Bug in colorbar()

Community matplotlib-users
1

I figured out you can pass in the rasterized keyword to all of those to change the rasterization in the output.

Also the docs say for pcolormesh it defaults to the backend if not set. Therefore, in the case of a vector based it would output vectors if not set to rasterize.

Haven’t tested but curious. Lets say I want to output at 600dpi but I display images interactively at 100dpi. Does it always rasterize the image to the higher dpi? I had noticed this didn’t seem to occur in specgram but figured because the specgram is a relatively low resolution image that outputing at 600dpi doesn’t do anything because original image is already a low resolution. I would expect the other modes do do this where the image isn’t already output and have to rasterize the image when saving like pcolormesh and contour plots.

Cheers,

Jeff

···

On Sat, Jul 28, 2012 at 1:43 PM, Jeffrey Spencer <jeffspencerd@…2015…87…> wrote:

Yes, specgram rasterizes and contourf is definately a vector specification which isn’t optimal for 100 levels.

I would switch to pcolormesh but the output file can’t be rasterizing the image. It outputs a huge file in .pdf (40X bigger than equivalent .png) and it looks like it is vector based not rasterized.

Basically, If you output specgram or imshow in .pdf or .png the file sizes are relatively comparable with .pdf, .eps, .svg being slightly larger due to embedding the picture.

If I output in pcolor, pcolormesh, contourf (with more than 100 levels) the file sizes are huge for .pdf, .eps, .svg which I’m assuming because vector based output. It also looks like vector based output because can see the lines it draws for contours. Could this possibly be a selection for these outputs to force raster based processing or is that not easy.

On Sat, Jul 28, 2012 at 3:26 AM, Eric Firing <efiring@…202…> wrote:

On 2012/07/26 10:26 PM, Jeffrey Spencer wrote:

Thanks, that is all good info to know. I change my data to log and

normalize it so the logNorm is just linear actually so specifying only

levels is fine. I’ll let you know if that doesn’t work properly for some

reason.

Ok, yeah I looked at pcolormesh quickly and can’t remember why I chose

originally when I wrote this to go with contourf but I use to only do

like 10 levels. I think it might be because use a log yaxis and think it

used to be a bit funky or couldn’t get it working properly but seemed

fine now.

No, I don’t want to modify the ticks but the black lines around that

like how they are removed on the major axis in this example:

https://dl.dropbox.com/u/13534143/example1.png

I want to remove the black lines also around the colorbar. Not the tick

marks. Does that make sense?

cbar.outline.set_color(‘none’)

or

cbar.outline.set_visible(False)

One more quick question out of curiosity noticing from saving plots to

.pdf from contourf and pcolormesh vs specgram. Specgram seems to output

the lines and text as vector graphics. Then imbeds the image. When

outputting from pcolormesh or contourf this isn’t the case. It tries to

write the lines or something else weird happens. Can you output to .pdf

from these and make the lines and text be vectors. Then the image output

as an image in the pdf like in specgram. Or is there a setting to do

this and specify the .dpi of the image in the .pdf.

Lines and text are output to pdf exactly the same by specgram, pcolormesh, and contourf. The difference should be only in the image part of the plot, which is rasterized for a specgram image and for the “quadmesh” produced by pcolormesh, but is a set of patches (vector specification, not rasterized) for contourf. Are you seeing results that are inconsistent with this expectation?

