RE: [eigen] overriding scalar_X_op for auto-dif?

[ Thread Index | Date Index | More Archives ]

I have a similar problem here. Although auto-dif is an excellent idea to calculate partial derivatives, it requires all arithmetic ops on the paired data object (if that's the case, usually). I can't image how slow my program will become if I do that, not even mentioning the work load to rewrite LAPACK routines or ODE solvers for the paired data class. Any smart suggestion to overcome this? Thanks.


-----Original Message-----
From: Listengine [mailto:listengine@xxxxxxxxxxxxxxxxx] On Behalf Of Bob Carpenter
Sent: Monday, May 23, 2011 2:09 PM
To: eigen@xxxxxxxxxxxxxxxxxxx
Subject: Re: [eigen] overriding scalar_X_op for auto-dif?

I'm glad to hear that a solution is possible.

Is there any way I could help?

It's important to us because it's holding up the efficient use of matrices in multivariate distributions like the normal, and we do a lot of hierarchical regression modeling.

We can code around this by promoting everything to an auto-dif variable, but it's hugely wasteful in both time and space.  Eventually, we plan to override the matrix arithmetic ops for auto-dif to save even more space and time, but we want to get everything working before worrying about further optimization.

- Bob

On 5/19/11 5:42 PM, Gael Guennebaud wrote:
> ok, actually it seems this is something we have to solve in Eigen 
> itself. Nothing difficult though.
> gael
> On Thu, May 19, 2011 at 10:48 PM, Bob Carpenter<carp@xxxxxxxxxxx>  wrote:
>> Thanks so much for getting back to me so quickly on this.  I tried 
>> overloading scalar_product_traits and NumTraits, but still no luck 
>> getting multiplication of an auto-dif matrix times a double matrix to 
>> compile.
>> I'd MUCH rather just overload the basic ops for now.
>> Later,  I can optimize with expressions and hand-coded gradients for 
>> some cases without disturbing the API.
>> I'm afraid your auto-dif scalar code's a bit beyond my current 
>> understanding of Eigen and C++ templates.
>> I've attached a small tar.gz ball.  Check out the README.txt, which 
>> shows how to run our reverse-mode auto-dif, has an example using 
>> Eigen auto-dif var multiplication which works fine, and then a demo 
>> that doesn't compile using double * auto-dif variables.
>> I include the specialization of scalar_product_traits you suggested 
>> (with both orders), as well as NumTraits, but still no luck.
>> - Bob Carpenter
>>   Columbia Uni, Dept of Statistics

Mail converted by MHonArc 2.6.19+