Ryan Curtin
4db93bfa4c
Refactor to use new set of rules. How many times will I restart writing this algorithm until I actually get it working well?
2015-03-12 15:42:45 -04:00
Ryan Curtin
bce6482563
Slightly tighter prune.
2015-03-12 15:42:45 -04:00
Ryan Curtin
aa36ea42dc
Shitty implementation of Elkan-type prune. It makes things faster... a little...
2015-03-12 15:42:45 -04:00
Ryan Curtin
cf56ad4a2a
A start at implementing pointwise Hamerly bounds.
2015-03-12 15:42:45 -04:00
Ryan Curtin
5a4b15cf0b
Remove vestigial code. After all, this code isn't generated by genetic algorithms. (Maybe it should be?)
2015-03-12 15:42:45 -04:00
Ryan Curtin
6300189ab4
Correct Pelleg-Moore prunes that finish a node. There were cases where a Pelleg-Moore prune would happen before committing the point. This is actually getting pretty fast in terms of base cases, so I am happy with that (for once).
2015-03-12 15:42:44 -04:00
Ryan Curtin
b094362c2b
Add some debugging output. This makes it a bit easier to figure out what is going wrong.
2015-03-12 15:42:44 -04:00
Ryan Curtin
c413ad2399
Better speedups, provide more output on prunes.
2015-03-12 15:42:44 -04:00
Ryan Curtin
50d4a325c7
Remove debug output; implement Pelleg-Moore prune. Now this is faster than the regular dtkm branch. Sometimes starting over is a good thing.
2015-03-12 15:42:44 -04:00
Ryan Curtin
2633583a44
Refactor into UpdateOwner(), instead of an ugly loop at the beginning of TreeUpdate().
2015-03-12 15:42:44 -04:00
Ryan Curtin
de01b8f67a
I don't think this is worth saving. It also doesn't work very well, but I learned a lot about the bookkeeping I need to do.
2015-03-12 15:42:44 -04:00
Ryan Curtin
ca7b1a38ff
Last changes for a little while.
2015-03-12 15:42:44 -04:00
Ryan Curtin
a06391e861
Fix second bound to not be the first bound. In some situations, MinQueryNodeDistance() was equal to SecondClosestBound(). This fixes that. This also resolves an issue where SecondClosestBound() wasn't properly updated after an Elkan prune.
2015-03-12 15:42:44 -04:00
Ryan Curtin
ca4066a3a8
Don't set CQN to NULL before recursing. This was an error in TreeUpdate() that caused... some problems that I'm only vaguely aware of as problems. Also refactor one of the second bound rules and comment it so that it makes some semblance of sense.
2015-03-12 15:42:44 -04:00
Ryan Curtin
87ea896277
Remove some debug output we don't need anymore.
2015-03-12 15:42:44 -04:00
Ryan Curtin
b6a1c3699e
Unset Owner() after failed Hamerly prune. This is better than setting it by hand when we're checking if we can prune.
2015-03-12 15:42:44 -04:00
Ryan Curtin
29136848ae
Fix compilation warnings. secondClosest will eventually be removed but is commented for now.
2015-03-12 15:42:44 -04:00
Ryan Curtin
27f3a36e93
Count number of points that are Hamerly pruned.
2015-03-12 15:42:44 -04:00
Ryan Curtin
eba305d23f
Comment out most of the debugging output. It still shows you what's been Hamerly pruned and what hasn't, which is what's important right now.
2015-03-12 15:42:44 -04:00
Ryan Curtin
79379ee09e
Um, I made it faster. We're well into the land of confusion. Confusion will reign until eventual refactoring. Hamerly prunes still don't appear to be working properly after a couple iterations.
2015-03-12 15:42:44 -04:00
Ryan Curtin
994b4858ba
Remove unnecessary check for parent's SCB. There actually isn't a need to take the parent's second closest bound preferentially, and it was preventing some valid prunes. This gives a little more (but not enough) Hamerly pruning, and provides (as usual) trivial speedup---but more than none!
2015-03-12 15:42:44 -04:00
Ryan Curtin
abfc0b8e4d
Be more careful about when to take the last SCB. This gives a trivial speedup, but hey, speedup! Hamerly prunes are still not working properly, but I'm getting there.
2015-03-12 15:42:44 -04:00
Ryan Curtin
db0d792ef8
Handle SecondClosestBound() a little better. Debugging information is still there, and it is going to need to be seriously refactored. We still don't have properly working Hamerly prunes; they go away after a couple iterations incorrectly.
2015-03-12 15:42:44 -04:00
Ryan Curtin
772dc745c4
Add debug output; don't adjust second bound. This provides another minor speedup, but this still is nowhere near as fast as it should be with a properly working Hamerly prune.
