Skip to content

Parallel predict now in spatial.tools

4 messages · Tim Howard, Jonathan Greenberg

Original
Tim Howard
# 
Jonathan,
Thank you for putting this together and for the example. I'm doing two things differently with randomForest ... I think perhaps one of them the function isn't handling.  
 
First, based on recommendations from Andy Liaw (and ?randomForest), I don't use the formula interface but use x=<many columns>, y=<a column> in the call. Does predict_rasterEngine handle the absence of a formula in the object?
 
Second, I have many large rasters I want to run the predict on, so making a brick would be difficult. I use a rasterStack instead. Does your example work with a rasterStack? 
 
I can dive deeper if any of this isn't clear or if these two tweaks work just fine for you. I was just trying to swap out this version of predict with another parallel version to evaluate speed and, while the alternate version works fine, predict_rasterEngine bailed on me. 
 
Thanks in advance. 
Tim Howard
Date: Tue, 18 Mar 2014 22:14:23 -0500
From: Jonathan Greenberg <jgrn at illinois.edu>
To: "r-sig-geo at r-project.org" <R-sig-Geo at r-project.org>
Subject: [R-sig-Geo] Parallel predict now in spatial.tools
Message-ID:
<CABG0rfseg+p0h4HdYOK+_Za=OLMeTKHAT+TQn7g_FkEdYiunFQ at mail.gmail.com>
Content-Type: text/plain; charset=ISO-8859-1

R-sig-geo'ers:

I finally got around to building a parallel predict statement that
I've included in version 1.3.7 (or later) of spatial.tools (check
http://r-forge.r-project.org/R/?group_id=1492 for the status of the
build), "predict_rasterEngine".  It should, in theory, be a direct
swap-in for the standard generic predict() statement.  Currently, it
will work on any predict.* statement that has the following features:
1) The data is passed to the predict as a data frame via a newdata
parameter, and
2) The data is returned from the predict statement as a vector/matrix.

When using predict_rasterEngine, the object= parameter is your model,
and the newdata= parameter is the raster/brick/stack to apply the
model to on a pixel-by-pixel basis (note that the names of the layers
must match the names of the predictor variables, in most cases).

I was hoping to get some stress-testing on this, since it is a fairly
oft-requested function.  If a predict.* function you'd like to use
doesn't work, let me know which function it is (with some test data)
and I'll see if I can tweak it to work.

Right now, I have confirmed this works with randomForest.  Here's an example:

######################

packages_required <- c("spatial.tools","doParallel","randomForest")
lapply(packages_required, require, character.only=T)

# Load up a 3-band image:
tahoe_highrez <- setMinMax(
brick(system.file("external/tahoe_highrez.tif", package="spatial.tools")))
tahoe_highrez
plotRGB(tahoe_highrez)

# Load up some training points:
tahoe_highrez_training_points <- readOGR(
dsn=system.file("external", package="spatial.tools"),
layer="tahoe_highrez_training_points")

# Extract data to train the randomForest model:
tahoe_highrez_training_extract <- extract(
tahoe_highrez,
tahoe_highrez_training_points,
df=TRUE)

# Fuse it back with the SPECIES info:
tahoe_highrez_training_extract$SPECIES <- tahoe_highrez_training_points$SPECIES

# Note the names of the bands:
names(tahoe_highrez_training_extract) # the extracted data
names(tahoe_highrez) # the brick

# Generate a randomForest model:
tahoe_rf <- randomForest(SPECIES~tahoe_highrez.1+tahoe_highrez.2+tahoe_highrez.3,
data=tahoe_highrez_training_extract)

tahoe_rf

# This will run the predict in parallel:
sfQuickInit()
prediction_rf_class <-
predict_rasterEngine(object=tahoe_rf,newdata=tahoe_highrez,type="response")
prediction_rf_prob <-
predict_rasterEngine(object=tahoe_rf,newdata=tahoe_highrez,type="prob")
sfQuickStop()

###############

--j
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20140326/360372bf/attachment.html>
Jonathan Greenberg
# 
Hi Tim:

Re: stack, yep, it should work.  In general, you get a decreased
performance from stacks since you are having to read from multiple
files rather than a single one, but it will still benefit from
parallel processing.

Re: formula -- the best way for me to test this is to crop out a piece
of your image and send me the random forest model (use ?save), and the
call you were using.  In theory you should be able to use anything you
would normally use on a data frame, but I'd have to play with it to
confirm!  If you can set up those on e.g. google drive I can test it
out.  Cheers!

--j
On Wed, Mar 26, 2014 at 6:49 AM, Tim Howard <tghoward at gw.dec.state.ny.us> wrote:

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Tim Howard
    

    

      
      #
    

  


  

      
Jonathan,
Thanks for your quick reply. I just tried your Tahoe example using x and y in the call rather than a formula and it worked fine, so I was barking up the wrong tree. Sorry. I'll try to delve deeper and put together a subset example for you to check out if I can't get anywhere. 
 
Thanks,
Tim

            

      
      
      
    

        
Hi Tim:

Re: stack, yep, it should work.  In general, you get a decreased
performance from stacks since you are having to read from multiple
files rather than a single one, but it will still benefit from
parallel processing.

Re: formula -- the best way for me to test this is to crop out a piece
of your image and send me the random forest model (use ?save), and the
call you were using.  In theory you should be able to use anything you
would normally use on a data frame, but I'd have to play with it to
confirm!  If you can set up those on e.g. google drive I can test it
out.  Cheers!

--j

        
On Wed, Mar 26, 2014 at 6:49 AM, Tim Howard <tghoward at gw.dec.state.ny.us> wrote:

            

      
      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Jonathan Greenberg
    

    

      
      #
    

  


  

      
Tim:

One thing to make sure is if the stack has the bands named properly
(they need to match the formula), e.g.:
names(mystack) <- c("b1","b2",...)

If you can also pass along the error you are getting, that will go a
long way towards figuring out the issue!  Cheers!

--j

        
On Wed, Mar 26, 2014 at 9:48 AM, Tim Howard <tghoward at gw.dec.state.ny.us> wrote: