int _svm_struct_learn (int argc, char* argv[])
{
SAMPLE sample; /* training sample */
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
STRUCT_LEARN_PARM struct_parm;
STRUCTMODEL structmodel;
int alg_type;
svm_struct_learn_api_init(argc,argv);
_svm_struct_learn_read_input_parameters(argc,argv,trainfile,modelfile,&verbosity,
&struct_verbosity,&struct_parm,&learn_parm,
&kernel_parm,&alg_type);
if(struct_verbosity>=1) {
printf("Reading training examples..."); fflush(stdout);
}
/* read the training examples */
sample=read_struct_examples(trainfile,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n"); fflush(stdout);
}
/* Do the learning and return structmodel. */
if(alg_type == 0)
svm_learn_struct(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,NSLACK_ALG);
else if(alg_type == 1)
svm_learn_struct(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,NSLACK_SHRINK_ALG);
else if(alg_type == 2)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,ONESLACK_PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,ONESLACK_DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,ONESLACK_DUAL_CACHE_ALG);
else if(alg_type == 9)
svm_learn_struct_joint_custom(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel);
else
exit(1);
/* Warning: The model contains references to the original data 'docs'.
If you want to free the original data, and only keep the model, you
have to make a deep copy of 'model'. */
if(struct_verbosity>=1) {
printf("Writing learned model...");fflush(stdout);
}
write_struct_model(modelfile,&structmodel,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n");fflush(stdout);
}
free_struct_sample(sample);
free_struct_model(structmodel);
svm_struct_learn_api_exit();
return 0;
}
int _svm_struct_learn (int argc, char* argv[])
{
char trainfile[200]; /* file with training examples */
char modelfile[200]; /* file for resulting classifier */
SAMPLE sample; /* training sample */
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
STRUCT_LEARN_PARM struct_parm;
STRUCTMODEL structmodel;
int alg_type;
HIDEO_ENV *hideo_env=create_env();
svm_struct_learn_api_init(argc,argv);
svm_struct_main_read_input_parameters(argc,argv,trainfile,modelfile,&verbosity,
&struct_verbosity,&struct_parm,&learn_parm,
&kernel_parm,&alg_type);
if(struct_verbosity>=1) {
printf("Reading training examples..."); fflush(stdout);
}
/* read the training examples */
sample=read_struct_examples(trainfile,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n"); fflush(stdout);
}
/* Do the learning and return structmodel. */
if(alg_type == 1)
svm_learn_struct(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,hideo_env);
else if(alg_type == 2)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,PRIMAL_ALG,hideo_env);
else if(alg_type == 3)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,DUAL_ALG,hideo_env);
else if(alg_type == 4)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,DUAL_CACHE_ALG,hideo_env);
else
exit(1);
/* Warning: The model contains references to the original data 'docs'.
If you want to free the original data, and only keep the model, you
have to make a deep copy of 'model'. */
if(struct_verbosity>=1) {
printf("Writing learned model...");fflush(stdout);
}
write_struct_model(modelfile,&structmodel,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n");fflush(stdout);
}
free_struct_sample(sample);
free_struct_model(structmodel);
svm_struct_learn_api_exit();
return 0;
}
int main (int argc, char* argv[])
{
SAMPLE sample; /* training sample */
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
STRUCT_LEARN_PARM struct_parm;
STRUCTMODEL structmodel;
int alg_type;
/* Allow the API to perform whatever initialization is required. */
api_initialize(argv[0]);
read_input_parameters(argc,argv,trainfile,modelfile,&verbosity,
&struct_verbosity,&struct_parm,&learn_parm,
&kernel_parm,&alg_type);
if(struct_verbosity>=1) {
printf("Reading training examples..."); fflush(stdout);
}
/* read the training examples */
sample=read_struct_examples(trainfile,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n"); fflush(stdout);
}
/* Do the learning and return structmodel. */
if(alg_type == 1)
svm_learn_struct(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,DUAL_CACHE_ALG);
else
exit(1);
/* Warning: The model contains references to the original data 'docs'.
