Tps adaptive mesh refinement by djkapner · Pull Request #132 · AllenInstitute/render-python

djkapner · 2018-12-14T19:57:36Z

Do you have a thin plate spline mesh that you suspect is overkill and is really chapping your hide? Maybe, something like this?

This PR adds an adaptive mesh estimator to create a new smaller transform that matches the original one on the control points. This can reduce the computation time associated with the tform() method, for example.

code below provided for example to reproduce this plot

from renderapi.transform import thin_plate_spline as tps
from scipy.spatial import Delaunay
import matplotlib.pyplot as plt
import json
reload(tps)

with open('/home/danielk/djk_render-python/test/test_files/thin_plate_spline.json', 'r') as f:
    tpsj = json.load(f)

tf = tps.ThinPlateSplineTransform(json=tpsj)
ntf = tf.adaptive_mesh_estimate(
        tol=1.0,
        max_iter=100,
        starting_grid=7,
        nworst=10)

d=[]
d.append(Delaunay(tf.srcPts.transpose()))
d.append(Delaunay(ntf.srcPts.transpose()))

plt.figure(1)
plt.clf()
for i in range(2):
    plt.subplot(1, 2, i + 1)
    plt.triplot(d[i].points[:, 0], d[i].points[:, 1], d[i].simplices.copy(), color='k', alpha=0.25)
    plt.scatter(d[i].points[:, 0], d[i].points[:, 1], marker='.', color='r')
    plt.gca().set_aspect('equal')
    plt.title('%d points' % d[i].points.shape[0])

djkapner · 2018-12-14T19:59:12Z

note: this PR started from #129 and should be rebased/merged after that one.

codecov · 2018-12-14T20:07:50Z

Codecov Report

Merging #132 into develop will decrease coverage by 0.06%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           develop    #132      +/-   ##
==========================================
- Coverage    95.66%   95.6%   -0.07%     
==========================================
  Files           29      29              
  Lines         2005    2023      +18     
==========================================
+ Hits          1918    1934      +16     
- Misses          87      89       +2

Impacted Files	Coverage Δ
renderapi/transform/leaf/thin_plate_spline.py	`100% <100%> (ø)`	⬆️
renderapi/utils.py	`94% <0%> (-2%)`	⬇️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6029528...daf8185. Read the comment docs.

codecov · 2018-12-14T20:07:50Z

Codecov Report

Merging #132 into develop will increase coverage by 0.06%.
The diff coverage is 100%.

@@             Coverage Diff             @@
##           develop     #132      +/-   ##
===========================================
+ Coverage    95.65%   95.71%   +0.06%     
===========================================
  Files           29       29              
  Lines         2000     2029      +29     
===========================================
+ Hits          1913     1942      +29     
  Misses          87       87

Impacted Files	Coverage Δ
renderapi/transform/leaf/thin_plate_spline.py	`100% <100%> (ø)`	⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 05719f1...88b5f5e. Read the comment docs.

djkapner · 2018-12-17T23:02:38Z

rebased to master

fcollman · 2018-12-17T23:16:07Z

renderapi/transform/leaf/thin_plate_spline.py

+        sortind = np.argsort(delta[ind])
+
+        if niter == max_iter:
+            logger.warning(


are we sure you want to not raise an exception here, seems the client of this should be in control of whether to accept the result and move on. It could raise an Exception and return the newtf in the exception.

also this should be tested.

added the exception... not sure how to raise and return the newtf at the same time

RussTorres · 2018-12-17T23:50:44Z

renderapi/transform/leaf/thin_plate_spline.py

+            return newtf
+
+        src = np.vstack((src, self.src[ind[sortind[0: nworst]]]))
+        return self.adaptive_mesh_estimate(


Could we wrap the computation here as a class/static method? It looks like there is a lot of setup (relating to origin points, etc) that can be run once and passed to another recursive method, which can make this more readable for the future.

trivial change to retrigger build

renderapi/transform/leaf/thin_plate_spline.py

…apner/render-python into TPS_adaptive_mesh_refinement

fcollman · 2018-12-19T22:42:28Z

renderapi/transform/leaf/thin_plate_spline.py

+logger.addHandler(logging.StreamHandler(sys.stdout))
+
+
+class AdaptiveMeshEstimationError(Exception):


sorry to be nitpicky, but this should subclass EstimationError, or we can make EstimationError have an keyword argument of transform=None, so that it works whether you want to pass on the transform calculated or not, but code can be written to generically catch all EstimationErrors

that's fine, I can't quite imagine a use case where you would/should use a failed result... but you seem to. So, I don't mind making it what you're after.

oh! well.. actually I was intuiting because you had it as a warning before that you had a use case in mind to use the transform, so i was trying to accommodate that.

so what do we want to do... 1) leave this as is 2) make it subclass EstimateError 3) simply remove all this pass up the transform in the exception business.

