Volume average #212
Volume average #212
Conversation
… updating matrix return.
Also make function in interputils return something
optimzations to static typing of functions in cython
Codecov Report
@@ Coverage Diff @@
## master #212 +/- ##
==========================================
+ Coverage 81.39% 81.48% +0.09%
==========================================
Files 24 24
Lines 5197 5228 +31
==========================================
+ Hits 4230 4260 +30
- Misses 967 968 +1
Continue to review full report at Codecov.
|
|
@prisae I know I never responded to it, but I did see your comment before, and implemented something that does the same thing, but bypasses the numpy boolean indexing. Let me know if that is sufficient. The idea was to speed it up when you're interpolating a large mesh extent onto a smaller mesh extent, without loosing too much time. |
I'll try to have a look before the meeting today. |
|
Yes, I like it now, great work @jcapriot! |
Improves speed if the new mesh is much smaller than the old mesh. See simpeg/discretize#212
Improves speed if the new mesh is much smaller than the old mesh. See simpeg/discretize#212
There was a problem hiding this comment.
Great to have this function built-in.
So I ran a small test (notebook attached) comparing the projection matrix between this one and the one used in simpeg.maps.TileMap. The good news is that both give the same answer, even after applying active cells. This one appears to be a lot slower, however, by about a factor 10x fairly consistently.
I am guessing there is a cost for having more general Tensor <-> Octree?
|
Yes, there is certainly a cost associated with the general Tensor to OcTree... but you do bring up a good point. There's no reason the function could not recognize that a Tensor and OcTree share the same base mesh and switch to the faster version. Also, which operation is the first timing operation doing there, where the discretize one is faster than the SimPEG one? |
|
YEa not sure why the first in the loop is different. Maybe some imports on the simpeg side? |
|
@jcapriot: would it be possible to do a beta release to make this available on conda-forge? I am having trouble installing discretize from source on a remote machine at the moment... |
| @@ -82,3 +83,109 @@ def interpmat(locs, x, y=None, z=None): | |||
| Q = sp.csr_matrix((vals, (I, J)), | |||
| shape=(npts, np.prod(shape))) | |||
| return Q | |||
|
|
|||
| def volume_average(mesh_in, mesh_out, values=None, output=None): | |||
| """Volume averaging interpolation between meshes. | |||
There was a problem hiding this comment.
Thanks for such a beautiful docstring :)
|
@domfournier Can you check the timings again for me? I added shortcuts for when meshes share the same base mesh edited: actually never mind I'll run your notebook, thanks for that |
This PR implements the volume averaging operation for
TensorMeshandTreeMesh, as well as combinations of the two. I have used the Tensor to Tensor myself, however the Tree to X operations could use some more test driving.Of course there's some simple tests in here to ensure it's doing what I think it should be doing.
I was also running into issues getting the numpydoc's plot directive to automatically work in the Examples section of docstrings, and the solution was to remove the napoleon docs, since we're already using the
numpydocextension, and I couldn't find any examples of Google style docstrings, I just removed that fromconf.py. So if this is okay, I'd like to just remove thenapoleonSphinx extension.