If you want to free the original data, and only keep the model, you
have to make a deep copy of 'model'. */
if(struct_verbosity>=1) {
printf("Writing learned model...");fflush(stdout);
}
write_struct_model(modelfile,&structmodel,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n");fflush(stdout);
}
free_struct_sample(sample);
free_struct_model(structmodel);
/* Allow the API to perform whatever cleanup is required. */
api_finalize();
return 0;
}
int main (int argc, char* argv[])
{
SAMPLE sample; /* training sample */
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
STRUCT_LEARN_PARM struct_parm;
STRUCTMODEL structmodel;
int alg_type;
svm_struct_learn_api_init(argc,argv);
read_input_parameters(argc,argv,trainfile,modelfile,&verbosity,
&struct_verbosity,&struct_parm,&learn_parm,
&kernel_parm,&alg_type);
if(struct_verbosity>=1) {
//verbose = true;
printf("Reading training examples..."); fflush(stdout);
}
/* read the training examples */
sample=read_struct_examples(trainfile,&struct_parm);
if(struct_verbosity>=1) {
printf("done\n"); fflush(stdout);
}
string config_tmp;
bool update_loss_function = false;
if(Config::Instance()->getParameter("update_loss_function", config_tmp)) {
update_loss_function = config_tmp.c_str()[0] == '1';
}
printf("[Main] update_loss_function=%d\n", (int)update_loss_function);
if(!update_loss_function) {
printf("update_loss_function should be true\n");
exit(-1);
}
eUpdateType updateType = UPDATE_NODE_EDGE;
if(Config::Instance()->getParameter("update_type", config_tmp)) {
updateType = (eUpdateType)atoi(config_tmp.c_str());
}
printf("[Main] update_type=%d\n", (int)updateType);
mkdir(loss_dir, 0777);
// check if parameter vector files exist
const char* parameter_vector_dir = "parameter_vector";
int idx = 0;
string parameter_vector_dir_last = findLastFile(parameter_vector_dir, "", &idx);
string parameter_vector_file_pattern = parameter_vector_dir_last + "/iteration_";
int idx_1 = 1;
string parameter_vector_file_last = findLastFile(parameter_vector_file_pattern, ".txt", &idx_1);
printf("[Main] Checking parameter vector file %s\n", parameter_vector_file_last.c_str());
// vectors used to store RF weights
vector<double>* alphas = new vector<double>[sample.n];
vector<double>* alphas_edge = 0;
if(updateType == UPDATE_NODE_EDGE) {
alphas_edge = new vector<double>[sample.n];
}
int alphas_idx = 0;
if(fileExists("alphas.txt") && fileExists(parameter_vector_file_last)) {
// Loading alpha coefficients
ifstream ifs("alphas.txt");
string line;
int lineIdx = 0;
while(lineIdx < sample.n && getline(ifs, line)) {
vector<string> tokens;
splitString(line, tokens);
for(vector<string>::iterator it = tokens.begin();
it != tokens.end(); ++it) {
alphas[lineIdx].push_back(atoi(it->c_str()));
}
++lineIdx;
}
ifs.close();
if(lineIdx > 0) {
alphas_idx = alphas[0].size();
}
// Loading parameters
printf("[Main] Found parameter vector file %s\n", parameter_vector_file_last.c_str());
struct_parm.ssvm_iteration = idx + 1;
update_output_dir(struct_parm.ssvm_iteration);
//EnergyParam param(parameter_vector_file_last.c_str());
//updateCoeffs(sample, param, struct_parm, updateType, alphas, alphas_idx);
//alphas_idx = 1;
} else {
struct_parm.ssvm_iteration = 0;
// insert alpha coefficients for first iteration
for(int i = 0; i < sample.n; ++i) {
alphas[i].push_back(1);
}
ofstream ofs("alphas.txt", ios::app);
int i = 0;
for(; i < sample.n - 1; ++i) {
ofs << alphas[i][alphas_idx] << " ";
}
if(i < sample.n) {
ofs << alphas[i][alphas_idx];
}
ofs << endl;
ofs.close();
// edges
for(int i = 0; i < sample.n; ++i) {
alphas_edge[i].push_back(1);
}
ofstream ofse("alphas_edge.txt", ios::app);
i = 0;
for(; i < sample.n - 1; ++i) {
