly, it is still using two interfaces (your "interactive" function namespace plus the original OO methods) to do this instead of the single interface that I am suggesting (embedding the functionality in the OO methods). One disadvantage is that when code is developed interactively, using your "interactive" namespace, more translation is required when the user wants to take the ipython history and edit it down to a module in which the OO interface is used. Independently, your suggestion to use wrappers instead of boilerplate.py could be considered as a way to make pyplot.py shorter; historically, boilerplate.py was used because the versions of python available when mpl was being developed lacked the ability to do everything that needed to be done with wrappers. I don't remember exactly what the sticking point was, but definitely there was one. I still haven't seen a prohibitive objection to my original suggestion, which I think would involve minimal API change, and which directly addresses precisely the problems that I called out. My own primary reservation about it is that it adds some complexity at the OO level, and as I said earlier, as a general rule, I want to see complexity reduced, not increased. Then the question is, can the added complexity be kept down to a low enough level that it is an acceptable tradeoff for the increased consistency and simplicity from the user's standpoint? I'm not sure; but I think it deserves consideration. As a side note, there was an earlier attempt to address a different aspect of pyplot-OO convergence here: https://github.com/matplotlib/matplotlib/pull/1457. It didn't work out, but it was in response to genuine pull from users. Notice also the last paragraph of @pelson's comment in that PR--a testament to the real-world advantages of pyplot. Eric
On Sun, Sep 28, 2014 at 7:52 PM, Eric Firing <ef...@ha...> wrote: > Regarding Matlab: it is justly popular for many reasons. It is > relatively easy to learn both by design and because of its consistent > high-quality documentation. Matplotlib's success has resulted in large > measure from its pyplot layer, which can shield learners and users from > mpl's complexity, which allows learners to build on their Matlab > knowledge, and which is particularly well suited to quick interactive > data exploration. The problem with the Matlab/pyplot approach is that > it doesn't scale well, so we see a chorus of advice along the lines of > "don't use pyplot except for subplots() and show(); use the nice, > explicit OO interface for everything else". But at present, this > doesn't work well, because the OO approach is not interactive enough, > and using the getters and setters is clumsy when typing at the > console--in demonstrating, teaching, learning, and exploring > interactively, every keystroke counts! > > Matlab is actually slowly trying to transition to an OO-style interface of their own. It has taken a LONG time, though.
One point which is often neglected while discussing oo- vs the pyplot-api is that today introspection tools quite often fail to work with mpl-oo but are perfectly fine with the pyplot module. E.g. if i am writing some kind of helper oder plotting function taking ax, one gets no auto-completion nor docstrings. I always have to use a interactive console to search for them or just type in the corresponding pyplot command for the function signature. Note that this is not matplotlibs fault, but another reason while beginners may prefer the pyplot-interface: Easier access to docstrings and available plotting functions. This is also why i am a bit wary of using properties in matplotlib and i don't think they are a good fit most of the time. Documenting them is hard, discoverablity is also worse. And most setter methods in mpl have very useful kwargs, something which is not doable with properties.
On 2014年09月28日, 12:42 PM, Till Stensitzki wrote: > One point which is often neglected while discussing > oo- vs the pyplot-api is that today introspection tools > quite often fail to work with mpl-oo but are > perfectly fine with the pyplot module. E.g. if i am writing > some kind of helper oder plotting function taking ax, one > gets no auto-completion nor docstrings. I always have to use > a interactive console to search for them or just type in > the corresponding pyplot command for the function > signature. I think I know what you mean here--I do the same thing. > Note that this is not matplotlibs fault, but > another reason while beginners may prefer the pyplot-interface: > Easier access to docstrings and available plotting functions. All of this sounds like an argument for considering a gradual move to plot functions rather than methods, as Tom is suggesting--correct? That is actually the way I implemented contour and quiver, and I have long thought that all of the moderately to very complicated plot functions, such as "plot", "hist", "boxplot", etc. should similarly be moved out of what used to be axes.py into modules containing related functionality; then it's easy to attach them as methods with a stub. > > This is also