Mark Bates comments

Results 40 comments of


                                            Mark Bates

CPU (scipy.optimize.curve fit) and GPU fit results are different.

It's strange, because the LM algorithm implemented in Gpufit has been tested extensively, and is as accurate as scipy, etc. Sometimes optimization routines can fail when there is zero noise...

If fit goes wrong convergence test can report positively although fit didn't converge

@jkfindeisen Please post an example code which reproduces the problem.

Criteria for pull requests of new models?

We are discussing options for the simplification of Gpufit customization. For the time being, we will not be merging new models into the master branch.

Criteria for pull requests of new models?

Fully agree with this. We have tried using function pointers in CUDA for passing references to the model functions, but there was a huge cost in processing speed which made...

Can GPU fit handle functions that returns vectors instead of a single value?

Can you be more specific? In general, a multidimensional fit can be reformulated as a normal, scalar fit, and solved accordingly.

Can GPU fit handle functions that returns vectors instead of a single value?

Hi, I think what you're asking for is already possible using the existing Gpufit framework. Your function returns a vector of values, let's say 3 values, and you are fitting...

Can GPU fit handle functions that returns vectors instead of a single value?

> I assume I'll need to create a function like this as unit test for my model, which convey the generation of samples according to the layout you mentioned, is...

Can GPU fit handle functions that returns vectors instead of a single value?

Continuing with the example, the final size of your derivatives array would be n_parameters * n_points * length_of_function_vector (three in this case).

Can GPU fit handle functions that returns vectors instead of a single value?

To see how to organize the derivatives, you can look into e.g. the Gaussian model function code. e.g. http://github.com/gpufit/Gpufit/blob/master/Gpufit/models/gauss_1d.cuh

Can GPU fit handle functions that returns vectors instead of a single value?

You don't need to calculate the Hessian - you only want the derivative of the function with respect to each of the parameters. You are essentially treating your function, which...