Eric

Thanks a lot,

Jeff

On Fri, Jul 27, 2012 at 5:51 PM, Eric Firing <efiring@…202… > > > mailto:efiring@...120.....202...> wrote:

On 2012/07/26 9:20 PM, Jeffrey Spencer wrote:



    import numpy as np

    import matplotlib as mpl

X, Y = np.meshgrid(arange(20),arange(__20))

    Z = np.arange(20*20)

    Z = Z.reshape(20,20)

logNorm = mpl.colors.Normalize(vmin=0,__vmax=200)

    fig = mpl.pyplot.figure(10)

    ax = fig.add_subplot(111)

    surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm =

    logNorm)

    cbar = fig.colorbar(surf, shrink=0.70, norm=logNorm)

    show()







OK, the basic problem here is that you are specifying 100 levels,

which are being auto-selected to cover the actual data range; and

the colorbar is doing what it is supposed to do, which is show the

levels you actually have.  Try leaving out the norm, and just

specify the levels to cover what you want, more like this:



surf = ax.contourf(X, Y, Z, np.arange(0, 200.1, 2), cmap=mpl.cm.jet,

extend='both')

cbar = fig.colorbar(surf, shrink=0.7)



If you actually do want a log norm, you can pass that in to contourf

and it will be passed on to colorbar; but most likely you should

still specify the levels you want as an array, and not specify vmin

and vmax in the norm.  If you want log scaling, it may work better

to simply plot the log of Z, and use the colorbar label to indicate

that this is what you are doing.



Note that with a recent change, you can use the set_under and

set_over methods of the cmap to specify arbitrary colors, or no

color, for the extended regions; or you can leave out the "extend"

kwarg and not color the regions outside the range of your contour

levels.



In general, contourf is most appropriate when there is a moderate

number of levels, well under 100; if you want that many gradations,

then you might do better with pcolormesh or ax.pcolorfast or imshow.

  For those image-like methods, it is appropriate to use vmin and

vmax, either directly, or in a norm.



Eric
2

Think I figured out an actual bug in the function: colors.Normalize().

The behavior states that if vmin or vmax is passed in as None it should take the minimum or maximum value respectively.

If only one value is passed into the function, both values are overwritten to the min and max instead of just the one that is not passed in. So the only case where it sets the limits is if both vmin and vmax are passed to the function. Is this the desired behavior because from the docstring it seems this is incorrect.

Cheers,

Jeff

···

On Sat, Jul 28, 2012 at 2:15 PM, Jeffrey Spencer <jeffspencerd@…287…> wrote:

I figured out you can pass in the rasterized keyword to all of those to change the rasterization in the output.

Also the docs say for pcolormesh it defaults to the backend if not set. Therefore, in the case of a vector based it would output vectors if not set to rasterize.

Haven’t tested but curious. Lets say I want to output at 600dpi but I display images interactively at 100dpi. Does it always rasterize the image to the higher dpi? I had noticed this didn’t seem to occur in specgram but figured because the specgram is a relatively low resolution image that outputing at 600dpi doesn’t do anything because original image is already a low resolution. I would expect the other modes do do this where the image isn’t already output and have to rasterize the image when saving like pcolormesh and contour plots.

Cheers,

Jeff

On Sat, Jul 28, 2012 at 1:43 PM, Jeffrey Spencer <jeffspencerd@…287…> wrote:

Yes, specgram rasterizes and contourf is definately a vector specification which isn’t optimal for 100 levels.

I would switch to pcolormesh but the output file can’t be rasterizing the image. It outputs a huge file in .pdf (40X bigger than equivalent .png) and it looks like it is vector based not rasterized.

Basically, If you output specgram or imshow in .pdf or .png the file sizes are relatively comparable with .pdf, .eps, .svg being slightly larger due to embedding the picture.

If I output in pcolor, pcolormesh, contourf (with more than 100 levels) the file sizes are huge for .pdf, .eps, .svg which I’m assuming because vector based output. It also looks like vector based output because can see the lines it draws for contours. Could this possibly be a selection for these outputs to force raster based processing or is that not easy.

On Sat, Jul 28, 2012 at 3:26 AM, Eric Firing <efiring@…202…> wrote:

On 2012/07/26 10:26 PM, Jeffrey Spencer wrote:

Thanks, that is all good info to know. I change my data to log and

normalize it so the logNorm is just linear actually so specifying only

levels is fine. I’ll let you know if that doesn’t work properly for some

reason.