why i am a bit wary of using properties in matplotlib and > i don't think they are a good fit most of the time. Documenting > them is hard, discoverablity is also worse. And most setter methods in > mpl have very useful kwargs, something which is not doable with > properties. Good point. I think that one of the problems with the getters and setters, though, apart from clunky names, is that there is a conflict between what the name suggests--which is consistent with using them as the basis for properties--and expanding their functionality with kwargs. When functionality is expanded, a name like "set_something" probably is not the best description. That's why we have some convenience functions that can affect a variety of related parameters. Maybe we need more of those; and maybe this can be addressed using Tom's Controller idea, but I'm hazy about that. Eric > > > ------------------------------------------------------------------------------ > Meet PCI DSS 3.0 Compliance Requirements with EventLog Analyzer > Achieve PCI DSS 3.0 Compliant Status with Out-of-the-box PCI DSS Reports > Are you Audit-Ready for PCI DSS 3.0 Compliance? Download White paper > Comply to PCI DSS 3.0 Requirement 10 and 11.5 with EventLog Analyzer > http://pubads.g.doubleclick.net/gampad/clk?id=154622311&iu=/4140/ostg.clktrk > _______________________________________________ > Matplotlib-devel mailing list > Mat...@li... > https://lists.sourceforge.net/lists/listinfo/matplotlib-devel >
n Sun, Sep 28, 2014 at 12:40 AM, Eric Firing <ef...@ha...> wrote:
> One of the biggest causes of controversy in mpl, and of difficulty in
> teaching and learning mpl, is the divide between pyplot and the rest of
> the library. There are at least two aspects:
>
> 1) plt.title() versus ax.set_title(), etc; that is, all the clunky
> getters and setters on the OO side. JDH used to note that they were a
> side-effect of his C++ heritage, and that he probably wouldn't have used
> them if he had had more Python experience when he started mpl.
>
> 2) For interactive use, such as in the ipython console, one really wants
> pyplot's draw_if_interactive() functionality; one doesn't want to have
> to type explicit plt.draw() commands. Worse, plt.draw() operates on the
> current figure, which might not be the figure that one just updated with
> "ax2.set_title('the other plot')".
I'm not very familiar with matplotlib's guts, but as a user I agree
100%. I'm sure there are Reasons why 90% of pylot functions aren't
simply:
def foo(*args, **kwargs):
return gca().foo(*args, **kwargs)
but I have no idea what they are, so whenever I have to memorize two
different APIs for doing the same thing it ends up feeling like a
pointless waste of time.
> I think that both of these speed bumps can be removed fairly easily, in
> an entirely backwards-compatible way.
>
> The first is just a matter of propagating some shorter-form pyplot
> function names back to Axes and Figure. This idea is mentioned at the
> end of MEP 13, as an alternative to properties.
I'd much rather write ax.xlim(...) than ax.xlim = .... The use of that
many magical properties feels unpythonic to me (too much magic!), and
pyplot.xlim() isn't going anywhere. So, it would still mean that
everyone has to learn both the pyplot and the OO APIs separately.
Learning 1 API is always going to be easier than learning two
different APIs, no matter how fancy and polished you make the second
one.
> The second requires accepting some behavior in the Axes and Figure
> classes that is conditional on the backend and the interactive state. I
> think it would look *roughly* like this:
>
> Add a method to Figure:
>
> def draw_if_interactive():
> if not is_interactive:
> return
> if not isinstance(self.canvas, interactive_canvases):
> return
> self.canvas.draw()
>
> Append this method to suitable Figure methods such as suptitle().
>
> Add a method to Axes:
>
> def draw_if_interactive():
> self.figure.draw_if_interactive()
>
> Append this method to suitable Axes methods--all those that execute
> changes, or at least those with corresponding pyplot functions.
In most scene-graphy frameworks, triggering redraws is simply not the
job of the mutator; it's the job of the rendering system to observe
the model for changes. (Using "observe" in the "observer pattern"
sense.) In this approach, every time an artist changes you'd call some
"note_change()" method, which would then propagate upwards. And then
some top-level piece of logic (the canvas or backend, I guess?) would
be responsible for deciding whether it wanted to respond to these
changes. So... very similar to what you're proposing, though with
different terminology :-).
> Some additional logic (either a kwarg, or temporary manipulation of the
> "interactive" flag, or of an Axes instance flag) would be needed to
> block the drawing at intermediate stages--e.g., when boxplot is drawing
> all its bits and pieces.
Presumably this would be a context manager, 'with artist.hold_changes(): ...'
--
Nathaniel J. Smith
Postdoctoral researcher - Informatics - University of Edinburgh
http://vorpus.org