2015-03-12 15:42:44 -04:00
Ryan Curtin
c83b94bc22
A first attempt at a working Hamerly prune. The bounds tighten too much and don't reset, so there's not much speedup, but it's a start.
2015-03-12 15:42:44 -04:00
Ryan Curtin
727714ea11
Avoid uninitialized owner field.
2015-03-12 15:42:44 -04:00
Ryan Curtin
96bcad550d
Fix invalid memory access.
2015-03-12 15:42:44 -04:00
Ryan Curtin
83f4201766
Add second bound.
2015-03-12 15:42:44 -04:00
Ryan Curtin
cd1765c30f
Rename to Hamerly-type score. Because that's what it is. Although Elkan's algorithm has a type of score that's sort of like this, it's Hamerly's algorithm that only keeps track of the second-closest cluster distance, and that's what we're doing here.
2015-03-12 15:42:44 -04:00
Ryan Curtin
6ad3f73535
Oops, bad parameter name.
2015-03-12 15:42:44 -04:00
Ryan Curtin
9515f01939
New stub method for (hopefully) good Elkan prune. This allows us to prune without even needing to recurse.
2015-03-12 15:42:44 -04:00
Ryan Curtin
f8a73a38ae
Elkan pruning can't prune the best query node. Bugfix. Trivial slowdown.
2015-03-12 15:42:44 -04:00
Ryan Curtin
dea4b29b71
Update FirstBound correctly. Trivial speedup.
2015-03-12 15:42:44 -04:00
Ryan Curtin
ee7eb43f01
Set FirstBound() after the allknn. Now, checking if FirstBound() == DBL_MAX is unnecessary, although this doesn't (yet) seem to give any additional prunes.
2015-03-12 15:42:44 -04:00
Ryan Curtin
fd948ea600
Fix typo, which actually gives a speedup. A fairly trivial speedup, but a speedup nonetheless.
2015-03-12 15:42:44 -04:00
Ryan Curtin
f9c48d2ce6
Fix warning for unused parameter.
2015-03-12 15:42:44 -04:00
Ryan Curtin
8a76218789
Try to do an Elkan prune not at a query leaf. It doesn't seem to work right now, though.
2015-03-12 15:42:44 -04:00
Ryan Curtin
37436c33c9
Remove unnecessary check; that can never happen.
2015-03-12 15:42:44 -04:00
Ryan Curtin
393be90675
Count the base cases and scores of the kNN run.
2015-03-12 15:42:44 -04:00
Ryan Curtin
61f65523b8
Don't calculate MaxDistance() unless we have to. Fairly significant time savings on this one.
2015-03-12 15:42:44 -04:00
Ryan Curtin
4a5add507f
Avoid calculating distances after an Elkan prune. Slight, nearly negligible performance gains.
2015-03-12 15:42:43 -04:00
Ryan Curtin
581a4a0dee
Don't accidentally take results before finished. We have to wait until all clusters are pruned, not all except one.
2015-03-12 15:42:43 -04:00
Ryan Curtin
1072053036
This output is less useful right now.
2015-03-12 15:42:43 -04:00
Ryan Curtin
e72ab4c744
Use ClosestQueryNode() not MaxQueryNodeDistance(). Eventually MaxQueryNodeDistance() will not be reset to DBL_MAX every iteration.
2015-03-12 15:42:33 -04:00
Ryan Curtin
56cbc1c3bb
Some fairly serious refactoring, but it works.
...
There are significant optimizations to still be made, but at this point we get
some acceleration over the naive algorithm.
2015-03-12 15:42:27 -04:00
Ryan Curtin
44fa159677
Perform the top-level scoring too.
2015-03-12 15:41:12 -04:00
Ryan Curtin
6366cc9349
Fix all -Wunused-parameter.
2015-03-12 15:40:02 -04:00
Ryan Curtin
347b8c08aa
Refactor to perform allknn on the clusters. This is done at the start of each iteration. This allows Elkan pruning without an additional distance calculation.
2015-03-12 15:39:29 -04:00
ryan
ee45212808
Beginning to refactor dual-tree k-means. So far the Pelleg-Moore rule is properly refactored (well, sort of, at least). Try to move IterationUpdate() into DualTreeKMeans::TreeUpdate(); for now this will make the algorithm simpler: instead of updating on the fly, update after an iteration. It may make more sense to do updates on demand later, but let's make the algorithm work first...
2015-03-12 15:39:26 -04:00
ryan
c63a518c8b
Slight refactoring of dual-tree k-means rules.
...
This is an attempt to clean up this code because it quickly became unwieldy and
hard to work with. A more modular approach could lead to a better, faster
solution.
2015-03-12 15:39:26 -04:00
Ryan Curtin