imshow rendering issue

[mcspritz]

Hi, I'm plotting this simple graph and I got non equal pixel sizes for the graph, it happens with all kind of sizes and dpi.
(I'm using matplotlib 1.5.0, python 3.5.1)
I can prove with some pictures as well if needed (I'm not sure if the bug is in matplotlib or in the renderer).

from matplotlib import use
use('Agg')
import matplotlib.pyplot as plt
import numpy as np

t = np.random.randn(64, 64)
fig, ax = plt.subplots()
ax.imshow(t, interpolation='none', origin="lower")
plt.savefig(
    'crap29.png',
    format='png'
)
plt.close()`

Thanks

[jenshnielsen]

Im not completely sure I understand what you issue is. Your code is plotting an image and not a graph? Please attach the image that you save not as a prof but to demonstrate what you are talking about.

[mcspritz]

Sorry, I knew it wasn't easy to explain.
I traced a line on the graphs and noticed that "pixes" of the imshow are not all of the same size.

[jenshnielsen]

The issue sounds somewhat related to #5520 which is fixed by the rather large image rewrite in #5718 but I am not sure

[tacaswell]

Does it work correctly with 'nearest'?

[mcspritz]

Nope, I also tried to raise the dpi but no luck.

[jenshnielsen]

We have merge large changes to the image algorithm in to master and 2.x @mcspritz if possible could you try testing this on the 2.x branch or master?

[mcspritz]

Hi, I'm a newbie and want to know a simple thing, I built the cloned v2.x branch and it shows version 1.5.1+282.g918e80e, is this correct or I'm doing something wrong?

Thanks.

[jenshnielsen]

Yes that means that its 1.5.1 + 282 new commits on top of that which is the current state of the 2.x branch

[mcspritz]

Thanks,

So, the image is much better, I found for low number of pixels that column 0 has a pixel less.

For bigger images things get more complicated.

[mdboom]

I think what we're seeing here is the case where the scaling factor from data points to on-screen pixels is not an integer multiple. Imagine a 64x64 array scaled up to 200x200pixels. Since 64 does not divide into 200 evenly, it's inevitable that not all data points will be rendered at the same size. If I were seeing data points that were more than one pixel different in size, I'd be worried, but in the first example I'm seeing data points of 6 or 7 pixels, and in the last example 1 or 2 pixels. So I'm not sure we could logically do any better (other than adding a mode and the inevitable complexity that would always ensure an integral scaling factor).

[mdboom]

Closing this, as I don't think there's a bug here, but feel free to continue to comment.