int main(int argc, char **argv)
{
int n_gaussians;
int n_states;
bool is_symbol;
char* file_states;
char* model_name1=NULL;
char* model_name2=NULL;
char* model_name3=NULL;
char* spacing_model;
char* lex_name1=NULL;
char* lex_name2=NULL;
char* lex_name3=NULL;
char *sent_start_symbol=NULL;
char *sent_end_symbol=NULL;
char *spacing_symbol=NULL;
char *model_file_in;
char *model_file_out;
char *dir_name;
real accuracy;
real threshold;
int max_iter;
real prior;
int max_load;
bool add_space_to_targets;
bool viterbi;
bool train_separate;
bool disk;
bool no_learn_means;
bool learn_var;
bool learn_weights;
real map_factor;
bool adapt_separate;
bool adapt_separate_set_data;
Allocator *allocator = new Allocator;
FileListCmdOption input_file_list("file name", "the list of inputs files or one data file");
input_file_list.isArgument(true);
FileListCmdOption target_file_list("file name", "the list of target files or one target file");
target_file_list.isArgument(true);
//===============================================================
//=================== The command-line ==========================
//===============================================================
// Construct the command line
CmdLine cmd;
cmd.setBOption("write log", false );
// Put the help line at the beginning
cmd.info(help);
// Adapt mode
cmd.addText("\nArguments:");
cmd.addSCmdArg("model_name1", &model_name1, "the list of the model models file");
cmd.addSCmdArg("lex_name1", &lex_name1, "the model lexicon file");
cmd.addSCmdArg("model_name2", &model_name2, "the list of the data models file");
cmd.addSCmdArg("lex_name2", &lex_name2, "the data lexicon file");
cmd.addSCmdArg("model_name3", &model_name3, "the list of the output models file");
cmd.addSCmdArg("lex_name3", &lex_name3, "the output lexicon file");
cmd.addCmdOption(&input_file_list);
cmd.addCmdOption(&target_file_list);
cmd.addSCmdArg("input model file", &model_file_in, "the input model file to be adapted");
cmd.addSCmdArg("output model file", &model_file_out, "the adapted output model file");
// Data Initialization
cmd.addText("\nInitialization Options:") ;
cmd.addICmdOption("-n_gaussians", &n_gaussians, 10, "number of Gaussians");
cmd.addICmdOption("-n_states", &n_states, 5, "number of states");
cmd.addSCmdOption("-file_states", &file_states, "", "file containing n_states per model");
cmd.addBCmdOption("-train_separate", &train_separate, false, "first train separate models");
// Data and Model Parameters
cmd.addText("\nData and Model Options:") ;
cmd.addBCmdOption("-symbol", &is_symbol, false, "targets are in symbol format");
cmd.addBCmdOption("-disk", &disk, false, "keep data on disk");
cmd.addICmdOption("-load", &max_load, -1, "max number of examples to load for train");
cmd.addSCmdOption("-spacing_model", &spacing_model,"", "name of silence phone");
cmd.addBCmdOption("-add_sil_to_targets", &add_space_to_targets, false, "add silence at begining of targets");
cmd.addSCmdOption("-sent_start_symbol" , &sent_start_symbol , SP_START , "symbol that will start every sentence" ) ;
cmd.addSCmdOption("-sent_end_symbol" , &sent_end_symbol , SP_END , "symbol that will end every sentence" ) ;
cmd.addSCmdOption("-spacing_symbol" , &spacing_symbol , SP_WORD ,"the silence dictionary symbol" ) ;
// Learning Options
cmd.addText("\nLearning Options:");
cmd.addBCmdOption("-viterbi", &viterbi, false, "viterbi training (instead of EM)");
cmd.addRCmdOption("-threshold", &threshold, 0.001, "variance threshold");
cmd.addRCmdOption("-prior", &prior, 0.001, "prior on the weights");
cmd.addICmdOption("-iter", &max_iter, 25, "max number of iterations of HMM");
cmd.addRCmdOption("-e", &accuracy, 0.00001, "end accuracy");
cmd.addBCmdOption("-adapt_separate", &adapt_separate, false, "adapt models separately");
cmd.addBCmdOption("-adapt_separate_set_data", &adapt_separate_set_data, true, "set data to non represented models when adapting models separately");
// MAP Options
cmd.addText("\nMAP Options:");
cmd.addRCmdOption("-map", &map_factor, 0.5, "adaptation factor [0-1]");
cmd.addBCmdOption("-no_means", &no_learn_means, false, "no enroll means");
cmd.addBCmdOption("-learn_var", &learn_var, false, "enroll var");
cmd.addBCmdOption("-learn_weights", &learn_weights, false, "enroll weights");
// Misc Options
cmd.addText("\nMisc Options:");
cmd.addSCmdOption("-dir", &dir_name, ".", "directory to save measures");
// Read the command line
cmd.read(argc, argv);
cmd.setWorkingDirectory(dir_name);
DiskXFile::setBigEndianMode();
