void initCacheLRU( int tailleCache, int ordre, lm_t *lmlocal , dict_t *dict , char const * nomMach) {
FILE *f=NULL;
char nom1[1024], nomVocabPondere[1024];
lm=lmlocal;
if (nomMach !=NULL) {
f=fopen(nomMach,"r");
if (f==NULL) {
fprintf(stderr,"can't open %s\n",nomMach);
exit(1);
}
fscanf(f,"%s",nomVocabPondere);
fscanf(f,"%i",&nbTrainer); //machine number
cslm= new TrainerNgramSlist *[nbTrainer];
poids= new double[nbTrainer];
cslm_wlist = new WordList;
if (cslm_wlist != NULL)
cslm_wlist->Read(nomVocabPondere); /* shortlist length will be set with TrainerNgramSlist class */
for (int i =0; i<nbTrainer; i++)
{
if (fscanf(f,"%s%lf",nom1,poids+i)!=2) {fprintf(stderr,"error reading machine %i\n",i);exit(1);}
std::ifstream ifs;
ifs.open(nom1,ios::binary);
CHECK_FILE(ifs,nom1);
Mach *m = Mach::Read(ifs);
ifs.close();
m->Info();
cslm[i] = new TrainerNgramSlist(m, cslm_wlist, (char*)"");
}
fclose(f);
lmMakeIndexCSLM(dict,lm,cslm_wlist);
}
TAILLEMAX=tailleCache+1;
resuCSLM=(float **) malloc(nbTrainer*sizeof(float*));
resuCSLM[0]=(float *) malloc(nbTrainer*sizeof(float)*TAILLEMAX);
for (int i=1 ; i<nbTrainer ;i++)
resuCSLM[i]=resuCSLM[i-1]+TAILLEMAX;
ordreMax=ordre;
if (ordre%2 ==1) ordre++;
ordreHash=ordre/2;
ram=(Gram * )malloc(sizeof( Gram) * TAILLEMAX);
ram->w=(unsigned int *) calloc(TAILLEMAX*ordre, sizeof(unsigned int) );
for (int i =1 ;i<TAILLEMAX; i++) ram[i].w=ram[i-1].w+ordre;
}
int main (int argc, char *argv[])
{
ifstream ifs;
Mach *m;
for (int i=1; i<argc; i++) {
ifs.open(argv[i],ios::binary);
CHECK_FILE(ifs,argv[i]);
cout << endl << "Information on machine: " << argv[i] << endl;
m = Mach::Read(ifs);
cout << "Using file version " << Mach::GetFileId() << endl;
m->Info();
ifs.close();
delete m;
}
GpuUnlock();
return 0;
}
int main3(int argc, char *argv[])
{
MachConfig mach_config(true);
string mach_fname, test_fname, dev_fname;
int curr_it = 0;
Mach *mlp;
// select available options
mach_config
.sel_cmdline_option<std::string>("mach,m" , true )
.sel_cmdline_option<std::string>("test-data,t" , true )
.sel_cmdline_option<std::string>("dev-data,d" , true)
;
// parse parameters
if (mach_config.parse_options(argc, argv)) {
// get parameters
mach_fname = mach_config.get_mach();
test_fname = mach_config.get_test_data();
dev_fname = mach_config.get_dev_data();
curr_it = mach_config.get_curr_iter();
//gold file
CONF_test_file = test_fname;
CONF_gold_file = dev_fname;
}
else if (mach_config.help_request())
usage2(mach_config);
else {
if (mach_config.parsing_error())
usage2(mach_config, false);
Error(mach_config.get_error_string().c_str());
}
// Check if existing machine exists
const char *mach_fname_cstr = mach_fname.c_str();
struct stat stat_struct;
if (stat(mach_fname_cstr, &stat_struct)==0) {
// read existing network
ifstream ifs;
ifs.open(mach_fname_cstr,ios::binary);
CHECK_FILE(ifs,mach_fname_cstr);
mlp = Mach::Read(ifs);
ifs.close();
cout << "Found existing machine with " << mlp->GetNbBackw()
<< " backward passes, continuing training at iteration " << curr_it+1 << endl;
}
else {
Error("No such machine for eval.");
}
//mlp->Info();
//evaluating the test files
cout << "Now evaluating the test file..." << endl;
CONLLReader* reader = new CONLLReader();
CONLLWriter* writer = new CONLLWriter();
reader->startReading(CONF_test_file.c_str());
writer->startWriting(CONF_output_file.c_str());
//the list
HashMap * wl = training_space::load_wordlist(CONF_wl_file.c_str());
//some variables
int oov_num = 0; //out of vocabulary
int sen_num = 0; //sentence number
int token_num = 0; //token number
int miss_count = 0; //only work if the testing file already have answers
int TIME_start = clock() / 1000;
int TIME_start_fine = 0;
//calculate
DependencyInstance* x = reader->getNext();
while(x != NULL){
if(sen_num%500 == 0){
cout << "Having processed " << sen_num << ";period is "<<