Ok, yeah I looked at pcolormesh quickly and can’t remember why I chose

originally when I wrote this to go with contourf but I use to only do

like 10 levels. I think it might be because use a log yaxis and think it

used to be a bit funky or couldn’t get it working properly but seemed

fine now.

No, I don’t want to modify the ticks but the black lines around that

like how they are removed on the major axis in this example:

https://dl.dropbox.com/u/13534143/example1.png

I want to remove the black lines also around the colorbar. Not the tick

marks. Does that make sense?

cbar.outline.set_color(‘none’)

or

cbar.outline.set_visible(False)

One more quick question out of curiosity noticing from saving plots to

.pdf from contourf and pcolormesh vs specgram. Specgram seems to output

the lines and text as vector graphics. Then imbeds the image. When

outputting from pcolormesh or contourf this isn’t the case. It tries to

write the lines or something else weird happens. Can you output to .pdf

from these and make the lines and text be vectors. Then the image output

as an image in the pdf like in specgram. Or is there a setting to do

this and specify the .dpi of the image in the .pdf.

Lines and text are output to pdf exactly the same by specgram, pcolormesh, and contourf. The difference should be only in the image part of the plot, which is rasterized for a specgram image and for the “quadmesh” produced by pcolormesh, but is a set of patches (vector specification, not rasterized) for contourf. Are you seeing results that are inconsistent with this expectation?

Eric

Thanks a lot,

Jeff

On Fri, Jul 27, 2012 at 5:51 PM, Eric Firing <efiring@…202… > > > > mailto:efiring@...120.....202...> wrote:

On 2012/07/26 9:20 PM, Jeffrey Spencer wrote:



    import numpy as np

    import matplotlib as mpl

X, Y = np.meshgrid(arange(20),arange(__20))

    Z = np.arange(20*20)

    Z = Z.reshape(20,20)

logNorm = mpl.colors.Normalize(vmin=0,__vmax=200)

    fig = mpl.pyplot.figure(10)

    ax = fig.add_subplot(111)

    surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm =

    logNorm)

    cbar = fig.colorbar(surf, shrink=0.70, norm=logNorm)

    show()







OK, the basic problem here is that you are specifying 100 levels,

which are being auto-selected to cover the actual data range; and

the colorbar is doing what it is supposed to do, which is show the

levels you actually have.  Try leaving out the norm, and just

specify the levels to cover what you want, more like this:



surf = ax.contourf(X, Y, Z, np.arange(0, 200.1, 2), cmap=mpl.cm.jet,

extend='both')

cbar = fig.colorbar(surf, shrink=0.7)



If you actually do want a log norm, you can pass that in to contourf

and it will be passed on to colorbar; but most likely you should

still specify the levels you want as an array, and not specify vmin

and vmax in the norm.  If you want log scaling, it may work better

to simply plot the log of Z, and use the colorbar label to indicate

that this is what you are doing.



Note that with a recent change, you can use the set_under and

set_over methods of the cmap to specify arbitrary colors, or no

color, for the extended regions; or you can leave out the "extend"

kwarg and not color the regions outside the range of your contour

levels.



In general, contourf is most appropriate when there is a moderate

number of levels, well under 100; if you want that many gradations,

then you might do better with pcolormesh or ax.pcolorfast or imshow.

  For those image-like methods, it is appropriate to use vmin and

vmax, either directly, or in a norm.



Eric
3

Jeff,

I don't see it, so you need to provide an actual example. Here is what happens (in ipython --pylab) when I try to reproduce what it sounds like you are describing:

In [2]: norm = Normalize(vmax=10)

In [3]: norm([-10, 10, 20])
Out[3]:
masked_array(data = [ 0. 1. 1.5],
              mask = False,
        fill_value = 1e+20)

In [4]: norm.vmin
Out[4]: -10.0

In [5]: norm.vmax
Out[5]: 10

This behavior is as described in the docstring.

