bd tree about rann

[jefferis]

So I tried a little debugging and it seems that the problem is in the recursive function building the bd tree i.e. rbd_tree:

library(RANN)
nn2(query=matrix(c(1,1),ncol=2), data=matrix(rep(2, 4),ncol=2), k=1L, treetype="bd")

[... repeats many times]
    frame #52356: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd030, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52357: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd0d0, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52358: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd170, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52359: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd210, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52360: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd2b0, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52361: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd350, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52362: 0x00000001023e877c RANN.so`rbd_tree(pa=<unavailable>, pidx=<unavailable>, n=<unavailable>, dim=2, bsp=1, bnd_box=0x00007fff5fbfd3d0, splitter=<unavailable>, shrink=<unavailable>)(double**, int*, ANNorthRect const&, int, int, int&, double&, int&), ANNshrinkRule) at bd_tree.cpp:399 [opt]
    frame #52363: 0x00000001023e8548 RANN.so`ANNbd_tree::ANNbd_tree(this=0x00000001025426c0, pa=0x0000000102546a40, n=2, dd=2, bs=1, split=ANN_KD_SUGGEST, shrink=<unavailable>) at bd_tree.cpp:144 [opt]
    frame #52364: 0x00000001023e7917 RANN.so`::get_NN_2Set(data=<unavailable>, query=0x00000001034cd278, D=<unavailable>, ND=<unavailable>, NQ=<unavailable>, K=<unavailable>, EPS=0x0000000103534dc8, SEARCHTYPE=0x0000000103534d90, USEBDTREE=0x0000000103535110, SQRAD=0x00000001035350d8, nn_index=<unavailable>, distances=<unavailable>) at NN.cc:48 [opt]
    frame #52365: 0x0000000100114fe4 libR.dylib`do_dotCode(call=0x0000000103776610, op=<unavailable>, args=<unavailable>, env=<unavailable>) at dotcode.c:0 [opt]
    frame #52366: 0x0000000100146ba8 libR.dylib`bcEval(body=<unavailable>, rho=0x0000000103537498, useCache=<unavailable>) at eval.c:6781 [opt]
    frame #52367: 0x00000001001417db libR.dylib`Rf_eval(e=<unavailable>, rho=<unavailable>) at eval.c:624 [opt]
    frame #52368: 0x000000010015414c libR.dylib`R_execClosure(call=0x00000001073fe3e8, newrho=0x0000000103537498, sysparent=<unavailable>, rho=<unavailable>, arglist=<unavailable>, op=<unavailable>) at eval.c:0 [opt]
    frame #52369: 0x0000000100141d0c libR.dylib`Rf_eval(e=0x00000001073fe618, rho=0x0000000101851b98) at eval.c:747 [opt]
    frame #52370: 0x0000000100184bda libR.dylib`Rf_ReplIteration(rho=0x0000000101851b98, savestack=<unavailable>, browselevel=0, state=0x00007fff5fbfe700) at main.c:258 [opt]
    frame #52371: 0x00000001001861bf libR.dylib`run_Rmainloop [inlined] R_ReplConsole(rho=<unavailable>, savestack=0, browselevel=0) at main.c:308 [opt]
    frame #52372: 0x0000000100186156 libR.dylib`run_Rmainloop at main.c:1082 [opt]
    frame #52373: 0x0000000100000f5b R`main + 27
    frame #52374: 0x00007fffd1e62235 libdyld.dylib`start + 1
    frame #52375: 0x00007fffd1e62235 libdyld.dylib`start + 1
(lldb)

[jefferis]

Thanks for the bug report. There seems to be an issue with the ANN library's bd method when there are duplicate points in the target matrix. It's not really something I can debug in a hurry and I have previously considers inactivating the bd method for this reason.

[gdkrmr]

you could simply test for duplicated points if the user chooses "bd"

[jefferis]

This is exactly what is done by another ANN wrapper:

https://github.com/cran/yaImpute/blob/master/R/ann.R#L17

but the time spent for the check makes using the bd tree pointless. So that is why I hesitated the last time this bug was reported. My feeling is that if the functionality is either unsafe or uneconomically slow it should just be eliminated.

