void _calculate_parameters(double h,my_point p[],double w[],int num) {
double H,I,J,K,L,A0, A1;
double x,y,d;
if (num > MAX_POINTS_NUM) {
fprintf(stderr,"Point number is larger than previous set!\n");
return;
}
is_set_ret = false;
compute_Aj(h,w,num);
H = compute_H(p,num);
I = compute_I(p,num);
J = compute_J(p,num);
K = compute_K(p,num);
L = compute_L(p,num);
A0 = H -h*h*J*J-K+h*h*L*L;
A1 = 2*(I-h*h*J*L);
// printf("H=%.3lf I=%.3lf J=%.3lf K=%.3lf L=%.3lf A0=%.3lf A1=%.3lf\n",
// H,I,J,K,L,A0,A1);
// printf("Calculated as follows:\n");
if (0 == A0) {
if (0 == A1) { // A0 A1 are0
// x,y could be any value
#if 1
printf("The distribution of the given points is a circle.\n");
x = y = sqrt(2.0) / 2;
d = -(h*h*(J*x+L*y));
compute_error(d,x,y,p,num);
#else
#endif
}
else { // A0 is 0 A1 is not 0，x2=1/2,x2+y2=1
double ar[2] = {sqrt(2.0)/2,-sqrt(2.0)/2};// possible values of x,y
int i,j;
for (i=0;i<2;i++) {
x = ar[i];
for (j=0;j<2;j++) {
y = ar[j];
d = -(h*h*(J*x+L*y));
compute_error(d,x,y,p,num);
}
}
}
}
else if (0 == A1) {
double x_ar[4] = {0,0,1,-1};
double y_ar[4] = {1,-1,0,0};//possible values of x,y
int i;
for (i=0;i<4;i++) {
x = x_ar[i];
y = y_ar[i];
d = -(h*h*(J*x+L*y));
compute_error(d,x,y,p,num);
}
}
else { // A0!=0 A1!=0
double t = A0 / sqrt (A1*A1+A0*A0); // 0 < t < 1
double x_ar[4] = {sqrt (0.5*(1+t)),sqrt (0.5*(1-t)),
-sqrt (0.5*(1+t)),-sqrt (0.5*(1-t))}; // possible values of x , x2 ≠ 0 or 1
int i;
for (i=0;i<4;i++) {
x = x_ar[i];
y = (A1/A0)* (x - 0.5/x);
d = -(h*h*(J*x+L*y));
compute_error(d,x,y,p,num);
}
}
}
void *LLW_train_thread(void *th_data) {
// Recover data
struct ThreadData *data = (struct ThreadData *)th_data;
const int thread_id = data->thread_id;
const int nprocs = data->nprocs;
struct Model *model = data->model;
struct KernelCache *kernelcache = data->kernelcache;
long chunk_size = data->chunk_size;
const double accuracy = data->accuracy;
double **gradient = data->gradient;
double **H_alpha = data->H_alpha;
double *best_primal_upper_bound = data->best_primal_upper_bound;
int *activeset = data->activeset;
long *nb_SV = data->nb_SV;
double *lp_rhs = data->lp_rhs;
FILE *fp = data->logfile_ptr;
pthread_mutex_unlock(&thread_data_mutex); // Release thread_data for next thread
// Local variables
int do_eval;
char yesno;
long long return_status = -1;
// Prepare the cache
struct TrainingCache cache;
cache.chunk_size = chunk_size;
LLW_alloc_memory(&cache, model->Q, model->nb_data, chunk_size);
cache.kc = kernelcache;
cache.activeset = activeset;
cache.lp_rhs = lp_rhs;
double **delta = matrix(chunk_size, model->Q);
double previous_ratio = 0.0;
double improvement = 1.0;
double theta_opt;
int jump = false;
if(accuracy == 0)
do_eval = 0;
else
do_eval = 1;
/*
Prepare parallel gradient computations:
- the gradient vector is split into NUMTHREADS_GRAD parts (along i)
- each part is updated by a different thread
*/
// max number of threads for gradient updates is nprocs
pthread_t *grad_threads = (pthread_t *)malloc(sizeof(pthread_t) * nprocs);
// start with 1 thread (main load on kernel evaluations)
int numthreads_grad = 1;
void *status;
int rc;
long k;
struct ThreadGradient_data *grad_data = (struct ThreadGradient_data *)malloc(sizeof(struct ThreadGradient_data) * nprocs);
// Disable parallel gradient computation for small data sets
int parallel_gradient_update = 1;
if(model->nb_data < 5000 || nprocs == 1)
parallel_gradient_update = 0;
if(parallel_gradient_update) {
for(k=0;k<nprocs;k++) {
grad_data[k].gradient = gradient;
grad_data[k].H_alpha = H_alpha;
grad_data[k].cache = &cache;
grad_data[k].model = model;
}
grad_data[0].start_i = 1;
grad_data[0].end_i = model->nb_data / numthreads_grad;
for(k=1;k<numthreads_grad-1;k++) {
grad_data[k].start_i = grad_data[k-1].end_i + 1;
grad_data[k].end_i = grad_data[k].start_i + model->nb_data / numthreads_grad -1;
}
if(numthreads_grad>1) {
grad_data[numthreads_grad-1].start_i = grad_data[numthreads_grad-2].end_i + 1;
grad_data[numthreads_grad-1].end_i = model->nb_data;
}
}
#ifdef _WIN32
// Init POOL
TP_WORK ** work;
if(parallel_gradient_update) {
work = malloc(sizeof(TP_WORK *) * nprocs);
for(k=0;k<nprocs;k++)
work[k] = CreateThreadpoolWork(LLW_update_gradient_thread2, (void *) &grad_data[k], NULL);
}
#endif
// Switch to nprocs/4 threads for gradient update when 25% of the kernel matrix is cached
int percentage_step = 1;
long percentage = model->nb_data / 4;
int next_numthreads_grad = nprocs/4;
if(next_numthreads_grad == 0)
next_numthreads_grad = 1;
// Main loop
int thread_stop = 0;
do {
if((TRAIN_SMALL_STEP < TRAIN_STEP) && (model->iter%TRAIN_SMALL_STEP) == 0) {
printf(".");
fflush(stdout);
}
// Select a random chunk of data to optimize
select_random_chunk(&cache,model);
// Compute the kernel submatrix for this chunk
compute_K(&cache,model);
// Enter Critical Section (using and modifying the model)
pthread_mutex_lock(&(model->mutex));
jump = LLW_solve_lp(gradient, &cache, model);
if(jump == false)
jump = LLW_check_opt_sol(gradient,&cache,model);
if(jump == false) {
LLW_compute_delta(delta,&cache,model);
theta_opt = LLW_compute_theta_opt(delta, &cache, model);
if (theta_opt > 0.0) {
*nb_SV += LLW_compute_new_alpha(theta_opt,&cache,model);
if(parallel_gradient_update) {
// Update gradient in parallel
for(k=0;k<numthreads_grad;k++) {
#ifdef _WIN32
SubmitThreadpoolWork(work[k]);
#else
rc = pthread_create(&grad_threads[k], NULL, LLW_update_gradient_thread, (void *) &grad_data[k]);
#endif
}
// Wait for gradient computations to terminate
for(k=0;k<numthreads_grad;k++) {
#ifdef _WIN32
WaitForThreadpoolWorkCallbacks(work[k], FALSE);
#else
rc = pthread_join(grad_threads[k],&status);
#endif
}
}
else {
// old-style non-threaded gradient update (for small data sets)
LLW_update_gradient(gradient,H_alpha, &cache,model);
}
}
}
if((do_eval && (model->iter%TRAIN_STEP) == 0) || EVAL || STOP || (do_eval && model->ratio >= accuracy) )
{
if(fp != NULL)
fprintf(fp,"%ld ",model->iter);
if(EVAL)
printf("\n\n*** Evaluating the model at iteration %ld...\n",model->iter);
// Evaluate how far we are in the optimization
// (prints more info if interrutped by user)