The one potentially confusing aspect that I see at the moment is that the Normalize class does not distinguish between explicit vmin, vmax, and the values that it calculates via its autoscale_None method when called. As soon as a None is replaced by an actual value, whether by being set explicitly or by autoscaling, the replacement is permanent. Perhaps this should be made explicit in the docstring. The docstrings could be re-arranged to clarify the roles of the class initializer and its __call__ method.

Eric

···

On 2012/07/28 10:17 PM, Jeffrey Spencer wrote:

Think I figured out an actual bug in the function: colors.Normalize().

The behavior states that if vmin or vmax is passed in as None it should
take the minimum or maximum value respectively.

If only one value is passed into the function, both values are
overwritten to the min and max instead of just the one that is not
passed in. So the only case where it sets the limits is if both vmin and
vmax are passed to the function. Is this the desired behavior because
from the docstring it seems this is incorrect.

Cheers,
Jeff

4

Eric,

Normalize appears to be working correctly and as you stated above but when passed into contourf appears to have inconsistent results not following the docstring by allowing the value to change.

Quick examples:

X, Y = meshgrid(arange(20),arange(20))

Z = arange(20*20)

Z = Z.reshape(20,20)

norm = colors.Normalize(vmin=200,vmax=None)

print norm.vmin

fig = figure(10)

ax = fig.add_subplot(111)

surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

This vmin has no effect where if you pass in:

X, Y = meshgrid(arange(20),arange(20))

Z = arange(20*20)

Z = Z.reshape(20,20)

norm = colors.Normalize(vmin=200,vmax=Z.max())

print norm.vmin

fig = figure(10)

ax = fig.add_subplot(111)

surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

it has the desired effect.

Let me know if this is correct or I’m missing something here.

Cheers,

Jeff

···

On Mon, Jul 30, 2012 at 12:32 AM, Eric Firing <efiring@…202…> wrote:

On 2012/07/28 10:17 PM, Jeffrey Spencer wrote:

Think I figured out an actual bug in the function: colors.Normalize().

The behavior states that if vmin or vmax is passed in as None it should

take the minimum or maximum value respectively.

If only one value is passed into the function, both values are

overwritten to the min and max instead of just the one that is not

passed in. So the only case where it sets the limits is if both vmin and

vmax are passed to the function. Is this the desired behavior because

from the docstring it seems this is incorrect.

Cheers,

Jeff

Jeff,

I don’t see it, so you need to provide an actual example. Here is what happens (in ipython --pylab) when I try to reproduce what it sounds like you are describing:

In [2]: norm = Normalize(vmax=10)

In [3]: norm([-10, 10, 20])

Out[3]:

masked_array(data = [ 0. 1. 1.5],

         mask = False,

   fill_value = 1e+20)

In [4]: norm.vmin

Out[4]: -10.0

In [5]: norm.vmax

Out[5]: 10

This behavior is as described in the docstring.

The one potentially confusing aspect that I see at the moment is that the Normalize class does not distinguish between explicit vmin, vmax, and the values that it calculates via its autoscale_None method when called. As soon as a None is replaced by an actual value, whether by being set explicitly or by autoscaling, the replacement is permanent. Perhaps this should be made explicit in the docstring. The docstrings could be re-arranged to clarify the roles of the class initializer and its call method.

Eric

5

You are correct. Contour and contourf are respecting vmin, vmax that have been set in the norm only if both have been set, so that it is fully scaled. I think that changing this to respect a single preset vmin or vmax would be reasonable.

Eric

···

On 2012/07/29 5:13 AM, Jeffrey Spencer wrote:

Eric,

Normalize appears to be working correctly and as you stated above but
when passed into contourf appears to have inconsistent results not
following the docstring by allowing the value to change.

