int write_unary_timeseries(time_series *ts_set, int n_series, char* filename){
int i, n, t;
int d;
int *record;
int n_observed;
nip_variable v;
nip_variable *observed;
nip_variable *observed_more;
time_series ts;
nip_model the_model;
FILE *f = NULL;
/* Check stuff */
if(!(n_series > 0 && ts_set && filename)){
nip_report_error(__FILE__, __LINE__, NIP_ERROR_INVALID_ARGUMENT, 1);
return NIP_ERROR_INVALID_ARGUMENT;
}
/* Check all the models are same */
the_model = ts_set[0]->model;
for(n = 1; n < n_series; n++){
if(ts_set[n]->model != the_model){
nip_report_error(__FILE__, __LINE__, NIP_ERROR_INVALID_ARGUMENT, 1);
return NIP_ERROR_INVALID_ARGUMENT;
}
}
/* Find out union of observed variables */
ts = ts_set[0];
observed = nip_variable_union(ts->observed, NULL,
the_model->num_of_vars - ts->num_of_hidden,
0, &n_observed);
if(n_observed < 0){
nip_report_error(__FILE__, __LINE__, NIP_ERROR_GENERAL, 1);
return NIP_ERROR_GENERAL;
}
for(n = 1; n < n_series; n++){
ts = ts_set[n];
observed_more = nip_variable_union(observed, ts->observed, n_observed,
the_model->num_of_vars -
ts->num_of_hidden,
&n_observed);
free(observed); /* nice to create the same array again and again? */
observed = observed_more;
}
if(n_observed < 1){ /* no observations in any time series? */
nip_report_error(__FILE__, __LINE__, NIP_ERROR_INVALID_ARGUMENT, 1);
return NIP_ERROR_INVALID_ARGUMENT;
}
/** NOTE: Actually this assumes there is only one observed variable! **/
/* Temporary space for a sorted record (unary values) */
v = observed[0];
record = (int*) calloc(NIP_CARDINALITY(v), sizeof(int));
if(!record){
nip_report_error(__FILE__, __LINE__, NIP_ERROR_OUTOFMEMORY, 1);
free(observed);
return NIP_ERROR_OUTOFMEMORY;
}
/* Try to open the file for write */
f = fopen(filename, "w");
if(!f){
nip_report_error(__FILE__, __LINE__, NIP_ERROR_IO, 1);
free(observed);
free(record);
return NIP_ERROR_IO;
}
/* Write names of the variable states */
for(i = 0; i < NIP_CARDINALITY(v); i++){
if(i > 0)
fprintf(f, "%c", NIP_FIELD_SEPARATOR);
fprintf(f, "%s", nip_variable_state_name(v, i));
}
fputs("\n", f);
/* Write the data */
for(n = 0; n < n_series; n++){ /* for each time series */
ts = ts_set[n];
for(t = 0; t < TIME_SERIES_LENGTH(ts); t++){ /* for each time step */
/* Fill record with indicators of missing data */
for(i = 0; i < NIP_CARDINALITY(v); i++)
record[i] = 0;
/* Extract data from the time series */
d = ts->data[t][0];
if(d >= 0 && d < NIP_CARDINALITY(v))
record[d] = 1;
/* Print the data */
for(i = 0; i < NIP_CARDINALITY(v); i++){
if(i > 0)
fprintf(f, "%c", NIP_FIELD_SEPARATOR);
if(record[i])
fputs("1", f);
else
fputs("0", f);
}
fputs("\n", f);
}
fputs("\n", f); /* TS separator */
}
free(observed);
free(record);
/* Close the file */
if(fclose(f)){
nip_report_error(__FILE__, __LINE__, NIP_ERROR_IO, 1);
return NIP_ERROR_IO;
}
return NIP_NO_ERROR;
}
int main(int argc, char *argv[]){
int i, j, k, n, t = 0;
double** quotient = NULL;
double*** result = NULL; /* probs of the hidden variables */
nip model = NULL;
nip_clique clique_of_interest = NULL;
nip_variable temp = NULL;
nip_variable interesting = NULL;
time_series *ts_set = NULL;
time_series ts = NULL;
/*************************************/
/* Some experimental timeslice stuff */
/*************************************/
/*****************************************/
/* Parse the model from a Hugin NET file */
/*****************************************/
/* -- Start parsing the network definition file */
if(argc < 3){
printf("Give the names of the net-file and data file, please!\n");
return 0;
}
else
model = parse_model(argv[1]);
if(model == NULL)
return -1;
/* The input file has been parsed. -- */
/*****************************/
/* read the data from a file */
/*****************************/
n = read_timeseries(model, argv[2], &ts_set); /* 1. Open */
if(n == 0){
free_model(model);
nip_report_error(__FILE__, __LINE__, NIP_ERROR_INVALID_ARGUMENT, 1);
fprintf(stderr, "%s\n", argv[2]);
return -1;
}
ts = ts_set[0];
/* Allocate some space for filtering */
assert(ts->num_of_hidden > 0);
result = (double***) calloc(TIME_SERIES_LENGTH(ts) + 1, sizeof(double**));
quotient = (double**) calloc(ts->num_of_hidden, sizeof(double*));
if(!(result && quotient)){
free_model(model);
nip_report_error(__FILE__, __LINE__, NIP_ERROR_OUTOFMEMORY, 1);
return 1;
}
for(t = 0; t < TIME_SERIES_LENGTH(ts) + 1; t++){
result[t] = (double**) calloc(ts->num_of_hidden, sizeof(double*));
if(!result[t]){
free_model(model);
nip_report_error(__FILE__, __LINE__, NIP_ERROR_OUTOFMEMORY, 1);
return 1;
}
for(i = 0; i < ts->num_of_hidden; i++){
result[t][i] = (double*) calloc(NIP_CARDINALITY(ts->hidden[i]),
sizeof(double));
if(!result[t][i]){
free_model(model);
nip_report_error(__FILE__, __LINE__, NIP_ERROR_OUTOFMEMORY, 1);
return 1;
}
}
}
for(i = 0; i < ts->num_of_hidden; i++){
quotient[i] = (double*) calloc(NIP_CARDINALITY(ts->hidden[i]),
sizeof(double));
if(!quotient[i]){
free_model(model);
nip_report_error(__FILE__, __LINE__, NIP_ERROR_OUTOFMEMORY, 1);
return 1;
}
}
/*****************/
/* Forward phase */
/*****************/
printf("## Forward phase ##\n");
reset_model(model); /* Reset the clique tree */
use_priors(model, !NIP_HAD_A_PREVIOUS_TIMESLICE);
for(t = 0; t < TIME_SERIES_LENGTH(ts); t++){ /* FOR EVERY TIMESLICE */
printf("-- t = %d --\n", t);
/********************/
/* Do the inference */
/********************/
make_consistent(model);
/* an experimental forward phase (a.k.a. filtering)... */
/* Calculates the result values */
for(i = 0; i < ts->num_of_hidden; i++){
/*********************************/
/* Check the result of inference */
/*********************************/
/* 1. Decide which variable you are interested in */
interesting = ts->hidden[i];
/* 2. Find the clique that contains the family of
* the interesting variable */
clique_of_interest = nip_find_family(model->cliques,
model->num_of_cliques,
interesting);
if(!clique_of_interest){
free_model(model);
free_timeseries(ts);
printf("In hmmtest.c : No clique found! Sorry.\n");
return 1;
}
/* 3. Marginalisation (memory for the result must have been allocated) */
nip_marginalise_clique(clique_of_interest, interesting, result[t][i]);
/* 4. Normalisation */
nip_normalise_array(result[t][i], NIP_CARDINALITY(interesting));
/* 5. Print the result */
for(j = 0; j < NIP_CARDINALITY(interesting); j++)
printf("P(%s=%s) = %f\n", nip_variable_symbol(interesting),
(interesting->state_names)[j], result[t][i][j]);
printf("\n");
}
if(t < TIME_SERIES_LENGTH(ts)){
/* forget old evidence */
reset_model(model);
use_priors(model, NIP_HAD_A_PREVIOUS_TIMESLICE);
for(i = 0; i < ts->num_of_hidden; i++){
/* old posteriors become new priors */
