Missing local R-squared and residuals in gwr output

On Wed, 9 May 2012, Maximilian Spro? wrote:

Thank you Roger! The gwr on the MPI cluster works fine.

However, now the output object includes the intially missing three 
data slots: "gwr.e","pred" and "localR2". Unfortunately, the latter 
contains only NA's.
Sorry for of any inconvenience, but do you think you can solve that?

I do not see any problem there, and indeed it is after the results 
have been returned from the cluster. You can tell whether you have 
been into the code block starting on line 261 in spgwr/R/gwr.R if 
there is no line beginning with "postprocess_localR2" in the timings 
component of the output object. The conditions are:

((!fp.given || fit_are_data) && is.null(fittedGWRobject))

where the first is FALSE, the second TRUE and the third TRUE in your 
case. If the "pred" column in your output contains values that are not 
finite, this may happen in this code block.

If you cannot see what is going on, we need a smaller test data set 
that replicates the problem.

Roger

Thanks in advance and all the best,

Max


On 05/07/2012 08:45 PM, Roger Bivand wrote:

On Mon, 7 May 2012, "Spro?, Johann" wrote:

-- 
Mag. J. Maximilian Spro?
Institute of Geography, University of Innsbruck
Innrain 52
A-6020 INNSBRUCK

Tel. +43 (0)512 507 5413
web: http://www.uibk.ac.at/geographie/projects/lidar/



-----Urspr?ngliche Nachricht-----
Von: Roger Bivand [mailto:Roger.Bivand at nhh.no]
Gesendet: Mo 07.05.2012 14:48
An: Maximilian Spro?
Cc: r-sig-geo
Betreff: Re: [R-sig-Geo] Missing local R-squared and residuals in 
gwr output

On Mon, 7 May 2012, Maximilian Spro? wrote:

Dear Roger!

Thank you very much for your fast reply and work!

I'm not really an expert in HPC-computing, but i will try to 
report as goog
as i can.

I updated spgwr and started a job on the cluster which takes 
normally 1,5 h.
So far, it run for 5 hours, which indicates that the 
parallelization does not
work efficient anymore. The function makeCluster(64, type="MPI") 
worked fine.
Our cluster runs with openMPI.

Correct. I'll try to add back an option to use snow instead of 
parallel.

I tried out the new version but it seems still using parallel.

code:

gwr_50 <- 
gwr(hef at data$DIF~hef at data$ELEVATION+hef at data$SKY+hef at data$SLOPE+hef at data$SOLAR+factor(asp_fac), 
data=hef, bandwidth=50, gweight=gwr.Gauss,fit.points=coords, 
hatmatrix=FALSE, cl=cl)

Add use_snow=TRUE to the command to switch to snow.

Roger

Loading required package: parallel

Attaching package: 'parallel'

The following object(s) are masked from 'package:snow':

   clusterApply, clusterApplyLB, clusterCall, clusterEvalQ,
   clusterExport, clusterMap, clusterSplit, makeCluster, parApply,
   parCapply, parLapply, parRapply, parSapply, splitIndices,
   stopCluster

Max


When it reaches R-forge, its revision number will be > 1252.

Roger

In that context, i found on the CRAN Task view: High-Performance 
and Parallel
Computing with R the following:
"<http://www.dict.cc/englisch-deutsch/parallelization.html>Direct 
support in
R is starting with release 2.14.0 which includes a new package 
parallel
incorporating (slightly revised) copies of packages multicore and 
snow (*but
excluding MPI, PVM and NWS clusters*). Does the new parallel 
support works
still in the openMPI environment?

regards,

Max

fyi:

sessionInfo()
R version 2.14.0 (2011-10-31)
Platform: x86_64-unknown-linux-gnu (64-bit)

locale:
[1] LC_CTYPE=en_US       LC_NUMERIC=C         LC_TIME=en_US
[4] LC_COLLATE=en_US     LC_MONETARY=en_US    LC_MESSAGES=en_US
[7] LC_PAPER=C           LC_NAME=C            LC_ADDRESS=C
[10] LC_TELEPHONE=C       LC_MEASUREMENT=en_US LC_IDENTIFICATION=C

attached base packages:
[1] parallel  stats     graphics  grDevices utils     datasets  
methods
[8] base

other attached packages:
[1] spgwr_0.6-15    spdep_0.5-45    coda_0.14-6     deldir_0.0-16
[5] maptools_0.8-10 foreign_0.8-46  nlme_3.1-102    MASS_7.3-16
[9] Matrix_1.0-1    lattice_0.20-0  boot_1.3-3      gstat_1.0-10
[13] spacetime_0.5-7 xts_0.8-2       zoo_1.7-6       sp_0.9-98
[17] snow_0.3-8      Rmpi_0.5-9

loaded via a namespace (and not attached):
[1] grid_2.14.0


On 05/05/2012 04:24 PM, Roger Bivand wrote:

On Fri, 4 May 2012, Maximilian Spro? wrote:

Dear r-sig-geo list!