ofse << alphas_edge[i][alphas_idx] << " ";
}
if(i < sample.n) {
ofse << alphas_edge[i][alphas_idx];
}
ofse << endl;
ofse.close();
++alphas_idx;
}
const int nMaxIterations = 5;
bool bIterate = true;
do {
printf("-----------------------------------------SSVM-ITERATION-%d-START\n",
struct_parm.ssvm_iteration);
struct_parm.iterationId = 1;
/* Do the learning and return structmodel. */
if(alg_type == 0)
svm_learn_struct(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,NSLACK_ALG);
else if(alg_type == 1)
svm_learn_struct(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,NSLACK_SHRINK_ALG);
else if(alg_type == 2)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,ONESLACK_PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,ONESLACK_DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel,ONESLACK_DUAL_CACHE_ALG);
else if(alg_type == 9)
svm_learn_struct_joint_custom(sample,&struct_parm,&learn_parm,&kernel_parm,&structmodel);
else
exit(1);
char _modelfile[BUFFER_SIZE];
//sprintf(_modelfile, "%s_%d", modelfile, struct_parm.ssvm_iteration);
sprintf(_modelfile, "%s_%d", modelfile, struct_parm.ssvm_iteration);
printf("[Main] Writing learned model to %s\n", _modelfile);
write_struct_model(_modelfile, &structmodel, &struct_parm);
// Run inference on training data and increase loss for misclassified points
printf("[Main] Loading learned model to %s\n", _modelfile);
EnergyParam param(_modelfile);
updateCoeffs(sample, param, struct_parm, updateType, alphas, alphas_edge,
struct_parm.ssvm_iteration + 1);
ofstream ofs("alphas.txt", ios::app);
int i = 0;
for(; i < sample.n - 1; ++i) {
ofs << alphas[i][alphas_idx] << " ";
}
if(i < sample.n) {
ofs << alphas[i][alphas_idx];
}
ofs << endl;
ofs.close();
ofstream ofse("alphas_edge.txt", ios::app);
i = 0;
for(; i < sample.n - 1; ++i) {
ofse << alphas_edge[i][alphas_idx] << " ";
}
if(i < sample.n) {
ofse << alphas_edge[i][alphas_idx];
}
ofse << endl;
ofse.close();
++alphas_idx;
printf("-----------------------------------------SSVM-ITERATION-%d-END\n",
struct_parm.ssvm_iteration);
++struct_parm.ssvm_iteration;
bIterate = (nMaxIterations == -1 || struct_parm.ssvm_iteration < nMaxIterations);
} while(bIterate);
// Output final segmentation for all examples
long nExamples = sample.n;
int nRFs = struct_parm.ssvm_iteration;
double* lossPerLabel = 0;
labelType* groundTruthLabels = 0;
for(int i = 0; i < nExamples; i++) { /*** example loop ***/
Slice_P* slice = sample.examples[i].x.slice;
Feature* feature = sample.examples[i].x.feature;
//map<sidType, nodeCoeffType>* nodeCoeffs = sample.examples[i].x.nodeCoeffs;
//map<sidType, edgeCoeffType>* edgeCoeffs = sample.examples[i].x.edgeCoeffs;
map<sidType, nodeCoeffType>* nodeCoeffs = 0;
map<sidType, edgeCoeffType>* edgeCoeffs = 0;
int nNodes = slice->getNbSupernodes();
stringstream soutPb;
soutPb << loss_dir;
soutPb << "pb_";
soutPb << getNameFromPathWithoutExtension(slice->getName());
soutPb << "_";
soutPb << "combined";
//soutPb << setw(5) << setfill('0') << ssvm_iteration;
soutPb << ".tif";
printf("[Main] Exporting %s\n", soutPb.str().c_str());
labelType* inferredLabels =
computeCombinedLabels(slice, feature, groundTruthLabels,
lossPerLabel, nRFs, alphas, i,
nodeCoeffs, edgeCoeffs, soutPb.str().c_str());
stringstream sout;
sout << loss_dir;
sout << getNameFromPathWithoutExtension(slice->getName());
sout << "_";
sout << "combined";
//sout << setw(5) << setfill('0') << ssvm_iteration;
sout << ".tif";
printf("[Main] Exporting %s\n", sout.str().c_str());
slice->exportOverlay(sout.str().c_str(), inferredLabels);
stringstream soutBW;
soutBW << loss_dir;
soutBW << getNameFromPathWithoutExtension(slice->getName());
