[Gsll-devel] Patch that exports symbols required to use non-linear fitting outside the gsl package
Hmm, so it does. I'm wondering if fit-test-delta is always called with last-step and current-position derived from the same fit object. In which case, would it make sense to just have one argument there, the fit object? E.g. (fit-test-delta fit 1.0d-4 1.0d-4) What do you think?
Liam
On Wed, Apr 21, 2010 at 10:15 PM, Russell Kliese russell@kliese.id.au wrote:
Hi Liam,
The example included at the bottom of gsll/solve-minimize-fit/nonlinear-least-squares.lisp, line 373 (fit-test-delta (last-step fit) (mpointer (solution fit)) 1.0d-4 1.0d-4)) uses (mpointer (solution fit)). Because mpointer isn't exported, I thought I would create a specific function like last-step and jacobian that return raw pointers. I don't really have a good working knowledge of the internals of gsll so this may be a naive fix.
Cheers,
Russell
2010/4/22 Liam Healy lhealy@common-lisp.net:
Russell,
Can you give an example showing why this function is needed? If I understand correctly, #'solution gives an marray which can be used like any other marray. As a general practice I don't define functions returning raw pointers unless the exclusive use of the returned value from the original function use is to pass to a GSL function that must take a raw pointer, and that is very rare.
Liam
Yes, I was thinking the same thing. I created a new patch and tested the example.
There is also the issue of the ls-covariance function requiring a raw pointer to the Jacobin matrix. I just exported the jacobian function in the attached patch.
Cheers,
Russell
On Thu, 2010-04-22 at 22:45 -0400, Liam Healy wrote:
Hmm, so it does. I'm wondering if fit-test-delta is always called with last-step and current-position derived from the same fit object. In which case, would it make sense to just have one argument there, the fit object? E.g. (fit-test-delta fit 1.0d-4 1.0d-4) What do you think?
Liam
On Wed, Apr 21, 2010 at 10:15 PM, Russell Kliese russell@kliese.id.au wrote:
Hi Liam,
The example included at the bottom of gsll/solve-minimize-fit/nonlinear-least-squares.lisp, line 373 (fit-test-delta (last-step fit) (mpointer (solution fit)) 1.0d-4 1.0d-4)) uses (mpointer (solution fit)). Because mpointer isn't exported, I thought I would create a specific function like last-step and jacobian that return raw pointers. I don't really have a good working knowledge of the internals of gsll so this may be a naive fix.
Cheers,
Russell
2010/4/22 Liam Healy lhealy@common-lisp.net:
Russell,
Can you give an example showing why this function is needed? If I understand correctly, #'solution gives an marray which can be used like any other marray. As a general practice I don't define functions returning raw pointers unless the exclusive use of the returned value from the original function use is to pass to a GSL function that must take a raw pointer, and that is very rare.
Liam
OK, I've applied your patch and made some further changes. I have generally followed a policy of making objects that are to be returned by the function be optional arguments, with the correctly sized object being made by default, thereby sparing the programmer from having to think about it. Toward this end, the covariance argument to ls-covariance is now an optional argument. In order to extract the dimension, I needed to pass in the solver object as the first argument to ls-covariance, so it calls #'jacobian instead of the programmer having to. So I've unexported it, as that's now all taken care of. I converted the example so that it no longer creates the covariance matrix. The test passes. There is probably more improvements that can be made along these lines, but I'll let this go for now.
[Note: for entirely independent reasons, you'll have to do fresh pull of GSD when you pull the new GSLL, if you are a SBCL user.]
Liam
On Fri, Apr 23, 2010 at 10:35 PM, Russell Kliese russell@kliese.id.au wrote:
Yes, I was thinking the same thing. I created a new patch and tested the example.
There is also the issue of the ls-covariance function requiring a raw pointer to the Jacobin matrix. I just exported the jacobian function in the attached patch.
Cheers,
Russell
On Thu, 2010-04-22 at 22:45 -0400, Liam Healy wrote:
Hmm, so it does. I'm wondering if fit-test-delta is always called with last-step and current-position derived from the same fit object. In which case, would it make sense to just have one argument there, the fit object? E.g. (fit-test-delta fit 1.0d-4 1.0d-4) What do you think?
Liam
On Wed, Apr 21, 2010 at 10:15 PM, Russell Kliese russell@kliese.id.au wrote:
Hi Liam,
The example included at the bottom of gsll/solve-minimize-fit/nonlinear-least-squares.lisp, line 373 (fit-test-delta (last-step fit) (mpointer (solution fit)) 1.0d-4 1.0d-4)) uses (mpointer (solution fit)). Because mpointer isn't exported, I thought I would create a specific function like last-step and jacobian that return raw pointers. I don't really have a good working knowledge of the internals of gsll so this may be a naive fix.
Cheers,
Russell
2010/4/22 Liam Healy lhealy@common-lisp.net:
Russell,
Can you give an example showing why this function is needed? If I understand correctly, #'solution gives an marray which can be used like any other marray. As a general practice I don't define functions returning raw pointers unless the exclusive use of the returned value from the original function use is to pass to a GSL function that must take a raw pointer, and that is very rare.
Liam
-
Liam Healy -
Russell Kliese