[yt-users] Problem when making a slice plot

[Matthew Turk]

Hi Markus,

On Sat, Sep 15, 2012 at 7:18 AM, Markus Haider <markus.haider at uibk.ac.at> wrote:
> Hi everyone,
>
> I am experiencing a strange problem when making a slice plot of a flash
> snapshot. From the first snapshot (which should be the initial condition) I
> plotted the temperature, which should just be the same value all over the
> simulation box.
>
> However, when I make a slice through the middle of the volume, this is how
> the result looks like, when yt automatically determines limits of the
> colorbar:
> http://ubuntuone.com/4I63O6KY7LDELHa0ZWvjem

Ah, yes.  This is a familiar problem -- the colorbar limits are so
tight as to not be visible on the colorbar itself!  What you're seeing
is actually the roundoff error in the antialiasing that yt applies,
which preferentially builds up at cell boundaries and in higher
refined areas.  I think the (relative) scatter should be on the order
of 1e-14 or so.  Since it's so small, it only shows up when *all* the
values are the same.  For reference, here's the source code:

https://bitbucket.org/yt_analysis/yt/src/tip/yt/visualization/_MPL.c#cl-143

The problem comes in during the calculation of the overlap factors,
which are meant as a simple means of anti-aliasing.  Please note, this
code is overdue for touching up; it's not only terse but a bit
obscure.  The overlap lines:

overlap2 = ((min(rypx, ysp+dysp) - max(lypx, (ysp-dysp)))*ipx_dy);
overlap1 = ((min(rxpx, xsp+dxsp) - max(lxpx, (xsp-dxsp)))*ipx_dx);
and then the value deposited is: (dsp*overlap1)*overlap2;

This is taking the overlap of a pixel with a cell in the y, then in
the x, and multiplying the two.  When there are multiply cells
overlapping a pixel, this becomes off by just a tiny bit.

You can disable this behavior by turning off antialiasing in the FRBs.
 Perhaps we should simply do this by default, although I think it will
lead to slightly less attractive plots in most cases.  This can be
done on a slice plot like:

slice_plot = SlicePlot( ... )
slice_plot.antialias = False
slice_plot._recreate_frb()

>
> If I specify the color-bar limits on my own, it looks like as it should be:
> http://ubuntuone.com/7E1YmEgKoV9XmR77Ddplhs
>
> Does someone know what is going wrong here?

Yup, in short -- the values are all correct to extremely tight
precision, but matplotlib will happily expose that tight precision
when it autosets colorbar limits.  I think one possibility we could
explore is actually having the SlicePlot / ProjectionPlot do a better
job of auto-calculating limits, and not letting matplotlib do it, when
values are quite tightly clustered.  One could imagine ensuring that
we get a bit of room on both sides, or perhaps not logging plots that
have < 1 OOM in spread.

Hope that helps -- let us know!

Matt

PS I just saw you had asked this in IRC -- hope to see you again soon!

>
> Here is the script I used to make the plot:
> http://ubuntuone.com/7icVc1nNDBaQfRRZBRTXEt
>
> And in case you want to reproduce the error, here is the snapshot file:
> http://ubuntuone.com/6wq4c6HALXdqyktHjWTtaB
>
> Thanks for your help,
> Markus
> _______________________________________________
> yt-users mailing list
> yt-users at lists.spacepope.org
> http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org