Sample binary grid and code to demonstrate imshow vs scipy problem.

imshowwhy.py

# Read a grid in the native projection and display with matplotlib and scipy.
import scipy.misc

import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np


def make_image():

    dpi = 100.

    fig = plt.figure(figsize=(600/dpi, 900/dpi), dpi=dpi, edgecolor='none')

    image_axes = fig.add_axes([0, 0, 1, 1])

    landmask = 'resampled-mask.600x900.bin'
    grid = np.fromfile(landmask, dtype=np.uint8).reshape((900, 600))
    grid[grid == 101] = 0       # reset permanent ice to land

    ct_cmap, ct_norm = mpl.colors.from_levels_and_colors(
        [0, 50, 256], ['#000000', '#FFFFFF']
    )

    image_axes.set_axis_off()
    image_axes.imshow(grid, origin='upper', interpolation='nearest',
                      cmap=ct_cmap, norm=ct_norm)

    fig.savefig('imshow.png', dpi=dpi, facecolor=fig.get_facecolor(),
                edgecolor='none', pad_inches=0.0)
    plt.close(fig)

    # Now the same grid with scipy imsave but also as an RGBA image.

    rgba = np.zeros((900, 600, 4))
    rgba[:, :, 0] = grid
    rgba[:, :, 1] = grid
    rgba[:, :, 2] = grid
    rgba[:, :, 3] = 255
    scipy.misc.imsave('scipyimsave.png', rgba)


if __name__ == '__main__':

    make_image()

make-gif.sh

#!/bin/bash

set -xe

python imshowwhy.py

convert imshow.png scipyimsave.png diff.gif

resampled-mask.600x900.bin

Sample code and data to demonstrate imshow problems.md

# Sample code and data for describing data issue in imshow

## Problem: imshow seems to move some of the data around in it's axes.

When I try to us axes.imshow in matplotlib with a grid of known size and shape,
and then later read the png data, the image's data changes compared to the
input gridded data.


I've included a script `make-gif.sh` that reads a sample 900x600 image and then
writes that to a file using matplotlib's imshow, as well as with scipy's
savefig.  The two png images are then fused (with imagemagick's convert) and
you can see flicking where the two different outputs show differences.


Examining the bands in the output PNGS, the data saved with scipy's
savefig are exactly aligned with the input data while there is a shift in three
quadrants of the data saved with imshow.


I would guess I'm missing some flag, but I have been unable to find something
to prevent this undesired shift.

Can someone help me out here?


Python 3.5.2:
```
matplotlib                1.5.3               np111py35_3    conda-forge/channel/main
scipy                     0.18.1          np111py35_blas_openblas_201  [blas_openblas]  conda-forge/channel/main
```


Thanks