External MIMO function works in C++, not in Python

[Fb]

Hi all,

I'm trying to implement a C++ MIMO external function in casadi, and to import it in Python.

Let us say that my function is:

f : (x,y,z) -> (p,q).

For this example let us assume that x,y,z,p and q are 1x6 vectors.

In C++, the function is implemented following the nomenclature described in 5.3.

In particular, fname_n_in (=3), fname_n_out (=2), fname_sparsity_in (={(6,1,1);(6,1,1);(6,1,1)}), fname_sparsity_out  (={(6,1,1);(6,1,1)}) are implemented.

In C++ everything seems to work as desired.

The strange behavior that I'm experiencing in Python is as follows :

from casadi import *

f = external('f','mylib.so',{'enable_fd':True})

input = MX.sym('input',3,6)

x = input[0,:]

y = input[1,:]
z = input[2,:]

fx = Function('fx',[input],[f(input)])

// Here I should be able to pass (x,y,z) as an input but :
print(f.n_in()) // Out: 1, should be 3
print(f.n_out()) // Out: 1, should be 2

print(f.sparsity_in(0)) // Out: 1x1, should be 6x1

print(f.sparsity_in(1)) // Out: Error M_range_check
print(f.sparsity_out(0)) // Out: 1x1, should be 6x1

print(f.sparsity_in(1)) // Out: Error M_range_check

// Hence I'm only able to pass the full input matrix, and the it results in:

print(fx.sparsity_in(0)) // Out: 3x6

print(fx.sparsity_out(0)) // Out: 3x6, should at least be 2x6

// Last point, despite the dimension mismatch, it succeeds in computing a

// symbolic Finite-Difference Jacobian

J = Function('J',[x],[jacobian(f(x),x)])

print(J.sparsity_out(0)) // Out: 18x18, wich is consitent with the rest

In conclusion, I am confused about what is computed through f in Python, since the output should not be of the same dimension as the input.

And last point, I am not able to debug numerically since the following :

_q = np.array([1,2,3,4,5,6])

_dq = np.array([1,2,3,4,5,6])

_ddq = np.array([1,2,3,4,5,6])

_x = vertcat(_q,_dq,_ddq).reshape((3,6))

print(f(_x))

produces a runtime error :

.../casadi/core/function_internal.cpp:2815: Failed to evaluate 'eval_dm' for f:
.../casadi/core/function_internal.cpp:1418: 'eval', 'eval_dm' not defined for External

Any help or advice would really be appreciated !

Thanks for the support,

FB

[Fb]

It turns out that the proper way to load the external function in Python is to name it after the shared library :

f = external('mylib','mylib.so',{'enable_fd':True})

This way, it works ! I don't get why though...

FB