[Numpy-discussion] looking for code advice
Gordon Wrigley
gordon@tolomea....
Wed Sep 29 17:57:32 CDT 2010
Hi Josef
Thanks for your comments.
> numpy.ptp(a, axis=None, out=None)
> Range of values (maximum - minimum) along an axis.
I feel silly for not having noticed that function, it's right next to amin
and amax in the docs.
Both your suggestions improved the performance, but only a little,
regardless they make the code simpler which is always a good thing.
This is what I have currently:
def decimate(data_in):
in_x, in_y, in_z = data_in.shape
out_x = in_x / 2
out_y = in_y / 2
out_z = in_z / 2
out_shape = out_x, out_y, out_z
out_size = product(out_shape)
# figure out which chunks are homogeneous
reshaped_array = data_in.reshape(out_x, 2, out_y, 2, out_z,
2).transpose(0,2,4,1,3,5).reshape(out_x, out_y, out_z, 8)
not_equal_array = numpy.ptp(reshaped_array, axis=3)
# select the actual value for the homogeneous chunks and MIXED for the
heterogeneous
decimated_array = reshaped_array[:,:,:,0]
data_out = numpy.where(not_equal_array, MIXED, decimated_array)
return data_out
Regards Gordon
On Thu, Sep 30, 2010 at 2:36 AM, <josef.pktd@gmail.com> wrote:
> On Wed, Sep 29, 2010 at 9:19 AM, <josef.pktd@gmail.com> wrote:
> > On Wed, Sep 29, 2010 at 8:25 AM, Gordon Wrigley <gordon@tolomea.com>
> wrote:
> >> Hi
> >> First the disclaimer: This is my first numpy experience, so I have next
> to
> >> no idea what I'm doing.
> >> I've muddled through and managed to put together some code for my
> current
> >> problem and now that I have it going I'd like to hear any comments
> people
> >> may have on both my solution and other ways of approaching the problem.
> >> I have two goals here, I'd like to make the process run faster and I'd
> like
> >> to broaden my understanding of numpy as I can see from my brief use of
> it
> >> that it is a remarkably powerful tool.
> >> Now to the problem at hand. I find this difficult to explain but will
> try as
> >> best I can.
> >> The best word I have for the process is decimation. The input and output
> are
> >> both 3 dimensional arrays of uint8's. The output is half the size of the
> >> input along each dimension. Each cell [x,y,z] in the output corresponds
> to
> >> the 2x2x2 block [2*x:2*x+2, 2*y:2*y+2, 2*z:2*z+2] in the input. The
> tricky
> >> bit is in how the correspondence works. If all the cells in the input
> block
> >> have the same value then the cell in the output block will also have
> that
> >> value. Otherwise the output cell will have the value MIXED.
> >> Here is my current solution, from my limited testing it seems to produce
> the
> >> result I'm after.
> >> def decimate(data_in):
> >> in_x, in_y, in_z = data_in.shape
> >> out_x = in_x / 2
> >> out_y = in_y / 2
> >> out_z = in_z / 2
> >> out_shape = out_x, out_y, out_z
> >> out_size = product(out_shape)
> >> # figure out which chunks are homogeneous
> >> reshaped_array = data_in.reshape(out_x, 2, out_y, 2, out_z,
> >> 2).transpose(0,2,4,1,3,5).reshape(out_x, out_y, out_z, 8)
> >> min_array = numpy.amin(reshaped_array, axis=3)
> >> max_array = numpy.amax(reshaped_array, axis=3)
> >> equal_array = numpy.equal(min_array, max_array)
>
> maybe ptp==0 is faster
>
> numpy.ptp(a, axis=None, out=None)
> Range of values (maximum - minimum) along an axis.
>
> Josef
>
> >> # select the actual value for the homogeneous chunks and MIXED for
> the
> >> heterogeneous
> >> decimated_array = data_in[::2,::2,::2]
> >> mixed_array = numpy.tile(MIXED, out_size).reshape(out_shape)
> >> data_out = numpy.where(equal_array, decimated_array, mixed_array)
> >
> > data_out = numpy.where(equal_array, decimated_array, MIXED)
> > should work
> >
> > I don't see anything else, unless there is something in scipy.ndimage.
> >
> > I have to remember your reshape trick for 3d. (I don't know how many
> > temporary arrays this creates.)
> >
> > Josef
> >
> >> return data_out
> >> For the curious this is will be used to build a voxel octtree for a 3d
> >> graphics application. The final setup will be more complicated, this is
> the
> >> minimum that will let me get up and running.
> >> Regards
> >> Gordon
> >> P.S. congrats on numpy, it is a very impressive tool, I've only scraped
> the
> >> surface and it's already impressed me several times over.
> >>
> >> _______________________________________________
> >> NumPy-Discussion mailing list
> >> NumPy-Discussion@scipy.org
> >> http://mail.scipy.org/mailman/listinfo/numpy-discussion
> >>
> >>
> >
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion@scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/numpy-discussion/attachments/20100930/5fa297c6/attachment-0001.html
More information about the NumPy-Discussion
mailing list