previous_ratio = model->ratio;
model->ratio = MSVM_eval(best_primal_upper_bound, gradient, H_alpha, NULL, model, EVAL, fp);
print_training_info(*nb_SV, model);
improvement = model->ratio - previous_ratio;
if(EVAL) // if interrupted by user (otherwise let the ratio decide if we go on training)
{
printf("\n *** Do you want to continue training ([y]/n)? ");
yesno = getchar();
if(yesno=='n') {
STOP = 1;
}
EVAL = 0; // reset interruption trigger
}
}
// Release kernel submatrix in cache
release_K(&cache);
// Check if a sufficient % of the kernel matrix is cached
if( parallel_gradient_update && cache.kc->max_idx >= percentage ) {
// and switch thread to compute gradient upates instead of kernel rows if it is
thread_stop = switch_thread(nprocs, &numthreads_grad, &next_numthreads_grad, &percentage, &percentage_step, grad_data, thread_id, model->nb_data);
// (threads are actually stopped to leave the CPUs
// to other threads that will compute gradient updates)
}
model->iter++;
// Release mutex: End of critical section
pthread_mutex_unlock(&(model->mutex));
} while(model->iter <= MSVM_TRAIN_MAXIT && (!do_eval || (model->ratio < accuracy && improvement != 0.0)) && !STOP && !thread_stop);
// Release mutex: End of critical section (see below)
pthread_mutex_unlock(&(model->mutex));
#ifdef _WIN32
if(parallel_gradient_update){
for(k=0;k<numthreads_grad;k++)
CloseThreadpoolWork(work[k]);
}
#endif
// compute return_status
if(do_eval && (model->ratio >= accuracy || improvement==0.0))
return_status = 0; // optimum reached or no more improvement.
// Free memory
LLW_free_memory(&cache);
free(delta[1]);free(delta);
free(grad_threads);
free(grad_data);
pthread_exit((void*)return_status);
}
bool extract_clips(const char *input_path,const char *cluster_path,const char *output_path,const char *index_out_path,int clip_size) {
DiskReadMda X(input_path);
DiskReadMda C(cluster_path);
if (X.totalSize()<=1) {
printf("Problem reading input file: %s\n",input_path);
return false;
}
if (C.totalSize()<=1) {
printf("Problem reading input file: %s\n",cluster_path);
return false;
}
int M=X.N1();
int T=clip_size;
int num_clips=C.N2();
int K=compute_K(C);
printf("K=%d\n",K);
Mda index_out;
index_out.allocate(1,K);
MDAIO_HEADER H_out;
H_out.data_type=MDAIO_TYPE_FLOAT32;
H_out.num_bytes_per_entry=4;
H_out.num_dims=3;
H_out.dims[0]=M;
H_out.dims[1]=T;
H_out.dims[2]=num_clips;
FILE *outf=fopen(output_path,"wb");
if (!outf) {
printf("Unable to open output file: %s\n",output_path);
return false;
}
mda_write_header(&H_out,outf);
float *buf=(float *)malloc(sizeof(float)*M*T);
int jj=0;
for (int k=1; k<=K; k++) {
index_out.setValue(jj,0,k-1);
for (int i=0; i<num_clips; i++) {
int ii=0;
int time0=(int)C.value(1,i);
int k0=(int)C.value(2,i);
if (k0==k) {
for (int t=0; t<T; t++) {
for (int m=0; m<M; m++) {
buf[ii]=X.value(m,t+time0-T/2);
ii++;
}
}
mda_write_float32(buf,&H_out,M*T,outf);
jj++;
}
}
}
free(buf);
fclose(outf);
if (!index_out.write(index_out_path)) {
printf("Unable to write output file: %s\n",index_out_path);
return false;
}
return true;
}