[Numpy-discussion] Indexing in Numpy vs. IDL?
Sun Nov 16 21:54:42 CST 2008
Jason Woolard wrote:
> hi all,
> I'm fairly new to Numpy and I've been trying to port over some IDL code
> to become more familiar. I've been moderately successful with
> numpy.where and numpy.compress to do some of things that were pretty
> easy to do in IDL. I'm a bit confused about how the indexing of arrays
> works though.
> This is pretty straightforward:
> in IDL
> data = [50.00, 100.00, 150.00, 200.00, 250.00, 300.00, 350.00]
> index = WHERE((data GT 100.00) AND (data LT 300.00))
> new_data = data[index]
> print, new_data
> 150.000 200.000 250.000
> in Python
> >>> import numpy
> >>> from numpy import *
> >>> data = [50.00, 100.00, 150.00, 200.00, 250.00, 300.00, 350.00]
> >>> data = array(data, dtype=float32) #Convert list to array
> >>> index_mask = numpy.where((data > 100.00) & (data < 300.00), 1,0)
> #Test for the condition.
> >>> index_one = numpy.compress(index_mask, data)
> >>> print index_one
> [ 150. 200. 250.]
> But I'm having a bit of trouble with the Python equivalent of this:
> in IDL:
> index_two = WHERE ((data[index_one] GT bottom) AND (data[index_one] LE
> and also this:
> result = MAX(data[index_one[index_two]])
> From what I've read it looks like numpy.take() might work to do the
> indexing. I've tried to test this but I'm not getting the answers I'd
> expect. Am I overlooking something obvious here?
> Thanks in advance for any responses.
I too am a former IDLer. There is a slight paradigm shift here. In IDL
you can index an array with another array of integer indices, and you
can do that too in numpy. But numpy also lets you index an array with an
array of booleans. So
data > 100.
creates an array of booleans the same size and shape as data, so you can
write your new array as
data[ (data > 100.) & (data < 300.) ]
Note we don't use a "where" function. In numpy, "where" is a completely
different thing than in IDL.
If you really wanted to generate a list of indices, you can use the
"nonzero" method, but the numpy book says this isn't as fast as boolean
More information about the Numpy-discussion