soutBW << "_";
soutBW << "combined";
//soutBW << setw(5) << setfill('0') << ssvm_iteration;
soutBW << "_BW.tif";
printf("[Main] Exporting %s\n", soutBW.str().c_str());
slice->exportSupernodeLabels(soutBW.str().c_str(),
struct_parm.nClasses,
inferredLabels, nNodes,
&(struct_parm.labelToClassIdx));
delete[] inferredLabels;
}
free_struct_sample(sample);
free_struct_model(structmodel);
svm_struct_learn_api_exit();
return 0;
}
int main(int argc, char* argv[]) {
double *w; /* weight vector */
long m, i;
double C, epsilon;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
char trainfile[1024];
char modelfile[1024];
int MAX_ITER;
/* new struct variables */
SVECTOR **fycache, *diff, *fy;
EXAMPLE *ex;
SAMPLE alldata;
SAMPLE sample;
SAMPLE val;
STRUCT_LEARN_PARM sparm;
STRUCTMODEL sm;
double primal_obj;
double stop_crit;
char itermodelfile[2000];
/* self-paced learning variables */
double init_spl_weight;
double spl_weight;
double spl_factor;
int *valid_examples;
/* read input parameters */
my_read_input_parameters(argc, argv, trainfile, modelfile, &learn_parm, &kernel_parm, &sparm,
&init_spl_weight, &spl_factor);
epsilon = learn_parm.eps;
C = learn_parm.svm_c;
MAX_ITER = learn_parm.maxiter;
/* read in examples */
alldata = read_struct_examples(trainfile,&sparm);
int ntrain = (int) round(1.0*alldata.n); /* no validation set */
if(ntrain < alldata.n)
{
long *perm = randperm(alldata.n);
sample = generate_train_set(alldata, perm, ntrain);
val = generate_validation_set(alldata, perm, ntrain);
free(perm);
}
else
{
sample = alldata;
}
ex = sample.examples;
m = sample.n;
/* initialization */
init_struct_model(alldata,&sm,&sparm,&learn_parm,&kernel_parm);
w = create_nvector(sm.sizePsi);
clear_nvector(w, sm.sizePsi);
sm.w = w; /* establish link to w, as long as w does not change pointer */
/* some training information */
printf("C: %.8g\n", C);
printf("spl weight: %.8g\n",init_spl_weight);
printf("epsilon: %.8g\n", epsilon);
printf("sample.n: %d\n", sample.n);
printf("sm.sizePsi: %ld\n", sm.sizePsi); fflush(stdout);
/* prepare feature vector cache for correct labels with imputed latent variables */
fycache = (SVECTOR**)malloc(m*sizeof(SVECTOR*));
for (i=0;i<m;i++) {
fy = psi(ex[i].x, ex[i].y, &sm, &sparm);
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
/* learn initial weight vector using all training examples */
valid_examples = (int *) malloc(m*sizeof(int));
/* errors for validation set */
double cur_loss, best_loss = DBL_MAX;
int loss_iter;
/* initializations */
spl_weight = init_spl_weight;
/* solve biconvex self-paced learning problem */
primal_obj = alternate_convex_search(w, m, MAX_ITER, C, epsilon, fycache, ex, &sm, &sparm, valid_examples, spl_weight);
printf("primal objective: %.4f\n", primal_obj);
fflush(stdout);
//alternate_convex_search(w, m, MAX_ITER, C, epsilon, fycache, ex, &sm, &sparm, valid_examples, spl_weight);
int nValid = 0;
for (i=0;i<m;i++) {
if(valid_examples[i]) {
nValid++;
}
}
if(ntrain < alldata.n) {
cur_loss = compute_current_loss(val,&sm,&sparm);
printf("CURRENT LOSS: %f\n",cur_loss);
}
/* write structural model */
write_struct_model(modelfile, &sm, &sparm);
// skip testing for the moment
/* free memory */
free_struct_sample(alldata);
if(ntrain < alldata.n)
{
free(sample.examples);
free(val.examples);
}
free_struct_model(sm, &sparm);
for(i=0;i<m;i++) {
free_svector(fycache[i]);
}
free(fycache);
free(valid_examples);
return(0);
}
int main (int argc, char* argv[])
{
int o, i, j, cad_num;
char filename[256];
CAD **cads, *cad;
SAMPLE sample, sample_part; /* training sample */
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
STRUCT_LEARN_PARM struct_parm;
STRUCTMODEL structmodel;
int alg_type;
int rank;
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
/* select GPU */
select_gpu(rank);
svm_struct_learn_api_init(argc,argv);
read_input_parameters(argc, argv, trainfile, cadfile, testfile, &verbosity,
&struct_verbosity, &struct_parm, &learn_parm,
&kernel_parm, &alg_type);
/* read cad models */
cads = read_cad_model(cadfile, &cad_num, 0, &struct_parm);
/* if cad_num == 1, train the final model */
if(cad_num == 1)
{
printf("Train the hierarchical model\n");
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = cads;
printf("Read training samples\n");
sample = read_struct_examples(trainfile, &struct_parm, &structmodel);
printf("Read training samples done\n");
if(struct_parm.is_root == 1 || struct_parm.is_aspectlet == 1)
{
/* first train weights for root parts */
printf("Train root models\n");
cad = cads[0];
struct_parm.cad_index = 0;
/* for each part */
for(i = 0; i < cad->part_num; i++)
{
/* choose what parts to train */
if((cad->roots[i] == -1 && struct_parm.is_root == 1) || (cad->roots[i] == 1 && struct_parm.is_aspectlet == 1))
{
printf("Train part %d\n", i);
/* training iteration index */
struct_parm.iter = 0;
struct_parm.part_index = i;
/* select training samples for part */
if(rank == 0)
select_examples_part(trainfile, sample, cad, 0, i);
MPI_Barrier(MPI_COMM_WORLD);
sample_part = read_struct_examples("temp_part.dat", &struct_parm, &structmodel);
/* train the root template */
struct_parm.deep = 0;
/* Do the learning and return structmodel. */
if(alg_type == 1)
svm_learn_struct(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
/* if aspectlet, train the subtree */
if(cad->roots[i] == 1 && struct_parm.is_aspectlet == 1)
{
printf("Train subtree for part %d\n", i);
struct_parm.deep = 1;
struct_parm.iter++;
/* select training samples for part */
if(rank == 0)
select_examples_part(trainfile, sample, cad, 0, i);
MPI_Barrier(MPI_COMM_WORLD);
free_struct_sample(sample_part);
sample_part = read_struct_examples("temp_part.dat", &struct_parm, &structmodel);
free_struct_model(structmodel);
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = cads;
if(alg_type == 1)
svm_learn_struct(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
}
/* data mining hard examples */
for(j = 0; j < struct_parm.hard_negative; j++)
{
/* increase iteration number */
struct_parm.iter++;
data_mining_hard_examples("temp_part.dat", testfile, &struct_parm, &structmodel);
/* read the training examples */
if(struct_verbosity>=1)
{
printf("Reading training examples...");
fflush(stdout);
}
free_struct_sample(sample_part);
sample_part = read_struct_examples("temp.dat", &struct_parm, &structmodel);
if(struct_verbosity>=1)
{
printf("done\n");
fflush(stdout);
}
/* Do the learning and return structmodel. */
free_struct_model(structmodel);
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = cads;
if(alg_type == 1)
svm_learn_struct(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
}
/* save constraints for training the full model */
if(rank == 0)
save_constraints(&struct_parm, &structmodel);
MPI_Barrier(MPI_COMM_WORLD);
free_struct_sample(sample_part);
free_struct_model(structmodel);
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = cads;
}
} /* end for each part */
if(rank == 0)
{
write_constraints(struct_parm.cset, &struct_parm);
free(struct_parm.cset.rhs);
for(i = 0; i < struct_parm.cset.m; i++)
free_example(struct_parm.cset.lhs[i], 1);
free(struct_parm.cset.lhs);
}
MPI_Barrier(MPI_COMM_WORLD);
} /* end if is_root == 1 */
/* train the full model */
printf("Train the full model\n");
struct_parm.iter = 0;
struct_parm.cad_index = -1;
struct_parm.part_index = -1;
struct_parm.deep = 1;
