[SOLVED] Nonlinear Least Squares Problem without Constraints doesn't work properly???

[renepreto]

Hello, I'm trying to train some FeedForward Neural Networks wthi the dtrnlsp subroutines from MKL, but it seems not to work properly, I've been searching for an error for three days already and I just can't figure it out. So I think I'm not using it right. The files from my program are added here, It is small program and has the files.

Rede.f -> Starts the subroutines

CalcErro.f -> Calculates the error of the neural network

FF.f -> The Neural network itself, that write the weights, bias and the output for the function

Gaussiana.f -> The activation function (Gaussian)

Entradas.dat -> The input file, in this case a sinus in a half period

Makefile -> Self explanatory

Well, the problem is! it returns me Segmentation fault every time. I think is something wrong with the fvec, but I couldn't figure it out! As i'm not really sure that the files are here I will paste the files here too.

Thanks

Rene

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Program Rede

Implicit none

integer i,j,k,l,m,n,numneu
parameter (m=200,numneu=8,n=3*numneu+1)
double precision Ent(m,2),X(n),eps(6),rs
double precision fvec(m),fjac(m,n),jac_eps,r1,r2
integer iter1,iter2,RCI_Request,sucesso,st_cr,iter
parameter (iter1=10000,iter2=100,rs=1d2,jac_eps=1d-8)
include "/opt/intel/Compiler/11.1/064/mkl/include/mkl_rci.fi"
integer*8 handle
external CalcErro

Call Random_Seed()

open (unit=1, file='Entradas.dat', status='unknown')

do i=1,m
read(1,*)(Ent(i,j),j=1,2)
enddo

Call Random_Number(X)
X=(X*2.d0)-1.d0

eps(1)=1d-5
eps(2)=1d-5
eps(3)=1d-5
eps(4)=1d-5
eps(5)=1d-5
eps(6)=1d-5

fvec=0.d0
fjac=0.d0

if (dtrnlsp_init(handle,n,m,x,eps,iter1,iter2,rs)/=TR_SUCCESS)then
print*, 'Deu pau na hora de iniciar'
Call MKL_FREEBUFFERS
stop
end if

RCI_Request=0
sucesso=0

do while (sucesso==0)

if (dtrnlsp_solve(handle, fvec, fjac, RCI_Request)/= TR_SUCCESS) then
print*,'Deu pau na hora de resolver'
Call MKL_FREEBUFFERS
stop
end if

Select Case (RCI_Request)

Case (-1,-2,-3,-4,-5,-6)
Sucesso=1

Case(1)
Call CalcErro(X,fvec,Ent,m,n,numneu)

Case(2)
If(djacobi(CalcErro,n,m,fjac,X,jac_eps)/=TR_SUCCESS) then
Print*,'Deu pau na matriz Jacobiana'
Call MKL_FREEBUFFERS
stop
end if
EndSelect
enddo

if (dtrnlsp_get(handle,iter,st_cr,r1,r2)/= TR_SUCCESS) then
Print*,'Deu pau no get'
Call MKL_FREEBUFFERS
stop
end if

if (dtrnlsp_delete(handle)/= TR_SUCCESS) then
Print *, 'Deu pau na hora de apagar'
Call MKL_FREEBUFFERS
stop
end if

Call MKL_FREEBUFFERS
if (r2<1d-5) then
Print *,'Deu certo'
else
Print*, 'Falhou'
end if

Print *, st_cr

Call FF(X,Ent,m,n,numneu)


end Program Rede

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

c ******************************************************************
c Calculo do Erro
c ******************************************************************

c Subrotina utilizada para calcular o erro da rede
c O erro calculado da seguinte forma
c
c Erro=(target-(saida da rede))

Subroutine Calcerro(X,Erro,XY,n,m,numneu)

Implicit none
c
type neuronio
double precision entrada
double precision ativacao
end type neuronio

integer i,j,k,l,m,n,numneu
Double precision Pesos(2*numneu), Erro(n), Bias(numneu+1),X(m)
Double precision Gauss, XY(n,2), Saida
type(neuronio) H1(numneu), Or

Erro=0.d0

k=0
do i=1,2*numneu
k=k+1
Pesos(i)=X(k)
enddo

do i=1,numneu+1
k=k+1
Bias(i)=X(k)
enddo
do l=1,n

do i=1,numneu
H1(i)%entrada=0.d0
H1(i)%ativacao=0.d0
enddo
c
Or%entrada=0.d0
Or%ativacao=0.d0

do i=1,numneu
H1(i)%entrada=XY(l,1)*Pesos(i)
H1(i)%entrada=H1(i)%entrada+Bias(i)
H1(i)%ativacao=Gauss(H1(i)%entrada)
enddo

do i=(numneu+1),2*numneu
Or%entrada=Or%entrada+(H1(i-numneu)%ativacao*Pesos(i))
enddo