//====================================================================
//=================== Data preparation ===============================
//====================================================================
Random::seed();
clock_t start_time , end_time ;
real total_time = 0.0 ;
start_time = clock() ;
MlHMM hmm;
//====================================================================
//=================== Set Options ====================================
//====================================================================
// initialization
hmm.setIOption("number of gaussians", n_gaussians );
hmm.setIOption("number of states", n_states );
hmm.setBOption("train separate", train_separate );
// data
hmm.setBOption("is symbol", is_symbol );
hmm.setBOption("add spacing", add_space_to_targets );
hmm.setIOption("max load", max_load );
hmm.setBOption("disk", disk );
// training
hmm.setROption("accuracy", accuracy );
hmm.setROption("threshold", threshold );
hmm.setIOption("iterations", max_iter );
hmm.setROption("prior", prior );
hmm.setBOption("viterbi", viterbi );
// map
hmm.setROption("map factor", map_factor );
hmm.setBOption("no means", no_learn_means );
hmm.setBOption("learn var", learn_var );
hmm.setBOption("learn weights", learn_weights );
hmm.setBOption("adapt separate", adapt_separate );
hmm.setBOption("adapt separate set data", adapt_separate_set_data );
hmm.setLexicon( model_name1, spacing_model, lex_name1, add_space_to_targets ? sent_start_symbol : NULL );
hmm.mapSetDataLexicon( model_name2, lex_name2, model_name3, lex_name3 );
//====================================================================
//=================== Create the DataSet ... =========================
//====================================================================
hmm.setData( input_file_list.file_names, input_file_list.n_files, target_file_list.file_names, target_file_list.n_files );
hmm.mapInit( file_states );
char* n_nll = strConcat(2,"hmm_train_val_",viterbi ? "viterbi" : "em");
allocator->retain(n_nll);
hmm.mapAdapt( model_file_in, model_file_out, cmd.getXFile(n_nll) );
end_time = clock() ;
total_time = (real)(end_time-start_time) / CLOCKS_PER_SEC ;
printf("\nTotal time spent adapting = %.2f secs\n", total_time) ;
delete allocator;
printf( SUCCESS );
return(0);
}
int main(int argc, char** argv) {
//signal(SIGINT, stop);
//signal(SIGTERM, stop);
std::set_terminate(handler);
int dimObservations;
int dimActions;
char* stringNActions;
int oscPort;
int oscRemotePort;
char* oscIP;
int agentId;
bool exportData;
int seed;
//=================== The command-line ==========================
// Construct the command line
CmdLine cmd;
// Put the help line at the beginning
cmd.info("Simple B-Tree agent");
cmd.addText("\nArguments:");
cmd.addICmdArg("agent_id", &agentId, "the id of the agent", true);
cmd.addICmdArg("dim_observations", &dimObservations, "observation dimension (without the reward)", true);
cmd.addICmdArg("dim_actions", &dimActions, "action dimension", true);
cmd.addSCmdArg("n_actions", &stringNActions, "number of actions as comma-separated values", true);
cmd.addText("\nOSC Options:");
cmd.addSCmdOption("-ip", &oscIP, "127.0.0.1", "the osc IP address", false);
cmd.addICmdOption("-port", &oscPort, 11000, "the osc local STARTING port (actual port will be port + agent_id)", false);
cmd.addICmdOption("-rport", &oscRemotePort, 12000, "the osc remote local port", false);
cmd.addText("\nMisc Options:");
cmd.addBCmdOption("-export-data", &exportData, false, "export the data to files", false );
cmd.addICmdOption("-seed", &seed, -1, "the random seed (-1 = based on time)");
// Read the command line
//int mode = cmd.read(argc, argv);
cmd.read(argc, argv);
// Parse n actions.
printf("Parsing actions: ");
unsigned int nActions[100];
char tmp[1000];
strcpy(tmp, stringNActions);
int k=0;
for (int i=0; i<dimActions-1; i++, k++) {
Q_ASSERT_ERROR_MESSAGE( sscanf(tmp, "%d,%s", &nActions[k], tmp) > 0, "Malformed argument <n_actions>: %s", stringNActions);
printf("%d ", nActions[k]);
}
Q_ASSERT_ERROR_MESSAGE( sscanf(tmp, "%d", &nActions[k]), "Malformed argument <n_actions>: %s", stringNActions);
printf("%d \n", nActions[k]);
oscPort += agentId;
char oscPortStr[100];
char oscRemotePortStr[100];
sprintf(oscPortStr, "%d", oscPort);
sprintf(oscRemotePortStr, "%d", oscRemotePort);
printf("Opening OSC connection (ip=%s, port=%s, remote_port=%s)\n", oscIP, oscPortStr, oscRemotePortStr);
OscManager::initOsc(oscIP, oscPortStr, oscRemotePortStr);
// Set random seed.
