[ADMB Users] string identifier error in ADMB-RE
John Sibert
sibert at hawaii.edu
Sat Jan 9 10:31:56 PST 2010
Hi Charlotte,
This seems like a really ugly "feature", one which I have never
encountered. Would you mind sharing some more information? What
operating system, compiler and architecture (32/64 bit) produced the
error. It would help us track down the problem, and at least produce a
more informative error message.
By the way, it is sometimes possible to use a funnel_dvariable to
reduce the amount of temporary data written to the disk, but it depends
on the structure of the model. Also, you can set the sizes of of memory
buffers to minimize wries to disk in the TOP_OF_MAIN_SECTION. In one
largish model the we run on a 64-bit linux system we use these settings
gradient_structure::set_GRADSTACK_BUFFER_SIZE( 89478485L);
gradient_structure::set_CMPDIF_BUFFER_SIZE( 400000000L);
gradient_structure gs( 20000000L);
Thanks,
John
Charlotte Boyd wrote:
> Hi Coilin
>
> That does seem to be it. Many thanks for the suggestion. I guess I should
> have been able to work that out myself, but the string identifier error
> message led me to think it was a coding issue so I've been looking in the
> wrong place.
>
> Many thanks for your help.
>
> Charlotte
>
>
>
>
> -----Original Message-----
> From: Coilin Minto [mailto:mintoc at mathstat.dal.ca]
> Sent: Friday, January 08, 2010 6:34 AM
> To: Charlotte Boyd
> Cc: users at admb-project.org
> Subject: Re: [ADMB Users] string identifier error in ADMB-RE
>
> Hi Charlotte,
>
> This might be a memory issue. Monitor the size of the temporary disk
> files created when ADMB-RE runs out of memory on one of your more
> complex models (section 3.12.1 of the ADMB-RE manual). If you are in
> GNU/Linux, start your model and open a new terminal, "cd" to the working
> directory and type:
>
> watch ls -alth
>
> This should refresh the file listing while your code is running so you
> can see which files are growing large. Useful commands to see what is
> going on with the memory are "top" and
>
> watch cat /proc/meminfo
>
> If you haven't already done so, you could allocate a greater amount of
> memory using the [-l1 -l2 -l3 -nl1 -nl2 -nl3] options when executing.
> There is also the option to split the job into N parts using -ndb N. The
> current ADMB-RE manual goes into greater detail. I would suggest trying
> these and, hopefully, other suggestions from more experienced users.
>
> Regards,
> Coilin
>
>
> Charlotte Boyd wrote:
>
>> Hi
>>
>> I've been working on adding random effects to a model that I developed
>> in ADMB. In ADMB-RE, I frequently (but not always) get the error message:
>>
>> 'Error in string identifer in list value should be UZ it appears to be
>> ...' during the Newton Raphson 2 phase.
>>
>> So far, I've failed to identify the cause of the error, and would
>> really appreciate some suggestions.
>>
>> The full model is a hidden Markov model (designed to partition animal
>> movement tracks into movement modes), solved with a forwards-backwards
>> algorithm.
>>
>> I started with a relatively simple version of the model (forwards
>> only, 12 estimated parameters, 6 random effects, standard deviation on
>> random effects fixed) and a simulated dataset. I can run this version
>> successfully on some datasets, but the error message occurs with
>> others. In general, more complex versions of the model (e.g. 16
>> estimated parameters) are more likely to produce the error, but
>> occasionally the more complex model runs to completion while the
>> simpler version fails. Models without the 'if' statement also tend to
>> fail at the same point with the same error message. (All datasets are
>> simulated according to the same parameters. Some have 8 trips with
>> 1,000 observations each, others 32 trips with 1,000 observations each.
>> The larger datasets are more likely to fail.) Sometimes a coding
>> change appears to resolve the problem, but it inevitably returns as I
>> free up more parameters.
>>
>> I've pasted the complete code below. This version works on some
>> datasets but not on others. I'm fairly new to ADMB, so I'm sure there
>> are many ways in which the code could be tightened up or made more
>> efficient, but resolving the string identifier error seems to be key.
>> If there is any additional information that would be useful, please
>> let me know.
>>
>> With many thanks in advance for your help.
