Improve performance of tissue classification #2174
Conversation
Codecov Report
@@ Coverage Diff @@
## master #2174 +/- ##
=======================================
Coverage 91.37% 91.37%
=======================================
Files 251 251
Lines 32449 32449
Branches 3408 3408
=======================================
Hits 29651 29651
Misses 2057 2057
Partials 741 741 |
| @@ -217,48 +239,7 @@ class ConstantObservationModel(object): | |||
| return mu_upd, var_upd | |||
|
|
|||
|
|
|||
| def update_param_new(self, image, P_L_Y, mu, nclasses): | |||
There was a problem hiding this comment.
It seems this was not used anywhere @grlee77 ?
There was a problem hiding this comment.
Yes, it seems to be some leftover dead code. I did a quick grep of the code base just to be sure and don't see any other files that refer to it.
| cnp.npy_intp [::1] num_vox_view | ||
|
|
||
| # ravel C-contiguous array so we can use 1D indexing in the loop below | ||
| input_view = np.ascontiguousarray(input_image, dtype=np.double).ravel() |
There was a problem hiding this comment.
Okay so here is the trick that speeds up the code. Have a look @villalonreina
|
+1 on my side. AFAIK this is ready to be merged. @villalonreina do have a look. If @villalonreina does not get back to us during the next 2 days then we can move ahead. This loos pretty safe. We still may need to wait for the CLI issues. Thank you @grlee77. |
|
Hi @grlee77, CI's are back to green ! So can you just rebase this PR to make sure that everything is ok? and then, the code can be merged. Thank you! |
grlee77
force-pushed
the
tissue_segment_speedup
branch
from
d3ae31b
to
6780573
Compare
|
Hi @skoudoro, I rebased this on master. The two Azure failures seem like unrelated errors due to a network error/timeout when downloading packages. Can you try restarting those two and see if it was just a fluke. |
|
Thanks ! I just restarted them |
|
merging, Thank you @grlee77! |
This PR improves the performance of tissue classification by modifying the Cython code in the
seg_statsmethod ofConstantObservationModelto avoid high overhead. There should be no change in behavior with this PR.This initial call to
seg_statsin theclassifymethod ofTissueClassifierHMRFfor the example data indoc/demo/tissue_classification.pytakes > 60 seconds on my system prior to this PR and ~60 ms afterwards. So, the setup prior to iteration is now nearly instantaneous. The iterated conditional modes loop itself still takes the same time as before (~2 minutes), giving an overall speedup of around 30% for the demo.I would recommend reviewing the first commit in isolation as it has all of the important changes.
The remaining commits are just minor maintenance/style fixes.