[SciPy-User] downsampling with max in 2d

[Dan Stowell]

Hi all,

I would like to downsample 2D arrays using max-pooling. In other words,
to downsample by a factor of 2 would reduce

 a b c d
 e f g h
 i j k l
 m n o p

to

 max(a,b,e,f) max(c,d,g,h)
 max(i,j,m,n) max(k,l,o,p)


I've searched through numpy and scipy and not found a method for this,
but I'd be grateful for pointers if I've missed it.
In the meantime, I wrote the following function. If you can spot a
faster/better ways to achieve this, please do say:


def maxdownsample2d(data, factor):
	"""Supply a 2D numpy array, and an integer factor by which to
downsample (by nonoverlapping maxpooling) in both directions"""
	# factor might not be exact fit, so we trim to this.
	trimrows = int(np.floor(data.shape[0] / float(factor))) * factor
	trimcols = int(np.floor(data.shape[1] / float(factor))) * factor
	first = True
	for coff in range(factor):
		for roff in range(factor):
			hopped = data[roff:trimrows:factor, coff:trimcols:factor]
			if first:
				maxsofar = hopped
				first = False
			else:
				maxsofar = np.maximum(maxsofar, hopped)
	return maxsofar



Best
Dan
-- 
Dan Stowell
Postdoctoral Research Assistant
Centre for Digital Music
Queen Mary, University of London
Mile End Road, London E1 4NS
http://c4dm.eecs.qmul.ac.uk/people/dans.htm
http://www.mcld.co.uk/