npstreams is an open-source Python package for streaming NumPy array operations. The goal is to provide tested routines that operate on streams (or generators) of arrays instead of dense arrays. Some routines are CUDA-enabled, based on PyCUDA's GPUArray (work-in-progress).
Streaming reduction operations (sums, averages, etc.) can be implemented in constant memory, which in turns allows for easy parallelization.
This approach has been a huge boon when working with lots of images; the images are read one-by-one from disk and combined/processed in a streaming fashion.
This package is developed in conjunction with other software projects in the Siwick research group.
Consider the following snippet to combine 50 images
from an iterable
import numpy as np images = np.empty( shape = (2048, 2048, 50) ) from index, im in enumerate(source): images[:,:,index] = im avg = np.average(images, axis = 2)
source iterable provided 1000 images, the above routine would
not work on most machines. Moreover, what if we want to transform the images
one by one before averaging them? What about looking at the average while it
is being computed? Let's look at an example:
import numpy as np from npstreams import iaverage from scipy.misc import imread stream = map(imread, list_of_filenames) averaged = iaverage(stream)
At this point, the generators
iaverage are 'wired'
but will not compute anything until it is requested. We can look at the average evolve:
import matplotlib.pyplot as plt for avg in average: plt.imshow(avg); plt.show()
We can also use
last to get at the final average:
from npstreams import last total = last(averaged) # average of the entire stream
npstreams comes with some streaming functions built-in. Some examples:
- Numerics :
- Statistics :
ivar(single-pass variance), etc.
- Stacking :
More importantly, npstreams gives you all the tools required to build your own streaming function. All routines are documented in the API Reference on readthedocs.io.
Creating your own: Streaming Maximum
Let's create a streaming maximum function for a stream. First, we have to choose how to handle NaNs:
- If we want to propagate NaNs, we should use
- If we want to ignore NaNs, we should use
Both of those functions are binary ufuncs, so we can use
ireduce_ufunc. We will
also want to make sure that anything in the stream that isn't an array will be made into one
Putting it all together:
from npstreams import array_stream, ireduce_ufunc from numpy import maximum, fmax @array_stream def imax(arrays, axis = -1, ignore_nan = False, **kwargs): """ Streaming maximum along an axis. Parameters ---------- arrays : iterable Stream of arrays to be compared. axis : int or None, optional Axis along which to compute the maximum. If None, arrays are flattened before reduction. ignore_nan : bool, optional If True, NaNs are ignored. Default is False. Yields ------ online_max : ndarray """ ufunc = fmax if ignore_nan else maximum yield from ireduce_ufunc(arrays, ufunc, axis = axis, **kwargs)
This will provide us with a streaming function, meaning that we can look at the progress
as it is being computer. We can also create a function that returns the max of the stream
numpy.ndarray.max() using the
from npstreams import last def smax(*args, **kwargs): # s for stream """ Maximum of all arrays in a stream, along an axis. Parameters ---------- arrays : iterable Stream of arrays to be compared. axis : int or None, optional Axis along which to compute the maximum. If None, arrays are flattened before reduction. ignore_nan : bool, optional If True, NaNs are ignored. Default is False. Returns ------- max : scalar or ndarray """ return last(imax(*args, **kwargs)
Some of the features I want to implement in this package in the near future:
- Benchmark section : how does the performance compare with NumPy functions, as array size increases?
- Optimize the CUDA-enabled routines
- More functions : more streaming functions borrowed from NumPy and SciPy.
The API Reference on readthedocs.io provides API-level documentation, as well as tutorials.
The only requirement is NumPy. To have access to CUDA-enabled routines, PyCUDA must also be installed. npstreams is available on PyPI; it can be installed with pip.:
python -m pip install npstreams
To install the latest development version from Github:
python -m pip install git+git://github.com/LaurentRDC/npstreams.git
Each version is tested against Python 3.4, 3.5 and 3.6. If you are using a different version, tests can be run using the standard library's unittest module.
Support / Report Issues
All support requests and issue reports should be filed on Github as an issue.
npstreams is made available under the BSD License, same as NumPy. For more details, see LICENSE.txt.