Re: [Cdk-devel] rJava and CDK

[Thomas K. <tho...@we...>]

> -----Urspr=FCngliche Nachricht-----
> Von: raj...@pr... [mailto:raj...@pr...]
> Gesendet: Mittwoch, 7. Juni 2006 15:15
> An: Thomas Kuhn
> Cc: 'Thomas Kuhn'; cdk...@li...
> Betreff: rJava and CDK
>=20
> On Wed, 2006-06-07 at 13:48 +0200, Thomas Kuhn wrote:
>=20
> > In the in the rModel.java class you use a .call method the execute R
> methods
> > like:
> > revaluator.call("saveModel", new Object[] { (Object)modelname,
> > (Object)filename });
> > But I can't find a similar method within the rJava project.
>=20
> The saveModel method is actually a function I have written. In fact =
the
> R-CDK interface has a number of functions I wrote that allow me to
> transfer complex R objects back to the Java side. In addition these
> functions, allow me to perform error checking or various =
pre-processing
> transforms on the data sent to an R method.
>=20
> If you look under src/org/openscience/cdk/qsar/model/data you'll find =
a
> number of R scripts. Some of them represent wrappers for specific
> modeling routines (lm, cnn, pls etc) and init_1.R provides some =
utility
> methods (such as saveModel, loadModel) etc.

I saw these scripts.=20

>=20
> A lot of the R code is necessary because of the nature of the R-Java
> bridge, so it might not be applicable to the rJava package.
>=20
> See the CDK News article for a more detailed description of what goes =
on
> under the hood.
>=20
> [ From what I have understood if you wanted to use R from within a CDK
> app, you need to use JRI (whereas rJava is for accessing Java from an =
R
> session) ]
=20
That's right. I have forgotten to mention this.=20

> Also if I understand correctly, the JRI API provides methods to call
> methods in an R session via the REngine class and classes for various
> types of R primitive objects (string, integer, numeric, factor, list).
>=20
> This would mean that for example if you are building a OLS model with
> lm() there is no easy way to get back an 'lm' object as a Java object.
> As far as I can see, you have to call lm() via REngine, and whenever =
you
> need something from the 'lm' object (such as say fitted(),
> coefficients() etc) you must make a call to the R session. This is
> certainly doable, but you need to have some mechanism of keeping track
> of objects in the R session.
>=20
> >
> > I understand I can only assign values to a symbol in R. Which I =
thought
> > would not be the big problem, I could assign the symbol first and =
than
> > execute the R method with this R symbol and the .eval method of the
> rJava
> > rengine.
>=20
> Correct.
>=20
> > But the problem is that I can only assign double[], int[], REXP,
> > and strings. Do I misunderstand this?
>=20
> For the case of 'building' models this is not necessarily a problem. =
In
> general all model building routines require matrix, data.frame, =
vector,
> or string arguments.
>=20
> Now, since you can only assign vectors, this means that on the R side
> you're going to have reshape the input vector to an appropriately
> dimensioned matrix (or data.frame if required).
>=20
> This implies you need to have wrapper methods on the R side. So =
instead
> of calling lm() you'd call a routine that you wrote called =
lm.wrapper()
> which would take a X vector rather than a X matrix, and would then
> reshape the X vector to a matrix and call lm().
>=20
> Now, your wrapper function would then have to assign the returned 'lm'
> object to the global environment (otherwise it would be lost after the
> wrapper method returns and you could not then query the result of =
lm())
> and essentially return the name of the 'lm' object to the Java caller.
>=20
> All this assumes that rJava does not allow one to make a R matrix =
object
> on the Java side.

As far as I understand this is his assumption correct.=20

> Of course an alternative approach without writing wrapper functions is
> to essentially perform the preprocessing using a sequence of eval(...)
> calls - which might be a better option a it would keep the code in one
> place, does not require fiddling with environments on the R side. But
> it's a pain to type :)
=20
It would really be a pain of type but I think it would be easier to =
maintain
or?
=20
> However as I noted above, there appears to be no way to get back =
complex
> R objects, so you must perform eval(...) calls each time you want to
> query some aspect of an R object (which is certainly a performance =
hit)

So what do you think should we do? Should we try to get a SJava version
running and try to get a reliable installation routine or tutorial for =
this
or should we switch to rJava which is easy to install but is still in an
early development status?