Dongryeol Lee
bd9dfdcf9e
Last generation of the Krylov subspace based method has been caught - the bug was due to reusing the same dual-tree computation function across three different phases (some phases were required to use squared kernel values, rather than the regular values)
2008-03-03 20:52:57 +00:00
Dongryeol Lee
221c1ce793
Bug fix in the dense version: same problem with the naive version since I forgot to divided by the normalization constant for the influence values
2008-03-03 04:28:20 +00:00
Dongryeol Lee
6a2a91e22b
Minor fixes in the user-level function Compute
2008-03-03 01:38:23 +00:00
Dongryeol Lee
c69e48ef2b
Some work for setting up the experiment for the dense version of the algorithm has been completed. I now need to look at the krylov subspace version for correctness since I do some fancy stuff for the normalization constants
2008-03-02 22:13:27 +00:00
Dongryeol Lee
d315ca87b5
Krylov subspace based LPR now computes full information required for crossvalidation: confidence bands, regression estimate, leave-one-out regression estimate, influence value, and magnitude of its weight diagram vector in the hat matrix
2008-03-02 19:27:39 +00:00
Dongryeol Lee
f39afd4da8
Fix in the compute function for naive code which contained erroneous extra argument
2008-03-02 06:11:51 +00:00
Dongryeol Lee
b8f8db9bf5
Now shuffles the values for weight diagram and influence values for Krylov version
2008-03-02 04:59:18 +00:00
Dongryeol Lee
f6dba35e8b
Fixed a bug in the dense version of the algorithm which did not shuffle the computed magnitude of weight diagrams and influence values accordingly due to the query tree construction; also there was a logic mistake in the design of the code which assumed that under variable-bandwidth case, the epanechnikov moment computed for the denominator matrix could be used for the weight diagram matrix, but this is not true since the effective normalization constant for weight diagram matrix requires squared normalization constants
2008-03-02 04:39:00 +00:00
Dongryeol Lee
28e2e2b364
Now computes the influence values for each reference point, need to add in the final dual-tree pass for computing the magnitude of each weight diagram vector for each reference point to construct the confidence interval
2008-03-01 21:51:31 +00:00
Dongryeol Lee
fe05acf2ed
Fixed the compilation error for general kd-tree code
2008-03-01 21:40:40 +00:00
Dongryeol Lee
e5e724a387
Now precomputes query point expansions before finalizing regression estimates, since they are needed in computing the confidence intervals (need to be passed to the Krylov solver). The Krylov solver needs to be modified, plus one final dual-tree pass over the computed solution vectors need to be implemented
2008-03-01 05:22:57 +00:00
Dongryeol Lee
6de282c263
Added the leave-one-out regression estimates for the reference set training stage
2008-03-01 04:34:38 +00:00
Dongryeol Lee
08e82abbce
Added a utility for computing the coordinate of the point that is farthest away from a given point inside a given bounding ball
2008-03-01 00:56:11 +00:00
Dongryeol Lee
58d7642d4d
One bug in multi-cg loop caught - needed to maintain separate r dot z (old) values, but the previous one shared a global one
2008-03-01 00:45:09 +00:00
Dongryeol Lee
31f08c253b
Added Modified Gram Schmidt and batch point mutivariate expansion utilities
2008-03-01 00:39:53 +00:00
Dongryeol Lee
df78dbf722
A little optimization in the stratified computation, but did not help much - going to save distance computations as well
2008-02-29 01:29:03 +00:00
Dongryeol Lee
29bec45e3c
Added the ball-tree
2008-02-29 00:51:28 +00:00
Dongryeol Lee
9afb15d7a9
Revived the proximity project under new CVS
2008-02-28 21:42:08 +00:00
Dongryeol Lee
5b3ab78a00
Now saves the first column of the linear operators for the matrix-free version, but this is still linear in the number of unknowns
2008-02-28 01:15:18 +00:00
Dongryeol Lee
f982e08f37
Optimized the performance of the CG-based LPR algorithm
2008-02-27 17:58:00 +00:00
Dongryeol Lee
c237ae1620
Need to optimize more by sharing stratification results
2008-02-27 03:55:40 +00:00
Dongryeol Lee
ea8ed122e2
Edited the status message a bit in the CG loop
2008-02-27 02:34:40 +00:00
Dongryeol Lee
c9d5eb13ac
Fix in the parameter retrieval in the matrix-free method
2008-02-27 02:08:15 +00:00
Dongryeol Lee
ffd2a11e26
Added a solver file to the build file description
2008-02-27 00:41:43 +00:00
Dongryeol Lee
a8de9f0fa4
Added the Krylov solver file
2008-02-27 00:40:48 +00:00
Dongryeol Lee
147691a951
Added the optimization for letting converged queries exit the Krylov loop.