I run gwr on a multi-node cluster(on 64 slots). In the gwr 
output (slot
"SDF"), the gwr residuals and the local R-squared are missing. When
performing the same model on the local machine, these components 
are
included. Unfortunately, the calculation in this way takes about 
5 days
instead of few hours when using the cluster.

Perhaps, that problem arises due to the argument "fit.points", 
which has
to be passed if the local coefficient estimates should be made on a
multi node cluster.

Does anyone have an idea how to solve that problem with the missing
local R-squared and residuals if the gwr is calculated on a 
cluster?

The understanding for use on a cluster was that the data points 
and the fit
points are different, so there is no observed dependent variable 
at the fit
point, hence no local R2. I've added logic in the code that 
checks for
equality between the fit and data points, and this for me 
resolves the
problem, but may break other things. I've committed to R-forge, 
project
rspatial, module spgwr. The source tarball and binary packages 
should be
available later this evening European time from:

https://r-forge.r-project.org/R/?group_id=1014

Could you please try it out, and report back? I should also 
migrate spgwr
from snow to parallel before I release it.

Best wishes,

Roger

Thank you very much in advance!

Kind regards,

Max


selected R-code:

### gwr on local machine:

gwr_50 <-
gwr(hef at data$DIF~hef at data$ELEVATION+hef at data$SKY+hef at data$SLOPE+hef at data$SOLAR, 
data=hef, bandwidth=50, gweight=gwr.Gauss)


# part of the  str(gwr_50) output...


List of 11
 $ SDF      :Formal class 'SpatialPointsDataFrame' [package 
"sp"] with 5
slots
  .. ..@ data       :'data.frame':    286288 obs. of  9 variables:
  .. .. ..$ sum.w      : num [1:286288] 2009 2003 2091 2089 2086 
...
  .. .. ..$ (Intercept): num [1:286288] -28.7 -28.5 -29.9 -29.7 
-29.5 ...
  .. .. ..$ elevation  : num [1:286288] 0.0139 0.0138 0.014 
0.014 0.014
...
  .. .. ..$ sky        : num [1:286288] -0.153 -0.155 -0.146 
-0.148 -0.149
...
  .. .. ..$ slope      : num [1:286288] -2.58 -2.61 -2.42 -2.45 
-2.48 ...
  .. .. ..$ solar      : num [1:286288] -0.00139 -0.00136 
-0.0015 -0.00147
-0.00144 ...
  .. .. ..$ gwr.e      : num [1:286288] -0.461 -0.683 -0.5987 
-0.2692
0.0406 ...
  .. .. ..$ pred       : num [1:286288] 0.806 0.833 0.507 0.514 
0.576 ...
  .. .. ..$ localR2    : num [1:286288] 0.621 0.618 0.638 0.635 
0.632 ...




### gwr on cluster :

cl <- makeCluster(32, type="MPI")

coords <- coordinates(hef)

gw <-
gwr(hef at data$DIF~hef at data$ELEVATION+hef at data$SKY+hef at data$SLOPE+hef at data$SOLAR, 
data=hef, bandwidth=50, gweight=gwr.Gauss,fit.points=coords,
hatmatrix=FALSE, cl=cl)

# part of the  str(gwr_50) output...

List of 11
 $ SDF      :Formal class 'SpatialPointsDataFrame' [package 
"sp"] with 5
slots
  .. ..@ data       :'data.frame':    286288 obs. of  6 variables:
  .. .. ..$ sum.w      : num [1:286288] 1 1 1 1 1 ...
  .. .. ..$ (Intercept): num [1:286288] 12541 1970 2057 -1505 
-1030 ...
  .. .. ..$ elevation  : num [1:286288] -3.891 -0.602 -0.738 
0.465 0.309
...
  .. .. ..$ sky        : num [1:286288] -0.954 -0.425 3.714 
0.159 0.152
...
  .. .. ..$ slope      : num [1:286288] 62.19 NA -27.21 1.95 
16.03 ...
  .. .. ..$ solar      : num [1:286288] NA NA NA NA 0.042 ...




    [[alternative HTML version deleted]]

-- 
Roger Bivand
Department of Economics, NHH Norwegian School of Economics,
Helleveien 30, N-5045 Bergen, Norway.
voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no