Quick examples:

X, Y = meshgrid(arange(20),arange(20))
Z = arange(20*20)
Z = Z.reshape(20,20)
norm = colors.Normalize(vmin=200,vmax=None)
print norm.vmin
fig = figure(10)
ax = fig.add_subplot(111)
surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

This vmin has no effect where if you pass in:

X, Y = meshgrid(arange(20),arange(20))
Z = arange(20*20)
Z = Z.reshape(20,20)
norm = colors.Normalize(vmin=200,vmax=Z.max())
print norm.vmin
fig = figure(10)
ax = fig.add_subplot(111)
surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

it has the desired effect.

Let me know if this is correct or I'm missing something here.

Cheers,
Jeff

6

I’d agree as then it would work as pcolormesh or imshow does which when passed in the same norm will only set vmin or vmax that wasn’t previously set.

Assume this would be a really easy fix but I can have a look at submitting a patch if someone else isn’t going to do so just let me know.

Cheers,

Jeff

···

On Mon, Jul 30, 2012 at 2:31 AM, Eric Firing <efiring@…202…> wrote:

On 2012/07/29 5:13 AM, Jeffrey Spencer wrote:

Eric,

Normalize appears to be working correctly and as you stated above but

when passed into contourf appears to have inconsistent results not

following the docstring by allowing the value to change.

Quick examples:

X, Y = meshgrid(arange(20),arange(20))

Z = arange(20*20)

Z = Z.reshape(20,20)

norm = colors.Normalize(vmin=200,vmax=None)

print norm.vmin

fig = figure(10)

ax = fig.add_subplot(111)

surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

This vmin has no effect where if you pass in:

X, Y = meshgrid(arange(20),arange(20))

Z = arange(20*20)

Z = Z.reshape(20,20)

norm = colors.Normalize(vmin=200,vmax=Z.max())

print norm.vmin

fig = figure(10)

ax = fig.add_subplot(111)

surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

it has the desired effect.

Let me know if this is correct or I’m missing something here.

Cheers,

Jeff

You are correct. Contour and contourf are respecting vmin, vmax that have been set in the norm only if both have been set, so that it is fully scaled. I think that changing this to respect a single preset vmin or vmax would be reasonable.

Eric

7

I'd agree as then it would work as pcolormesh or imshow does which when
passed in the same norm will only set vmin or vmax that wasn't
previously set.

Assume this would be a really easy fix but I can have a look at
submitting a patch if someone else isn't going to do so just let me know.

I will take care of it, thanks.

Eric

···

On 2012/07/30 1:17 AM, Jeffrey Spencer wrote:

Cheers,
Jeff

On Mon, Jul 30, 2012 at 2:31 AM, Eric Firing <efiring@...202... > <mailto:efiring@…202…>> wrote:

    On 2012/07/29 5:13 AM, Jeffrey Spencer wrote:

        Eric,

        Normalize appears to be working correctly and as you stated
        above but
        when passed into contourf appears to have inconsistent results not
        following the docstring by allowing the value to change.

        Quick examples:

        X, Y = meshgrid(arange(20),arange(20)__)
        Z = arange(20*20)
        Z = Z.reshape(20,20)
        norm = colors.Normalize(vmin=200,__vmax=None)
        print norm.vmin
        fig = figure(10)
        ax = fig.add_subplot(111)
        surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

        This vmin has no effect where if you pass in:

        X, Y = meshgrid(arange(20),arange(20)__)
        Z = arange(20*20)
        Z = Z.reshape(20,20)
        norm = colors.Normalize(vmin=200,__vmax=Z.max())
        print norm.vmin
        fig = figure(10)
        ax = fig.add_subplot(111)
        surf = ax.contourf(X,Y,Z, 100, cmap=matplotlib.cm.jet, norm = norm)

        it has the desired effect.

        Let me know if this is correct or I'm missing something here.

        Cheers,
        Jeff

    You are correct. Contour and contourf are respecting vmin, vmax
    that have been set in the norm only if both have been set, so that
    it is fully scaled. I think that changing this to respect a single
    preset vmin or vmax would be reasonable.

    Eric