In a scipy iterative solver situation, I've come to expect the behavior that a solution is always returned, but a max iter message pops up. That's what I was originally going for with the logger message:

from scipy.optimize import minimize, rosen, rosen_der x0 = [1.3, 0.7, 0.8, 1.9, 1.2] res = minimize(rosen, x0, method='Nelder-Mead', tol=1e-6, options={'maxiter': 1000, 'disp': True}) Optimization terminated successfully. Current function value: 0.000000 Iterations: 295 Function evaluations: 494 print(res.__class__) <class 'scipy.optimize.optimize.OptimizeResult'> del res res = minimize(rosen, x0, method='Nelder-Mead', tol=1e-6, options={'maxiter': 10, 'disp': True}) Warning: Maximum number of iterations has been exceeded. print(res.__class__) <class 'scipy.optimize.optimize.OptimizeResult'>

but maybe a better precedent is just bombing out like in our polynomial transform:
https://github.com/fcollman/render-python/blob/d5ccfafd75e5a3694e989e9c97397d315c9b0c52/renderapi/transform/leaf/polynomial_models.py#L231-L234

yeah i like bombing out better personally.

RussTorres · 2018-12-20T00:12:08Z

renderapi/transform/leaf/thin_plate_spline.py

+
+
+class AdaptiveMeshEstimationError(EstimationError):
+    def __init__(self, value, transform):


This is a tricky thing to do in python. I'm pretty sure that trying to raise something like this in e.g. a multiprocessing pool will cause hanging since the reduction function pickle uses will not treat it nicely. There is a way around this, but correct behavior should be tested if we go this route (which I think is okay).

djkapner · 2019-01-24T15:48:53Z

nudge.
I'd like to use this in a TEMCA/GPU lens correction thing pretty soon

Dan Kapner added 5 commits December 17, 2018 15:01

adding adaptive mesh estimate for TPS size reduction

3b57824

fixing className caps in json

894cc6a

added some documentation header

9fa5c7c

trivial change to retrigger build

66bccbe

trivial change to retrigger build

e6ba38c

djkapner force-pushed the TPS_adaptive_mesh_refinement branch from a73c24c to e6ba38c Compare December 17, 2018 23:02

fcollman reviewed Dec 17, 2018

View reviewed changes

RussTorres reviewed Dec 17, 2018

View reviewed changes

Dan Kapner and others added 3 commits December 18, 2018 03:54

separating recursion from setup, raising max_iter exception

5de474e

Update thin_plate_spline.py

ae42318

trivial change to retrigger build

another trivial change for rebuild

f0e067f

fcollman reviewed Dec 19, 2018

View reviewed changes

renderapi/transform/leaf/thin_plate_spline.py Show resolved Hide resolved

fcollman reviewed Dec 19, 2018

View reviewed changes

renderapi/transform/leaf/thin_plate_spline.py Show resolved Hide resolved

Dan Kapner added 2 commits December 19, 2018 08:46

added doc string, default pass self to recursion, custom exception

4ad3e61

Merge branch 'TPS_adaptive_mesh_refinement' of https://github.com/djk…

1bfd1e1

…apner/render-python into TPS_adaptive_mesh_refinement

fcollman reviewed Dec 19, 2018

View reviewed changes

Dan Kapner added 3 commits December 19, 2018 14:47

subclassing custom exception as EstimationError

809b376

removing custom class __str__ method

268453c

removing test of exception __str__

83ec597

RussTorres reviewed Dec 20, 2018

View reviewed changes

reverting to simpler estimation error exception

88b5f5e

fcollman approved these changes Jan 24, 2019

View reviewed changes

RussTorres approved these changes Jan 25, 2019

View reviewed changes

fcollman merged commit 43d2387 into AllenInstitute:develop Jan 25, 2019

		logger.addHandler(logging.StreamHandler(sys.stdout))


		class AdaptiveMeshEstimationError(Exception):



		class AdaptiveMeshEstimationError(EstimationError):
		def __init__(self, value, transform):

Conversation

djkapner commented Dec 14, 2018

Uh oh!

djkapner commented Dec 14, 2018

Uh oh!

codecov bot commented Dec 14, 2018

Codecov Report

Uh oh!

codecov bot commented Dec 14, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

djkapner commented Dec 17, 2018

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

djkapner commented Jan 24, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

codecov bot commented Dec 14, 2018 •

edited

Loading