temp = ts->hidden[i];
if(temp->next != NULL)
nip_update_likelihood(temp->next, result[t][i]);
}
nip_global_retraction(model->variables, model->num_of_vars,
model->cliques, model->num_of_cliques);
/* 0. Put some data in */
for(i = 0; i < model->num_of_vars - ts->num_of_hidden; i++)
if(ts->data[t][i] >= 0)
nip_enter_index_observation(model->variables, model->num_of_vars,
model->cliques, model->num_of_cliques,
ts->observed[i],
ts->data[t][i]);
}
}
/******************/
/* Backward phase */
/******************/
printf("## Backward phase ##\n");
/* forget old evidence */
reset_model(model);
use_priors(model, NIP_HAD_A_PREVIOUS_TIMESLICE); /* JJT: Not sure... */
for(t = TIME_SERIES_LENGTH(ts)-1; t >= 0; t--){ /* FOR EVERY TIMESLICE */
printf("-- t = %d --\n", t);
if(t > 0){
for(i = 0; i < model->num_of_vars - ts->num_of_hidden; i++){
temp = ts->observed[i];
assert(temp);
if(ts->data[t - 1][i] >= 0)
nip_enter_index_observation(model->variables, model->num_of_vars,
model->cliques, model->num_of_cliques,
temp,
ts->data[t - 1][i]);
}
for(i = 0; i < ts->num_of_hidden; i++){
temp = ts->hidden[i];
if(temp->next != NULL)
nip_enter_evidence(model->variables, model->num_of_vars,
model->cliques, model->num_of_cliques,
temp->next, result[t-1][i]);
}
}
if(t < TIME_SERIES_LENGTH(ts)){
for(i = 0; i < ts->num_of_hidden; i++){
temp = ts->hidden[i];
if(temp->previous != NULL){
/* search for the other index */
for(k = 0; k < ts->num_of_hidden; k++)
if(nip_equal_variables(temp->previous, ts->hidden[k]))
break;
/* FIXME: Get rid of the quotient array */
printf("result[%d][%d][%d] / result[%d][%d][%d]\n",
t+1, i, j, t, k, j);
for(j = 0; j < NIP_CARDINALITY(temp); j++)
quotient[i][j] = result[t + 1][i][j] / result[t][k][j];
nip_enter_evidence(model->variables, model->num_of_vars,
model->cliques, model->num_of_cliques,
temp->previous, quotient[i]);
}
}
}
/********************/
/* Do the inference */
/********************/
make_consistent(model);
/*********************************/
/* Check the result of inference */
/*********************************/
for(i = 0; i < ts->num_of_hidden; i++){
/* 1. Decide which variable you are interested in */
interesting = ts->hidden[i];
/* 2. Find the clique that contains the family of
* the interesting variable */
clique_of_interest = nip_find_family(model->cliques,
model->num_of_cliques,
interesting);
if(!clique_of_interest){
free_model(model);
free_timeseries(ts);
printf("In hmmtest.c : No clique found! Sorry.\n");
return 1;
}
/* 3. Marginalisation (the memory must have been allocated) */
nip_marginalise_clique(clique_of_interest, interesting, result[t][i]);
/* 4. Normalisation */
nip_normalise_array(result[t][i], NIP_CARDINALITY(interesting));
/* 5. Print the result */
for(j = 0; j < NIP_CARDINALITY(interesting); j++)
printf("P(%s=%s) = %f\n", nip_variable_symbol(interesting),
(interesting->state_names)[j], result[t][i][j]);
printf("\n");
}
/* forget old evidence */
reset_model(model);
use_priors(model, NIP_HAD_A_PREVIOUS_TIMESLICE);
}
for(t = 0; t < TIME_SERIES_LENGTH(ts) + 1; t++){
for(i = 0; i < ts->num_of_hidden; i++)
free(result[t][i]);
free(result[t]);
}
for(i = 0; i < ts->num_of_hidden; i++)
free(quotient[i]);
free(result);
free(quotient);
for(i = 0; i < n; i++)
free_timeseries(ts_set[i]);
free(ts_set);
free_model(model);
return 0;
}