(clock()/1000)-TIME_start_fine << "ms."<< endl;
TIME_start_fine = clock() / 1000;
}
sen_num++;
int length = x->forms->size();
token_num += length - 1;
double *tmp_scores = new double[length*length*2];
//construct scores using nn
int num_pair = length*(length-1); //2 * (0+(l-1))*l/2
REAL *mach_x = new REAL[num_pair*CONF_X_dim];
REAL *mach_y = new REAL[num_pair*CONF_Y_dim];
int* word_index = new int[length+2]; //including <s> and </s>
for(int i=0;i<length;i++){
HashMap::iterator iter = wl->find(x->forms->at(i));
if(iter == wl->end()){
oov_num++;
word_index[i+1] = wl->find(&unknown_token)->second;
}
else
word_index[i+1] = iter->second;
}
string sen_s = SENTENCE_START;
string sen_e = SENTENCE_END;
word_index[0] = wl->find(&sen_s)->second;
word_index[length+1] = wl->find(&sen_e)->second;
int pair_count = 0;
REAL* assign_x = mach_x;
for(int ii=0;ii<length;ii++){
for(int j=ii+1;j<length;j++){
for(int lr=0;lr<2;lr++){
#ifdef WHICH_TWO
//2 word pair
if(lr==E_RIGHT){
*assign_x++ = (REAL)word_index[ii+1];
*assign_x++ = (REAL)word_index[j+1];
}
else{
*assign_x++ = (REAL)word_index[j+1];
*assign_x++ = (REAL)word_index[ii+1];
}
#endif
#ifdef WHICH_SIX
//build mach_x : 6 words group(h-1 h h+1 m-1 m m+1)
if(lr==E_RIGHT){
*assign_x++ = (REAL)word_index[ii];
*assign_x++ = (REAL)word_index[ii+1];
*assign_x++ = (REAL)word_index[ii+2];
*assign_x++ = (REAL)word_index[j];
*assign_x++ = (REAL)word_index[j+1];
*assign_x++ = (REAL)word_index[j+2];
}
else{
*assign_x++ = (REAL)word_index[j];
*assign_x++ = (REAL)word_index[j+1];
*assign_x++ = (REAL)word_index[j+2];
*assign_x++ = (REAL)word_index[ii];
*assign_x++ = (REAL)word_index[ii+1];
*assign_x++ = (REAL)word_index[ii+2];
}
#endif
}
}
}
//- give it to nn
mlp->evaluate(mach_x,mach_y,num_pair,CONF_X_dim,CONF_Y_dim);
REAL* assign_y = mach_y;
for(int ii=0;ii<length;ii++){
for(int j=ii+1;j<length;j++){
for(int lr=0;lr<2;lr++){
int index = get_index2(length,ii,j,lr);
#ifdef ULTIMATE_DEBUG
cout << ii << "to" << j << ((lr==E_RIGHT)?'r':'l')
<< ":" << *assign_y << endl;
#endif
tmp_scores[index] = *assign_y++;
}
}
}
//- decode and write
vector<int> *ret = decodeProjective(length,tmp_scores);
for(int i2=1;i2<length;i2++){ //ignore root
if((*ret)[i2] != (*(x->heads))[i2])
miss_count ++;
}
#ifdef ULTIMATE_DEBUG
{
double score_here=0;
double score_right=0;
for(int i2=1;i2<length;i2++){ //ignore root
int head_here = (*ret)[i2];
int head_right = (*(x->heads))[i2];
score_here += get_score_from_array(length,head_here,i2,tmp_scores);
score_right += get_score_from_array(length,head_right,i2,tmp_scores);
}
cout << "Right:" << score_right << ";Output:" << score_here << endl;
}
#endif
delete x->heads;
x->heads = ret;
writer->write(x);
delete x;
delete []tmp_scores;
delete []mach_x;
delete []mach_y;
delete []word_index;
x = reader->getNext();
}
reader->finishReading();
writer->finishWriting();
delete reader;
delete writer;
cout << "Finished testing in " << (clock()/1000-TIME_start) << "ms" << endl;
//conclude and evaluate
cout << "Testing data description:\n"
<< "Sentences: " << sen_num << '\n'
<< "Tokens: " << token_num << '\n'
<< "OOV token: " << oov_num << '\n'
<< "Miss token: " << miss_count << endl;
string t;
DependencyEvaluator::evaluate(CONF_gold_file,CONF_output_file,t,false);
return 0;
}
int main2(int argc, char *argv[])
{
MachConfig mach_config(true);
string mach_fname, train_fname, dev_fname;
int curr_it = 0;
Mach *mlp;
// select available options
mach_config
.sel_cmdline_option<std::string>("mach,m" , true )
.sel_cmdline_option<std::string>("train-data,t" , true )
.sel_cmdline_option<std::string>("dev-data,d" , false)
.sel_cmdline_option<REAL> ("lrate-beg,L" , false)
.sel_cmdline_option<REAL> ("lrate-mult,M" , false)
.sel_cmdline_option<REAL> ("weight-decay,W", false)
.sel_cmdline_option<int> ("curr-iter,C" , false)