/* Do the learning and return structmodel. */
if(alg_type == 1)
svm_learn_struct(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
/* data mining hard examples */
while(struct_parm.hard_negative > 0)
{
/* increase iteration number */
struct_parm.iter++;
data_mining_hard_examples(trainfile, testfile, &struct_parm, &structmodel);
/* read the training examples */
if(struct_verbosity>=1)
{
printf("Reading training examples...");
fflush(stdout);
}
free_struct_sample(sample);
sample = read_struct_examples("temp.dat", &struct_parm, &structmodel);
if(struct_verbosity>=1)
{
printf("done\n");
fflush(stdout);
}
/* Do the learning and return structmodel. */
free_struct_model(structmodel);
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = cads;
if(alg_type == 1)
svm_learn_struct(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
struct_parm.hard_negative--;
}
if(rank == 0)
{
if(struct_verbosity >= 1)
{
printf("Writing learned model...");
fflush(stdout);
}
sprintf(modelfile, "%s.mod", struct_parm.cls);
write_struct_model(modelfile, &structmodel, &struct_parm);
if(struct_verbosity>=1)
{
printf("done\n");
fflush(stdout);
}
}
MPI_Barrier(MPI_COMM_WORLD);
free_struct_sample(sample);
free_struct_model(structmodel);
}
/* train weights for aspectlets */
/* start with the second aspectlet, the first one is the whole object */
for(o = 1; o < cad_num; o++)
{
printf("Train aspectlet %d\n", o);
cad = cads[o];
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = &cad;
/* training iteration index */
struct_parm.iter = 0;
struct_parm.cad_index = -1;
struct_parm.part_index = -1;
struct_parm.deep = 1;
/* select training samples for the aspectlet */
if(rank == 0)
select_examples_aspectlet(trainfile, cads, o);
MPI_Barrier(MPI_COMM_WORLD);
printf("Read training samples\n");
sample_part = read_struct_examples("temp_part.dat", &struct_parm, &structmodel);
printf("Read training samples done\n");
/* Do the learning and return structmodel. */
if(alg_type == 1)
svm_learn_struct(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
/* data mining hard examples */
for(i = 0; i < struct_parm.hard_negative; i++)
{
/* increase iteration number */
struct_parm.iter++;
data_mining_hard_examples("temp_part.dat", testfile, &struct_parm, &structmodel);
/* read the training examples */
if(struct_verbosity>=1)
{
printf("Reading training examples...");
fflush(stdout);
}
free_struct_sample(sample_part);
sample_part = read_struct_examples("temp.dat", &struct_parm, &structmodel);
if(struct_verbosity>=1)
{
printf("done\n");
fflush(stdout);
}
/* Do the learning and return structmodel. */
free_struct_model(structmodel);
/* set the cad model for structmodel */
structmodel.cad_num = 1;
structmodel.cads = &cad;
if(alg_type == 1)
svm_learn_struct(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel);
else if(alg_type == 2)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, PRIMAL_ALG);
else if(alg_type == 3)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_ALG);
else if(alg_type == 4)
svm_learn_struct_joint(sample_part, &struct_parm, &learn_parm, &kernel_parm, &structmodel, DUAL_CACHE_ALG);
else
exit(1);
}
if(rank == 0)
{
if(struct_verbosity>=1)
{
printf("Writing learned model...");
fflush(stdout);
}
sprintf(modelfile, "%s_cad%03d.mod", struct_parm.cls, o);
write_struct_model(modelfile, &structmodel, &struct_parm);
if(struct_verbosity>=1)
{
printf("done\n");
fflush(stdout);
}
}
MPI_Barrier(MPI_COMM_WORLD);
free_struct_sample(sample_part);
free_struct_model(structmodel);
} /* end for each cad model */
for(i = 0; i < cad_num; i++)
destroy_cad(cads[i]);
free(cads);
svm_struct_learn_api_exit();
MPI_Finalize();
return 0;
}
int main(int argc, char* argv[]) {