>>
>> Charlotte Boyd
>>
>> ///School of Aquatic and Fishery Sciences/
>>
>> /University of Washington/
>>
>> /1122 NE Boat St, Rm 116/
>>
>> /Seattle, WA 98105///
>>
>> TOP_OF_MAIN_SECTION
>>
>> arrmblsize=20000000;
>>
>> DATA_SECTION
>>
>> init_int Nobs; // number of observations
>>
>> init_int Ntrips; // number of trips
>>
>> init_int Ntripobs; // number of observations per trip
>>
>> init_matrix Obs(1,Nobs,1,4);
>>
>> vector treatment(1,Nobs);
>>
>> vector trip(1,Nobs);
>>
>> vector distt(1,Nobs);
>>
>> vector theta(1,Nobs);
>>
>> !!treatment = column(Obs,1);
>>
>> !!trip = column(Obs,2);
>>
>> !!distt = column(Obs,3);
>>
>> !!theta = column(Obs,4);
>>
>> number pi;
>>
>> //stdevs on random effects - fixed in this version
>>
>> number sigma_scale2;
>>
>> number sigma_shape2;
>>
>> number sigma_rho2;
>>
>> number sigma_scale3;
>>
>> number sigma_shape3;
>>
>> number sigma_rho3;
>>
>> //transition parameters - fixed in this version
>>
>> number pswitch11;
>>
>> number pswitch12;
>>
>> number pswitch13;
>>
>> number pswitch21;
>>
>> number pswitch22;
>>
>> number pswitch23;
>>
>> number pswitch31;
>>
>> number pswitch32;
>>
>> number pswitch33;
>>
>> !! cout<<"done reading data"<<endl; // indicates data has been read in
>>
>> INITIALIZATION_SECTION
>>
>> //global parameters for weibull on distance and wrapped cauchy on
>> turning angle
>>
>> input_mode2_scale -2.3
>>
>> input_mode2_shape -0.22
>>
>> input_mode2_rho -0.32
>>
>> input_mode3_scale -0.22
>>
>> input_mode3_shape 1.61
>>
>> input_mode3_rho 2.4
>>
>> //treatment effects
>>
>> mode2_scale_treat 0.69
>>
>> mode2_shape_treat 0
>>
>> mode2_rho_treat 0.5
>>
>> mode3_scale_treat 0.41
>>
>> mode3_shape_treat 0.18
>>
>> mode3_rho_treat 0.69
>>
>> PARAMETER_SECTION
>>
>> init_number input_mode2_scale(1); //global parameters
>>
>> init_number input_mode2_shape(1);
>>
>> init_number input_mode2_rho(1);
>>
>> init_number input_mode3_scale(1);
>>
>> init_number input_mode3_shape(1);
>>
>> init_number input_mode3_rho(1);
>>
>> init_number mode2_scale_treat(1); //treatment effects
>>
>> init_number mode2_shape_treat(1);
>>
>> init_number mode2_rho_treat(1);
>>
>> init_number mode3_scale_treat(1);
>>
>> init_number mode3_shape_treat(1);
>>
>> init_number mode3_rho_treat(1);
>>
>> random_effects_matrix u(1,Ntrips,1,6,2); // matrix of unscaled random
>> effects: re=sigma_re*u where u~N(0,1)
>>
>> objective_function_value obj_fun;
>>
>> PROCEDURE_SECTION
>>
>> int t, obs; //counters for trip and all observations
>>
>> obs=0;
>>
>> obj_fun=0; //resets objective function
>>
>> // loop over trips
>>
>> for (t=1;t<=Ntrips;t++)
>>
>> fit_trip(input_mode2_scale, input_mode2_shape, input_mode2_rho,
>> input_mode3_scale, input_mode3_shape, input_mode3_rho,
>> mode2_scale_treat, mode2_shape_treat, mode2_rho_treat,
>> mode3_scale_treat, mode3_shape_treat, mode3_rho_treat, u(t), t, obs);
>>
>> SEPARABLE_FUNCTION void fit_trip(const dvariable& input_mode2_scale,
>> const dvariable& input_mode2_shape, const dvariable& input_mode2_rho,
>> const dvariable& input_mode3_scale, const dvariable&
>> input_mode3_shape, const dvariable& input_mode3_rho, const dvariable&
>> mode2_scale_treat, const dvariable& mode2_shape_treat, const
>> dvariable& mode2_rho_treat, const dvariable& mode3_scale_treat, const
>> dvariable& mode3_shape_treat, const dvariable& mode3_rho_treat, const
>> dvar_vector& ut, const int& t, int obs)
>>
>> obs = (t-1)*Ntripobs; //sets obs for beginning of each trip
>>
>> int i; // loop counter
>>
>> pi = 3.141593; //specifies pi
>>
>> sigma_scale2=0.01;
>>
>> sigma_shape2=0.01;
>>
>> sigma_rho2 = 0.01;
>>
>> sigma_scale3=0.01;
>>
>> sigma_shape3=0.1;
>>
>> sigma_rho3 = 0.001;
>>
>> dvar_vector priorfor_mode1(1,Nobs);
>>
>> dvar_vector priorfor_mode2(1,Nobs);
