Python porting questions

[roya0045]

Greetings,

First, I would like to thank you for creating such a nice project with so many optimization algorythms!

I'm currently trying to port the online memory limited BFGS algorythm to python ( yes scipy has many algorythm coded in haskell but it's lacking some that I would like to test, namely this one.)

I've got most of it done but I'm confused about the gradient descent part using the weight on the input.

I'm not familiar with matlab and I want to know if calling .cost , .grad or anything else on the input (the variable problem) are integrated method in the problem code itself as a method or is it a matlab built-in?

P.S.:Thanks for all your hard work!

[hiroyuki-kasai]

Thank you for using the project.

I'm not familiar with matlab and I want to know if calling .cost , .grad or anything else on the
input (the variable problem) are integrated method in the problem code itself as a method or is > it a matlab built-in?

I am sorry that I cannot understand your question. Could you provide me the clearer question so that I can answer to?? I am looking forward to hearing from you soon.

Regards,

Hiro

[roya0045]

To be more clear here is what I'm trying to do using this method:

I have a random function with n parameters. The input being [X1,X2....Xn] and the output is only one float (y).

What I am trying to figure out is how to adapt the code. What should calculate the gradient, creating a n by n tensor with every possible y would be impossible. This problem doesn't have the usual mapping that we can see on line 226 of your obfgs code( grad = problem.grad(w,indice_j);) . In this case I would only have an input vector ( w ).

That way the optimizer would perform on any solutions.

[hiroyuki-kasai]

Hi Alex,

I thank you for your question.

I have not yet understood your question clearly. But, the below is my answer.

As you might already understand, the optimization solver (such as obfgs.m) and the problem are completely separately defined. As in demo.m file in the project,

problem = logistic_regression(data.x_train, data.y_train, data.x_test, data.y_test);

is defined for the "problem" definition.

Then, the solver is executed with this problem as

[w_svrg, info_svrg] = svrg(problem, options);

Thus, the solver, which is "svrg" in this demo.m, binds itself with the problem (logistic regression problem in this example).

Therefore, if the solver file calls like

f_val = problem.cost(w);

This calculates the cost value defined in the "problem" definition (in logistic regression problem in this example).

They are not the pure Matlab functions or libraries.

Does this make it clearer in the first place??

Thanks,

Hiro

[roya0045]

Greetings Hiro,

Yes this makes more sense, I am currently trying to find a way to move from predefined functions ( such as problem.cost, problem.grad and problem.full_grad) to find a general solver.

This would mean that for a function that inputs a vector of n lenght , it would output a scalar.:
y1=foo([2,5,8,6,7])
y1==2.56

y2=foo([5,7,6,2,10])
y2==1.25

My goal would be to optimize foo in order to obtain the biggest or slower value as the output.
But foo can be anything, from a simple function to a neural network or any other black box algorithms.

For this to work I would need a way to adapt the current code.

Do you have any idea how this could be done?

Thanks again for your precious time.