Or%entrada=Or%entrada+Bias(numneu+1)
Or%ativacao=Or%entrada

Saida=0.d0
Saida=Or%ativacao
Erro(l)=0.d0
Erro(l)=(XY(l,2)-Saida)
enddo

Print *, 'Passei por aqui'

return

end subroutine Calcerro

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Function Gauss(W)

Implicit none
double precision W, Gauss

Gauss=dexp(-(W**2.d0))

end Function Gauss

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

Subroutine FF(X,XY,n,m,numneu)

Implicit none
c
type neuronio
double precision entrada
double precision ativacao
end type neuronio

integer i,j,k,l,m,n,numneu
Double precision Pesos(2*numneu), Bias(numneu+1),X(3*numneu+1)
Double precision Gauss, XY(n,2), Saida(n)
type(neuronio) H1(numneu), Or

Saida=0.d0

k=0
do i=1,2*numneu
k=k+1
Pesos(i)=X(k)
enddo

do i=1,numneu+1
k=k+1
Bias(i)=X(k)
enddo
do l=1,n

do i=1,numneu
H1(i)%entrada=0.d0
H1(i)%ativacao=0.d0
enddo
c
Or%entrada=0.d0
Or%ativacao=0.d0

do i=1,numneu
H1(i)%entrada=XY(l,1)*Pesos(i)
H1(i)%entrada=H1(i)%entrada+Bias(i)
H1(i)%ativacao=Gauss(H1(i)%entrada)
enddo

do i=(numneu+1),2*numneu
Or%entrada=Or%entrada+(H1(i-numneu)%ativacao*Pesos(i))
enddo

Or%entrada=Or%entrada+Bias(numneu+1)
Or%ativacao=Or%entrada

Saida(l)=0.d0
Saida(l)=Gauss(Or%ativacao)
enddo

open (unit=5, file='Saida.dat', status='Unknown')

do i=1,n
write(5,*) XY(i,1),Saida(i)
enddo

open (unit=3, file='Pesos.dat', status='Unknown')

do i=1,2*numneu
write(3,*) Pesos(i)
enddo

open (unit=4, file='Bias.dat', status='Unknown')

do i=1,numneu+1
write(4,*) Bias(i)
enddo


return

end subroutine FF

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

SOURCE = Rede.f CalcErro.f Gaussiana.f FF.f

FFLAGS =

MKLINCLUDE= /opt/intel/Compiler/11.1/064/mkl/include

MKLPATH = /opt/intel/Compiler/11.1/064/mkl/lib/em64t/

LIBS = -L$(MKLPATH) -I$(MKLINCLUDE) -Wl,--start-group $(MKLPATH)libmkl_intel_lp64.a $(MKLPATH)libmkl_sequential.a $(MKLPATH)libmkl_core.a -Wl,--end-group -lpthread

FC = ifort
all:
$(FC) $(FFLAGS) $(SOURCE) $(LIBS) -o Rede

clean:
rm -rf Rede Pesos.dat Bias.dat Saida.dat

[ArturGuzik]

Segmentation fault means that you're trying to access a memory location that it is not allowed to access or in not allowed way (e.g. write to a read-only location). Check whether all arrays are correctly allocated and passed etc. Does the calling program "see" the FF function the right way. try to switch on automatic (or code it manually) interfaces for the routines, so compiler will have a chance to check couple of things for you.

If this doesn't help, then post here exact error message and place where it appears.

A.

[Nikita_S_Intel]

Hi Rene,

Ive looked over your codes. At first sight here is a problem with djacobi calling. The first parameter is pointer to function which should has the following interface:

fcn (m, n, x, f) with the following parameters:

m - INTEGER. Input parameter. Length of f

n - INTEGER. Input parameter. Length of x.

x - DOUBLE PRECISION. Input parameter. Array of size n. Vector, at which the function is evaluated. The fcn function should not change this parameter.

f - DOUBLE PRECISION. Output parameter. Array of size m; contains the function values at x.

Declare fcn as EXTERNAL in the calling program.

In your codes:

If(djacobi(CalcErro,n,m,fjac,X,jac_eps)/=TR_SUCCESS) then

Where: Subroutine Calcerro(X,Erro,XY,n,m,numneu)

So, please change the interface or use the RCI version djacobi routine.

If it doesnt work, please let me know.

Thank you!

--Nikita

[renepreto]

Thank You very much! It's working really fine! I've made the fcn subroutine and followed the example programs too! But your help was really important! I'm really grateful

Rene