>>
>> dvar_vector priorfor_mode3(1,Nobs);
>>
>> dvar_vector likefor_mode1(1,Nobs);
>>
>> dvar_vector likefor_mode2(1,Nobs);
>>
>> dvar_vector likefor_mode3(1,Nobs);
>>
>> dvar_vector probfor_mode1(1,Nobs);
>>
>> dvar_vector probfor_mode2(1,Nobs);
>>
>> dvar_vector probfor_mode3(1,Nobs);
>>
>> obj_fun-=(-0.5*norm2(ut)); // likelihood contribution from unscaled
>> random effects based on the prior ~N(0,1)
>>
>> //scale & shape passed thro log filter, rho thro arctan filter
>>
>> dvariable
>>
>>
> mode2_scale=mfexp(input_mode2_scale+(treatment(obs+1)*mode2_scale_treat))+(s
> igma_scale2*ut(1));
>
>> dvariable
>>
>>
> mode2_shape=mfexp(input_mode2_shape+(treatment(obs+1)*mode2_shape_treat))+(s
> igma_shape2*ut(2));
>
>> dvariable
>>
>>
> mode2_rho=(0.5+atan(input_mode2_rho+(treatment(obs+1)*mode2_rho_treat))/pi)+
> (sigma_rho2*ut(3));
>
>> dvariable
>>
>>
> mode3_scale=mfexp(input_mode3_scale+(treatment(obs+1)*mode3_scale_treat))+(s
> igma_scale3*ut(4));
>
>> dvariable
>>
>>
> mode3_shape=mfexp(input_mode3_shape+(treatment(obs+1)*mode3_shape_treat))+(s
> igma_shape3*ut(5));
>
>> dvariable
>>
>>
> mode3_rho=(0.5+atan(input_mode3_rho+(treatment(obs+1)*mode3_rho_treat))/pi)+
> (sigma_rho3*ut(6));
>
>> pswitch11= 0.1 + (treatment(obs+1)*(-0.05));
>>
>> pswitch12= 0.6 + (treatment(obs+1)*0.1);
>>
>> pswitch13= 1 - pswitch11 - pswitch12;
>>
>> pswitch21= 0.03 + (treatment(obs+1)*0.02);
>>
>> pswitch22= 0.9 + (treatment(obs+1)*(-0.1));
>>
>> pswitch23= 1 - pswitch21 - pswitch22;
>>
>> pswitch31= 0.05 + (treatment(obs+1)*0);
>>
>> pswitch32= 0.1 + (treatment(obs+1)*(-0.05));
>>
>> pswitch33= 1 - pswitch31 - pswitch32;
>>
>> //first obs'n does not contribute to the likelihood, but contributes
>> to priors for second obs'n
>>
>> probfor_mode1(obs+1)=0;
>>
>> probfor_mode2(obs+1)=0.5;
>>
>> probfor_mode3(obs+1)=0.5;
>>
>> for (i=obs+2;i<=obs+Ntripobs-1;i++)
>>
>> {
>>
>> if (theta(i)==0)
>>
>> {
>>
>> probfor_mode1(i)=1; //obs'ns with incomplete movement data
>> automatically assigned to mode1; do not contribute to likelihood, but
>> contribute to priors for subsequent obs'ns
>>
>> probfor_mode2(i)=0;
>>
>> probfor_mode3(i)=0;
>>
>> } else
>>
>> {
>>
>> priorfor_mode2(i)=(probfor_mode1(i-1)*pswitch12) +
>> (probfor_mode2(i-1)*pswitch22) + (probfor_mode3(i-1)*pswitch32);
>>
>> priorfor_mode3(i)=(probfor_mode1(i-1)*pswitch12) +
>> (probfor_mode2(i-1)*pswitch23) + (probfor_mode3(i-1)*pswitch33);
>>
>> likefor_mode2(i)=(mode2_shape/mode2_scale)*(pow(distt(i)/mode2_scale,
>> mode2_shape-1)*mfexp(-1*pow(distt(i)/mode2_scale,mode2_shape))) *
>>
>>
> (1/(2*pi))*(1-pow(mode2_rho,2))/(1+pow(mode2_rho,2)-(2*mode2_rho*cos(theta(i
> ))))*priorfor_mode2(i);
>
>> // likelihood for each observation under mode 2
>>
>> likefor_mode3(i)=(mode3_shape/mode3_scale)*(pow(distt(i)/mode3_scale,
>> mode3_shape-1)*mfexp(-1*pow(distt(i)/mode3_scale,mode3_shape))) *
>>
>>
> (1/(2*pi))*(1-pow(mode3_rho,2))/(1+pow(mode3_rho,2)-(2*mode3_rho*cos(theta(i
> ))))*priorfor_mode3(i);
>
>> probfor_mode1(i)=0;
>>
>> probfor_mode2(i)=likefor_mode2(i)/(likefor_mode2(i) + likefor_mode3(i));
>>
>> probfor_mode3(i)=likefor_mode3(i)/(likefor_mode2(i) + likefor_mode3(i));
>>
>> obj_fun-= log(likefor_mode2(i) + likefor_mode3(i));
>>
>> }
>>
>> }
>>
>> /// // /
>>
>> ------------------------------------------------------------------------
>>
>> _______________________________________________
>> Users mailing list
>> Users at admb-project.org
>> http://lists.admb-project.org/mailman/listinfo/users
>>
>>
>
>
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users
>
>
--
John Sibert
Emeritus Researcher, SOEST
University of Hawaii at Manoa
Visit the ADMB project http://admb-project.org/
More information about the Users
mailing list