Skip to content

Friendlier matplotlib interaction with large images

License

Notifications You must be signed in to change notification settings

NSLS-II-CHX/mpl-modest-image

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

48 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ModestImage

Friendlier matplotlib interaction with large images

ModestImage extends the matplotlib AxesImage class, and avoids unnecessary calculation and memory when rendering large images (where most image pixels aren't visible on the screen). It has the following benefits over AxesImage:

  • Draw time is (roughly) independent of image size
  • Large numpy.memmap arrays can be visualized, without making an in-memory copy of the entire array. This enables visualization of images too large to fit in memory.

Installation

pip install ModestImage

or

easy_install ModestImage

Using ModestImage

The easiest way is to use the modified imshow function:

import matplotlib.pyplot as plt
from modest_image import ModestImage, imshow

ax = plt.gca()
imshow(ax, image_array, vmin=0, vmax=10)
plt.show()

imshow accepts all the keyword arguments that the matplotlib function does. The vmin and vmax keywords aren't necessary but, if they are not provided, the entire image will be scanned to determine the min/max values. This can be slow if the array is huge.

To create a ModestImage artist directly:

artist = ModestImage(data=array)

Looking at very big FITS images

import matplotlib.pyplot as plt
import pyfits
from modest_image import imshow

ax = plt.gca()
huge_array = pyfits.open('file_name.fits', memmap=True)[0].data
artist = imshow(ax, huge_array, vmin=0, vmax=10)
plt.show()

This opens almost instantly, with a modest memory footprint.

Why is Matplotlib Image Drawing Slow?

For the first draw request after setting the color mapping or data array, AxesImage (the default matplotlib image class) calculates the RGBA value for every pixel in the data array. That's a lot of work for large images, and usually overkill given that the final rendering is limited by screen resolution (usually 100K-1M pixels) and not image resolution (often much more).

AxesImage compensates for this by saving the results of this scaling. This means that subsequent renderings that only change the position or zoom level are very fast. However, in interactive situations where the data array or intensity scale change often, AxesImage wastes lots of time calculating RGBA values for every pixel in a (potentially large) data set. It also makes several temporary arrays with size comparable to the original array, wasting memory.

How is ModestImage faster?

ModestImage resamples the image array at each draw request, extracting a smaller image whose resolution and extent are matched to the screen resolution. Thus, the RGBA scaling step is much faster, since it takes place only for pixels relevant for the current rendering.

This scheme does not take advantage of AxesImage's caching, and thus redraws after move and zoom operations are slightly slower. However, draws after colormap and data changes are substantially faster, and most redraws are fast enough for interactive use.

Performance and Tests

speed_test.py compares the peformance of ModestImage and AxesImage. For a 1000x1000 pixel image:

    Performace Tests for AxesImage

           time_draw: 186 ms per operation
           time_move: 19 ms per operation
      time_move_zoom: 28 ms per operation

    Performace Tests for ModestImage

          time_draw: 25 ms per operation
          time_move: 20 ms per operation
     time_move_zoom: 28 ms per operation

time_draw is the render time after the cache has been cleared (e.g. after set_data has been called, or the colormap has been changed). ModestImage is slightly slower than, though still competetive with, AxesImage for move and zoom operations where AxesImage uses cached data.

Unit tests can be found in the tests directory. ModestImage does not always produce results identical to AxesImage at the pixel level, due to how it downsamples images. The discrepancy is minor, however, and disappears if no downsampling takes place (i.e. a screen pixel samples <= 1 data pixel)

About

Friendlier matplotlib interaction with large images

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 100.0%