if (seed == -1)
{
randomSeed(time(NULL) + agentId);
}
else
{
randomSeed(seed);
}
ActionProperties actionProperties(ACTION_DIM, N_ACTIONS);
#if ATTRACTION_MODE
int nRepeat = (agentId + 1) * 10;
BehaviorTreeNode* root =
BT.priority()->setChildren(
BT.sequential()->setChildren(
Q_NEW(BoolCondition<TestBTreeAgent>)(&TestBTreeAgent::hasClosest, false),
BT.repeat(-1)->setChild(
BT.sequential()->setChildren(
Q_NEW(FloatCondition<TestBTreeAgent>)(&TestBTreeAgent::getClosestDistance, GREATER_OR_CLOSE, 0.05),
Q_NEW(ChooseAction)(&actionProperties, ATTRACT),
BT_END)
),
BT_END),
BT.repeat(-1)->setChild(
BT.probability()->setWeightedChildren(
BT.weighted(0.99, Q_NEW(ChooseAction)(&actionProperties, ATTRACT)),
BT.weighted(0.01, BT.failure()),
BT_END)
),
BT.sequential()->setChildren(
Q_NEW(BoolCondition<TestBTreeAgent>)(&TestBTreeAgent::hasClosest, true),
BT.repeat(500)->setChild(
Q_NEW(ChooseAction)(&actionProperties, REPEL)
),
BT_END),
BT_END);
// BehaviorTreeInternalNode* root = new PriorityNode();
// ProbabilityNode* probNode = new ProbabilityNode();
// probNode->addChild(new ChooseAction(&actionProperties, ATTRACT), 0.99f);
// probNode->addChild(new AlwaysFailure(), 0.01f);
//
// root->addChild((new SequentialNode())
// ->addChild(new BoolCondition<TestBTreeAgent>(&TestBTreeAgent::hasClosest, false))
// ->addChild((new PriorityNode())
// ->addChild((new RepeatNode(-1))
// ->addChild((new SequentialNode)
// ->addChild(new FloatCondition<TestBTreeAgent>(&TestBTreeAgent::getClosestDistance, GREATER_OR_CLOSE, 0.05))
// ->addChild(new ChooseAction(&actionProperties, ATTRACT))))
// ->addChild((new RepeatNode(-1))
// ->addChild(probNode))))
//
// ->addChild((new SequentialNode())
// ->addChild(new BoolCondition<TestBTreeAgent>(&TestBTreeAgent::hasClosest, true))
// ->addChild((new RepeatNode(500))
// ->addChild(new ChooseAction(&actionProperties, REPEL))));
#else
int nRepeat = (agentId + 1) * 10;
BehaviorTreeNode* root =
sequential()->CHILDREN(
repeat(nRepeat)->CHILDREN(
BT_NEW(ChooseAction)(&actionProperties, UP)
),
repeat(nRepeat)->CHILDREN(
BT_NEW(ChooseAction)(&actionProperties, RIGHT)
),
repeat(nRepeat)->CHILDREN(
BT_NEW(ChooseAction)(&actionProperties, DOWN)
),
repeat(nRepeat)->CHILDREN(
BT_NEW(ChooseAction)(&actionProperties, LEFT)
)
);
// BehaviorTreeInternalNode* root = new SequentialNode();
// root->addChild((new RepeatNode(nRepeat))
// ->addChild((new ChooseAction(&actionProperties, UP))))
// ->addChild((new RepeatNode(nRepeat))
// ->addChild((new ChooseAction(&actionProperties, RIGHT))))
// ->addChild((new RepeatNode(nRepeat))
// ->addChild((new ChooseAction(&actionProperties, DOWN))))
// ->addChild((new RepeatNode(nRepeat))
// ->addChild((new ChooseAction(&actionProperties, LEFT))));
#endif
Agent* agent = new TestBTreeAgent(&actionProperties, root);
printf("--- Creating environment ---\n");
Environment* env;
Environment* oscEnv = new OscRLEnvironment(agentId, dimObservations, actionProperties.dim());
if (exportData) {
char fileName[1000];
sprintf(fileName, "export-%d.raw", agentId);
DiskXFile* f = new(Alloc::instance()) DiskXFile(fileName, "w+");
env = new FileExportEnvironment(oscEnv, f, dimObservations, actionProperties.dim());
} else
env = oscEnv;
Qualia* qualia = new RLQualia(agent, env);
printf("--- Initialising ---\n");
qualia->init();
printf("--- Starting ---\n");
qualia->start();
printf("--- Looping ---\n");
stopLoop = false;
while (!stopLoop) {
qualia->step();
}
delete qualia;
delete agent;
delete env;
Q_DELETE(root);
//delete root;
return 0;
}
