How to create an iteration callback function in CasADi v3.0.0

[Silvia Manara]

Hi everybody,

I am using CasADi v3.0.0 with the Python interface and I would like to create a callback function such that it allows me to see what's going on with the solution at each iteration of IPOPT.
In CasADi v2.4.0 I used to do this by using:

@pycallback
def current_solution(f):
    opt = V(f.getOutput("x"))
    return 0

[...]

nlp = MXFunction( nlpIn( x = V ), nlpOut( f = f, g = g ) )
nlp.init()
solver = NlpSolver( "ipopt", nlp )
solver.setOption("iteration_callback", current_solution)

I tried to do something analogous with the new version of CasADi, but it does not seem to work.. What I get is the following error:

"RuntimeError:  on line 254 of file "/home/travis/build/casadi/binaries/casadi/casadi/interfaces/ipopt/ipopt_interface.cpp"
No such IPOPT option: iteration_callback"

Does anybody know how to tackle the problem?

Best regards,
Silvia

[Joris Gillis]

Dear Silvia,

The way we implement Callbacks internally drastically altered with 3.0.0.
This caused a syntax change.
It's easiest if you pattern-match off the example: http://casadi.sourceforge.net/api/html/d7/df0/solvers_2callback_8py-example.html

Best regards,
  Joris

[Mikhail Katliar]

Hi Joris,

Does the callback mechanism also work in MATLAB?

вторник, 28 июня 2016 г., 16:38:22 UTC+2 пользователь Joris Gillis написал:

[Peter Listov]

Hello Joris,

Do I understand correctly, that the "iteration_callback" functionality available only for CasADi and IPOPT built from sources? (as 'enableipoptCallback' fix might imply) 

Best regards,

Peter

[Joris Gillis]

The functionality should work from the binaries, and on Matlab as well.

Best,
  Joris

[Peter Listov]

I then got a bit confused with assertion I am getting when trying to create an iteration callback in Matlab (v3.0 binaries):

Error using casadiMEX

The assertion "fcallback_.size_in(i)==size_out(i)" on line 245 of file "/Users/travis/build/casadi/binaries/casadi/casadi/core/function/nlpsol.cpp" failed.

Callback function input size mismatch

Here is code for Callback class I use. It generally mimics the aforementioned python tutorial:

classdef IterationCallback < casadi.Callback

  properties

    nx

    ng

    np

  end

  methods

    function self = IterationCallback(name, nx, ng, np)

      

      se...@casadi.Callback();

 

      self.nx = nx;

      self.ng = ng;

      self.np = np;

      

      opts.input_scheme = casadi.nlpsol_out();

      opts.output_scheme = 'ret';

      self.construct(name,opts);

      disp('Callback instance created')

    end

    

    function v=get_n_in(self)

        v=casadi.nlpsol_n_out;

    end

    function v = get_n_out(self)

        v = 1;

    end

    

    function v = get_sparsity_in(self, i)

      n = casadi.nlpsol_out(i);

      if n=='f'

        v =  casadi.Sparsity.scalar();

      elseif strcmp(n,'x') || strcmp(n,'lam_x')

        v = casadi.Sparsity.dense(self.nx);

      elseif strcmp(n,'g') || strcmp(n,'lam_g')

        v = casadi.Sparsity.dense(self.ng);

      elseif strcmp(n,'lam_p')

        v = casadi.Sparsity.dense(self.np);

      else

        v = casadi.Sparsity(0,0);

      end

    end

 

    function v = eval(self, arg)

      disp('Solver iteration')

      v = {0};

    end

  end

end

I see it comes from sanity assertion in nlpsol.cpp but further output is not available in Matlab. Would be very great full for any hint why is it happening.

Best regards,

Peter

[Joris Gillis]

Let's debug.
In get_sparsity_in, make it print 'i' and 'n'.
Also, give a printout of solver.print_dimensions()/solver.printDimensions() when you disable the callback.

We are fully aware that the Callbacks are not userfriendly yet.

Best,
  Joris

[Peter Listov]

Thanks a lot for helping out.

solver.printDimensions() output is:

 Number of inputs: 8

  Input 0, a.k.a. "x0", (243, 1)

  Input 1, a.k.a. "p", (0, 0)

  Input 2, a.k.a. "lbx", (243, 1)

  Input 3, a.k.a. "ubx", (243, 1)

  Input 4, a.k.a. "lbg", (180, 1)

  Input 5, a.k.a. "ubg", (180, 1)

  Input 6, a.k.a. "lam_x0", (243, 1)

  Input 7, a.k.a. "lam_g0", (180, 1)

 Number of outputs: 6

  Output 0, a.k.a. "x", (243, 1)

  Output 1, a.k.a. "f", (1, 1)

  Output 2, a.k.a. "g", (180, 1)

  Output 3, a.k.a. "lam_x", (243, 1)

  Output 4, a.k.a. "lam_g", (180, 1)

  Output 5, a.k.a. "lam_p", (0, 0)

Here is output on callback creation:

get_n_in

v = 6

get_n_out

Callback instance created

It does not seem to call 'get_sparsity_in' function.

Best regards,

Peter

[Peter Listov]

Hello Joris,

I also implemented an example from the tutorial:

x   = SX.sym('x');

y   = SX.sym('y');

f   = (1-x)^2+100*(y-x^2)^2;

nlp = struct('x',vertcat(x,y), 'f', f, 'g', x+y);

fcn = Function('f', {x, y}, {f});

mycallback = IterationCallback('mycallback', 2, 1, 0);

opts = struct;

opts.iteration_callback = mycallback;

opts.ipopt.tol = 1e-8;

opts.ipopt.max_iter = 50;

solver = nlpsol('solver', 'ipopt', nlp, opts);

args = struct;

args.lbx = -10;

args.ubx = 10;

args.lbg = -10;

args.ubg = 10;

sol = solver(args)

Unfortunately it also does not pass "size_in(i) == size_out(i)" assertion.