The only other alternative I see is to add an argument like bd.check.dups=T to give people the option to trade safety for speed if they are convinced there are no duplicates.

[krlmlr]

bd tree should really be able to handle this. Can we request a patch upstream, or do it ourselves?

[jefferis]

You're right of course @krlmlr. I'm on holiday right now so won't be doing this myself, but if you want to try, I think you could contact the senior author of ANN for input. Best, Greg.

Sent from my iPhone

On 31 Aug 2016, at 12:36, Kirill Müller [email protected] wrote:

bd tree should really be able to handle this. Can we request a patch upstream, or do it ourselves?

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub, or mute the thread.

[twolodzko]

@jefferis I wanted to ask if there was any progress on this? The bug seems to persist in the CRAN version. At minimum, there should be a safeguard for the code not to give segfaults.

[Jean-Romain]

Hi, In my case I get a segfault with bd.

[jefferis]

To reiterate, the bug is in the upstream library and I don't have time to chase it down. There must be a comparison logic error somewhere that is revealed with two identical points.

I can either remove bd functionality altogether or provide a check for identical points that can be turned off at the user's discretion. But as a reminder, checking for identical points is so expensive that it makes the bd tree pointless.

Votes?

[twolodzko]

@jefferis then maybe: remove, document it in help and e-mail the author of the algorithm?

[Jean-Romain]

Well a segfault is a big problem! I think you could document this bug. I don't know how bd can speed up the search compared to kd (indeed it crashed) but if the gain is big, document the bug, otherwise remove the feature.

[Jean-Romain]

I made few benchmark with 300k 3D points

RANN::nn2(X, X, k = k, treetype = "kd", searchtype = "radius", radius = 2)

• Check for duplicates 22 ms
• k = 10
  • kd: 1.0 s
  • bd: 1.6 s
• k = 60
  • kd: 1.6 s
  • bd: 2.2 s

Maybe I miss something. Maybe 300k is not big enough to take advantage of bd. In any case checking for duplicates is ways faster than the search. I made other tests and kdout performed bd every time.

[jefferis]

I haven't had a compelling use case for bd but I expect there will be some situation in which it is faster. How are you checking for dupes and what is your estimate for the tree build time (can estimate by time taken to search for just one point)?

…

Sent from my iPhone

On 2 Oct 2017, at 22:22, Jean-Romain Roussel ***@***.***> wrote: I made few benchmark with 300k 3D points RANN::nn2(X, X, k = k, treetype = "kd", searchtype = "radius", radius = 2) Check for duplicates 22 ms nn2 with kd 1.0 s for k = 10 nn2 with bd 1.6 s for k = 10 nn2 with kd 1.6 s for k = 60 nn2 with bd 2.2 s for k = 60 Maybe I miss something. Maybe 300k is not big enough to take advantage of bd. In any case checking for duplicates is ways faster than the search. I made other tests and kdout performed bd every time. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[Jean-Romain]

I check dupes with the data.table tools.

  dupes  = duplicated(X, by = c("x", "y", "z"))

x being a data.table. Then I removed the dupes to use bd. I did not estimate the tree build time. I only estimated the total search time (build + search). The ultimate goal being to speed up the search I tried to estimated how much time is lost in finding duplicates against how much time is gained using bd instead of kd in a given situation. For the last one the result was not the one expected.

[MattMyint]

Hi,

Just as some extra info, this issue pops up in another function that uses nn2, but it seems in that case, there don't seem to be any duplicate entries in that matrix.

These are the steps I took

> sample_expr1 <- read.table("~/sample_expr_matrix1.txt", header = TRUE, stringsAsFactors = FALSE)
> any(duplicated(sample_expr1))
[1] FALSE
> dim(sample_expr1)
[1] 7615   15
> dim(unique(sample_expr1))
[1] 7615   15

Perhaps I'm searching for duplicates the wrong way? Either way, we've switched to using kd

[Jean-Romain]

You are search for duplicated over the 15 columns. I don't have a clue of what your data contains but I guess over the 15 columns there are no duplicates i.e not line with twice the same 15 numbers. The question is: is there duplicates over the 2 or 3 columns used as coordinates.