Jaime,
One thing that might be a good selling point would be if you built-in
threaded operations on image stacks, which I think are pretty common in all
forms of microscopy now. I'm thinking particularly of `ni.convolve()`,
where one might want to repeatedly call the function on every image frame
in the stack. Generally multi-threading kernel operations like convolve on
single images is tricky, but if you have a stack it's much more
straight-forward to parallelize. Right now I would do such a thing with a
thread pool in Python, but it would scale a lot better if there was C-level
support. E.g.:
import multiprocessing as mp
import numpy as np
import scipy.ndimage as ni
# funcHandle must release the GIL
def stackFunc( funcHandle, images, funcArgs ):
tPool = mp.ThreadPool( self.n_threads )
slices = self.images.shape[0]
# Build parameters list for the threaded processes, consisting of
index
tArgs = [None] * slices
for J in np.arange(slices):
tArgs[J] = (J, images, funcArgs)
# All operations are done 'in-place'
tPool.map( funcHandle, tArgs )
tPool.close()
tPool.join()
Sincerely,
Robert
On Mon, Jun 26, 2017 at 2:30 AM, Jaime Fernández del Río <
[email protected]> wrote:
> Hi all,
>
> I have started working on an ambitious refactor of scipy.ndimage. Some
> more context on what I would like to achieve can be found here
> <https://mail.python.org/pipermail/scipy-dev/2017-June/021986.html>.
>
> scikit-image is probably the biggest "customer" of ndimage, so even though
> the API and performance are not expected to change, you may want to have a
> say.
>
> Assuming there is no major concern with the overall plan, I would also
> like to poke this community for help with:
>
> - figuring out a sensible test strategy: Ralf Gommers points out that
> we should probably use scikit-image's test suite regularly to complement
> the relatively few tests currently in ndimage: is the development version
> of scikit-image stable enough for this purpose or do you suggest using a
> specific stable release? Any other ideas?
> - PR reviewing. The idea is to make this refactor as incremental as
> possible. That means many smaller interdependent PRs. Even if you don't
> have commit rights to the scipy repository you are probably better versed
> than most of us in image processing algorithms, so if anyone can keep an
> eye on ndimage PRs
>
> <https://github.com/scipy/scipy/pulls?q=is%3Apr+is%3Aopen+label%3Ascipy.ndimage>
> and provide occasional reviews it would be greatly appreciated.
> - expert consulting: some of the code in ndimage is pretty hard to
> understand, e.g. I struggle with spline interpolation. If anyone with the
> proper academic knowledge is willing to answer stupid questions, please let
> me know.
> - the refactoring itself: if anyone would like to be a part of this,
> it would be my pleasure! A basic command of the C language is needed, but
> I'm more than willing to mentor beginners.
>
> Thanks!
>
> Jaime
>
> --
> (\__/)
> ( O.o)
> ( > <) Este es Conejo. Copia a Conejo en tu firma y ayúdale en sus planes
> de dominación mundial.
>
> _______________________________________________
> scikit-image mailing list
> [email protected]
> https://mail.python.org/mailman/listinfo/scikit-image
>
>
--
Robert McLeod, Ph.D.
[email protected]
[email protected] <[email protected]>
[email protected]
_______________________________________________
scikit-image mailing list
[email protected]
https://mail.python.org/mailman/listinfo/scikit-image
