matplotlib-devel Mailing List for matplotlib

Tim,
Based on a *very superficial* quick look, I have two comments:
1) Since the plan is (or was?) to have one more release before 
specializing to numpy-only support, the ".T" won't work at present.
2) In the following (possibly mangled by mailer),
 > + if Nsizes == 1:
 > + xy = asarray([xverts, yverts]).T * sizes[0]
 > + polys = [xy + offset for offset in 
transOffset.seq_xy_tups(self._offsets)]
I think you can gain additional speedup by eliminating the list 
comprehension. Instead of using seq_xy_tups you can use numerix_xy to 
return a 2-D array (with columns x and y), and simply add that to xy.
Eric
Tim Leslie wrote:
> Hi All,
> 
> I've attached a patch which optimizes a particular case of the
> RegularPolyCollection.get_transformed_patches method. This is
> particularly useful when using a picker on a scatter plot with ~40,000
> points (as I happen to be doing) and cuts the time spent in this
> function from ~4s to ~1s, which accounts for about 50% of the lag in
> my particular use case.
> 
> Similar improvements might be possible in the N > 1 case, but I don't
> have a) a data set or b) the time to work them out with, so hopefully
> someone is happy to check in this patch as it stands, as I think the
> N==1 is probably the common case.
> 
> If there are any problems with the patch let me know and I'll try to
> fix it up. It's against latest svn (3262)
> 
> Cheers,
> 
> Tim
> 
> 
> ------------------------------------------------------------------------
> 
> Index: lib/matplotlib/collections.py
> ===================================================================
> --- lib/matplotlib/collections.py	(revision 3262)
> +++ lib/matplotlib/collections.py	(working copy)
> @@ -153,7 +153,7 @@
> if not self.pickable(): return
> ind = []
> x, y = mouseevent.x, mouseevent.y
> - for i, thispoly in enumerate(self.get_transformed_patches()): 
> + for i, thispoly in enumerate(self.get_transformed_patches()):
> inside = nxutils.pnpoly(x, y, thispoly)
> if inside: ind.append(i)
> if len(ind):
> @@ -167,7 +167,6 @@
> The ith element in the returned sequence is a list of x,y
> vertices defining the ith polygon
> """
> -
> verts = self._verts
> offsets = self._offsets
> usingOffsets = offsets is not None
> @@ -451,13 +450,19 @@
> __init__.__doc__ = dedent(__init__.__doc__) % kwdocd
> 
> def get_transformed_patches(self):
> - 
> +
> xverts, yverts = zip(*self._verts)
> xverts = asarray(xverts)
> yverts = asarray(yverts)
> sizes = sqrt(asarray(self._sizes)*self._dpi.get()/72.0)
> Nsizes = len(sizes)
> transOffset = self.get_transoffset()
> +
> + # optimize this case for an approx 4x speedup
> + if Nsizes == 1:
> + xy = asarray([xverts, yverts]).T * sizes[0]
> + polys = [xy + offset for offset in transOffset.seq_xy_tups(self._offsets)]
> +
> polys = []
> for i, loc in enumerate(self._offsets):
> xo,yo = transOffset.xy_tup(loc)
> 
> 
> ------------------------------------------------------------------------
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by DB2 Express
> Download DB2 Express C - the FREE version of DB2 express and take
> control of your XML. No limits. Just data. Click to get it now.
> http://sourceforge.net/powerbar/db2/
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Matplotlib-devel mailing list
> Mat...@li...
> https://lists.sourceforge.net/lists/listinfo/matplotlib-devel

Hi All,
I've attached a patch which optimizes a particular case of the
RegularPolyCollection.get_transformed_patches method. This is
particularly useful when using a picker on a scatter plot with ~40,000
points (as I happen to be doing) and cuts the time spent in this
function from ~4s to ~1s, which accounts for about 50% of the lag in
my particular use case.
Similar improvements might be possible in the N > 1 case, but I don't
have a) a data set or b) the time to work them out with, so hopefully
someone is happy to check in this patch as it stands, as I think the
N==1 is probably the common case.
If there are any problems with the patch let me know and I'll try to
fix it up. It's against latest svn (3262)
Cheers,
Tim