2008-02-27 00:39:38 +00:00
Dongryeol Lee
f2c60c4a46
A minor bug fix in CG method - needed to subtract off the number of queries at one point
2008-02-26 22:24:29 +00:00
Dongryeol Lee
4f86df69e8
Some more tweaks
2008-02-26 22:08:41 +00:00
Dongryeol Lee
e19a0bb46b
Initial check in of the Multi Conjugate-Gradient algorithm
2008-02-26 19:20:31 +00:00
Dongryeol Lee
83a4a21e1f
There was a problem in reference-side initialization - needed to clear out the Epanechnikov moments, but didn't. Now the reference stats are reset everytime the dual-tree multiplier call is made
2008-02-26 18:52:02 +00:00
Dongryeol Lee
306aef24c5
A bug in the matrix-vector multiplier. Need to fix now
2008-02-26 18:49:14 +00:00
vasiloglou
445c11a971
changes in the cap
2008-02-26 08:39:18 +00:00
Dongryeol Lee
67fd4fd261
Everything is fixed, and it works now
2008-02-26 02:51:20 +00:00
Dongryeol Lee
3a22e17b8d
Still does not work - possibly misuse or a bug in AztecOO
2008-02-26 01:47:29 +00:00
Dongryeol Lee
8dcae4adca
Compiled the iterative method... Now debugging
2008-02-25 21:34:24 +00:00
Dongryeol Lee
6bb0d2834a
Got Epetra stuff to compile
2008-02-25 18:36:47 +00:00
Dongryeol Lee
fb5adf58b7
Removed the custom-built Krylov solver, now going to replace it with Epetra solver
2008-02-24 21:52:16 +00:00
vasiloglou
363de2ed28
uo and running
2008-02-24 18:54:06 +00:00
vasiloglou
9252a734b5
compiled
2008-02-24 18:25:21 +00:00
Dongryeol Lee
0737de01cb
Fixed the bug on the knn bandwidth
2008-02-24 07:19:29 +00:00
Dongryeol Lee
edc7236d24
Fix for variable bandwidth
2008-02-24 06:02:06 +00:00
Dongryeol Lee
6ce1bc7d47
Need to fix matrix-free version now
2008-02-24 02:51:55 +00:00
Dongryeol Lee
be92268c22
Memory consumption cut in half
2008-02-24 01:23:24 +00:00
Dongryeol Lee
aff1938da8
Finally done with matrix-free version
2008-02-23 22:48:45 +00:00
Dongryeol Lee
8d87d60814
Sparse version now stores dense far-field moments
2008-02-23 22:24:31 +00:00
Dongryeol Lee
4e2228c19c
Epanechnikov acceleration added to Krylov method
2008-02-23 21:38:45 +00:00
vasiloglou
648ff4ad17
After a major reorganization that didn't necessarily give better code
...
we have SPE converging, need more on fixing the gain
Things to add:
Linearly updating in SPE
Inequality constraints on the furthest neighbors
Hessian Update based n the links
2008-02-23 07:40:15 +00:00
rriegel
879be538f3
fixed non-BLAS LengthEuclidean
2008-02-22 21:33:47 +00:00
Dongryeol Lee
dd37ca6c6a
Last attempt on krylov-based method
2008-02-22 20:08:25 +00:00
Dongryeol Lee
41474a58b8
Phase 1 has Epanechnikov acceleration added
2008-02-22 17:40:11 +00:00
Dongryeol Lee