On Sun, Sep 12, 2010 at 12:12 PM, Jeff Eastman
<[email protected]>wrote:

>  Wow, thanks Jake, that is really definitive. There's a lot of gears
> whirring silently under the hood here. So Grant's command line script would
> need to add a "mahout cleaneigens <...>" line to really get the valid
> eigenvectors (or use the second approach). And too, that step will yield the
> eigenvalues so that either decomposition approach may be used. The problem
> I'm seeing with doing this programmatically is there is no ready Java method
> (e.g. job() called by run() in the clustering stuff) which I could use.
>

Not a problem.  Sorry to be out of this loop so long.


> +1 on folding EigenVerificationJob into DistributedLanczosSolver. Or, at
> least implement a job() method on EVJ.
>

+1 for having the latter, with a boolean flag in DLS to optionally call EJV
after it's done.


> +1 on renaming DistributedMatrix.times() to transposeTimes() to avoid
> confusion.
>

+1  I actually though I'd put it even in the javadocs, but apparently I
didn't.


> +1 on adding DistributedMatrix.timesDiagonal(Matrix) and perhaps also
> timesDiagonal(Vector)? Perhaps after the 20.2 retrofit?
>

Like I replied to Ted, I think we can have a transparent api and just check
whether the matrix we've been passed is able to be passed around
via-side-channel serialization to the M/R nodes.

  -jake

Reply via email to