double *w; /* weight vector */
int outer_iter;
long m, i;
double C, epsilon;
LEARN_PARM learn_parm;
KERNEL_PARM kernel_parm;
char trainfile[1024];
char modelfile[1024];
int MAX_ITER;
/* new struct variables */
SVECTOR **fycache, *diff, *fy;
EXAMPLE *ex;
SAMPLE sample;
STRUCT_LEARN_PARM sparm;
STRUCTMODEL sm;
//double decrement;
double primal_obj;//, last_primal_obj;
//double cooling_eps;
//double stop_crit;
DebugConfiguration::VerbosityLevel = VerbosityLevel::None;
/* read input parameters */
my_read_input_parameters(argc, argv, trainfile, modelfile, &learn_parm, &kernel_parm, &sparm);
epsilon = learn_parm.eps;
C = learn_parm.svm_c;
MAX_ITER = learn_parm.maxiter;
/* read in examples */
sample = read_struct_examples(trainfile,&sparm);
ex = sample.examples;
m = sample.n;
/* initialization */
init_struct_model(sample,&sm,&sparm,&learn_parm,&kernel_parm);
w = sm.w;
//w = create_nvector(sm.sizePsi);
//clear_nvector(w, sm.sizePsi);
//sm.w = w; /* establish link to w, as long as w does not change pointer */
/* some training information */
printf("C: %.8g\n", C);
printf("epsilon: %.8g\n", epsilon);
printf("sample.n: %ld\n", sample.n);
printf("sm.sizePsi: %ld\n", sm.sizePsi);
fflush(stdout);
/* impute latent variable for first iteration */
init_latent_variables(&sample,&learn_parm,&sm,&sparm);
/* prepare feature vector cache for correct labels with imputed latent variables */
fycache = (SVECTOR**)malloc(m*sizeof(SVECTOR*));
for (i=0; i<m; i++) {
fy = psi(ex[i].x, ex[i].y, ex[i].h, &sm, &sparm);
/* DEBUG */
printf("true_psi[%d]=", i);
for (int j = 0; j < sm.sizePsi; ++j)
printf("%.4lf ", fy->words[j].weight);
printf("\n");
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
/* outer loop: latent variable imputation */
outer_iter = 1;
//last_primal_obj = 0;
//decrement = 0;
//cooling_eps = 0.5*C*epsilon;
//while ((outer_iter<=MIN_OUTER_ITER)||((!stop_crit)&&(outer_iter<MAX_OUTER_ITER))) {
while (outer_iter<MAX_OUTER_ITER) {
LearningTracker::NextOuterIteration();
printf("OUTER ITER %d\n", outer_iter);
/* cutting plane algorithm */
primal_obj = cutting_plane_algorithm(w, m, MAX_ITER, C, /*cooling_eps, */fycache, ex, &sm, &sparm);
/* compute decrement in objective in this outer iteration */
/*
decrement = last_primal_obj - primal_obj;
last_primal_obj = primal_obj;
printf("primal objective: %.4f\n", primal_obj);
printf("decrement: %.4f\n", decrement); fflush(stdout);
stop_crit = (decrement<C*epsilon)&&(cooling_eps<0.5*C*epsilon+1E-8);
cooling_eps = -decrement*0.01;
cooling_eps = MAX(cooling_eps, 0.5*C*epsilon);
printf("cooling_eps: %.8g\n", cooling_eps); */
/* print new weights */
printf("W=");
for (i = 1; i <= sm.sizePsi; ++i)
printf("%.3f ", sm.w[i]);
printf("\n");
/* Save model */
char modelfile_tmp[1024];
sprintf(modelfile_tmp, "%s.%d", modelfile, outer_iter);
write_struct_model(modelfile_tmp, &sm, &sparm);
/* impute latent variable using updated weight vector */
for (i=0; i<m; i++) {
free_latent_var(ex[i].h);
ex[i].h = infer_latent_variables(ex[i].x, ex[i].y, &sm, &sparm);
}
/* re-compute feature vector cache */
for (i=0; i<m; i++) {
free_svector(fycache[i]);
fy = psi(ex[i].x, ex[i].y, ex[i].h, &sm, &sparm);
/* DEBUG */
printf("true_psi[%d]=", i);
for (int j = 0; j < sm.sizePsi; ++j)
printf("%.4lf ", fy->words[j].weight);
printf("\n");
diff = add_list_ss(fy);
free_svector(fy);
fy = diff;
fycache[i] = fy;
}
outer_iter++;
} // end outer loop
/* write structural model */
write_struct_model(modelfile, &sm, &sparm);
// skip testing for the moment
/* free memory */
free_struct_sample(sample);
free_struct_model(sm, &sparm);
for(i=0; i<m; i++) {
free_svector(fycache[i]);
}
free(fycache);
return(0);
}