.sel_cmdline_option<int> ("last-iter,I" , false)
.sel_cmdline_option<int> ("block-size,B" , false)
;
// parse parameters
if (mach_config.parse_options(argc, argv)) {
// get parameters
mach_fname = mach_config.get_mach();
train_fname = mach_config.get_train_data();
dev_fname = mach_config.get_dev_data();
curr_it = mach_config.get_curr_iter();
}
else if (mach_config.help_request())
usage(mach_config);
else {
if (mach_config.parsing_error())
usage(mach_config, false);
Error(mach_config.get_error_string().c_str());
}
// Check if existing machine exists
const char *mach_fname_cstr = mach_fname.c_str();
struct stat stat_struct;
if (stat(mach_fname_cstr, &stat_struct)==0) {
// read existing network
ifstream ifs;
ifs.open(mach_fname_cstr,ios::binary);
CHECK_FILE(ifs,mach_fname_cstr);
mlp = Mach::Read(ifs);
ifs.close();
cout << "Found existing machine with " << mlp->GetNbBackw()
<< " backward passes, continuing training at iteration " << curr_it+1 << endl;
}
else {
mlp=mach_config.get_machine();
if (mlp == NULL)
Error(mach_config.get_error_string().c_str());
}
mlp->Info();
ErrFctMSE errfct(*mlp);
//ErrFctCrossEnt errfct(*mlp);
Trainer trainer(mlp, &errfct, (char *)train_fname.c_str(),
(dev_fname.empty() ? NULL : (char *)dev_fname.c_str()),
mach_config.get_lrate_beg(), mach_config.get_lrate_mult(),
mach_config.get_weight_decay(), mach_config.get_last_iter(), curr_it);
cout << "Initial error rate: " << 100.0*trainer.TestDev() << "%" << endl;
char sfname[1024], *p;
strcpy(sfname, mach_fname_cstr);
p=strstr(sfname, ".mach");
if (p) { *p=0; strcat(sfname,".best.mach"); }
trainer.TrainAndTest(sfname);
// save machine at the end
ofstream fs;
fs.open(mach_fname_cstr,ios::binary);
CHECK_FILE(fs,argv[4]);
mlp->Write(fs);
fs.close();
GpuUnlock();
if (mlp) delete mlp;
return 0;
}
void eval()
{
const char *mach_fname_cstr = CONF_mach_file.c_str();
Mach *mlp;
struct stat stat_struct;
if (stat(mach_fname_cstr, &stat_struct)==0) {
// read existing network
ifstream ifs;
ifs.open(mach_fname_cstr,ios::binary);
CHECK_FILE(ifs,mach_fname_cstr);
mlp = Mach::Read(ifs);
ifs.close();
cout << "Found existing machine" << endl;
}
else {
Error("No such machine for eval.");
}
//evaluating the test files
cout << "Now evaluating the test file..." << endl;
CONLLReader* reader = new CONLLReader();
CONLLWriter* writer = new CONLLWriter();
reader->startReading(CONF_test_file.c_str());
writer->startWriting(CONF_output_file.c_str());
//the list
HashMap * wl = load_wordlist(CONF_vocab_file.c_str());
#ifdef INSANE_DEBUG
FILE* x_file = fdopen(3,"w");
#endif
//some variables
int oov_num = 0; //out of vocabulary
int sen_num = 0; //sentence number
int token_num = 0; //token number
int miss_count = 0; //only work if the testing file already have answers
int TIME_start = clock() / 1000;
int TIME_start_fine = 0;
//calculate
DependencyInstance* x = reader->getNext();
while(x != NULL){
if(sen_num%1000 == 0){
cout << "Having processed " << sen_num << ";period is "<<
(clock()/1000)-TIME_start_fine << "ms."<< endl;
TIME_start_fine = clock() / 1000;
}
sen_num++;
int length = x->forms->size();
token_num += length - 1;
double *tmp_scores = new double[length*length*2];
//construct scores using nn
int num_pair = length*(length-1); //2 * (0+(l-1))*l/2
REAL *mach_x = new REAL[num_pair*IND_CONF_x_dim_final];
REAL *mach_y = new REAL[num_pair*CONF_y_class_size];
int *word_index = get_word_index(length,x,wl,&oov_num);
int pair_count = 0;
REAL* assign_x = mach_x;
for(int ii=0;ii<length;ii++){
for(int j=ii+1;j<length;j++){
for(int lr=0;lr<2;lr++){
//build mach_x
if(lr==E_RIGHT)
fill_feature(length,ii,j,word_index,assign_x);
else
fill_feature(length,j,ii,word_index,assign_x);
assign_x += IND_CONF_x_dim_final;
}
}
}
//- give it to nn
mlp->evaluate(mach_x,mach_y,num_pair,IND_CONF_x_dim_final,CONF_y_class_size);
REAL* assign_y = mach_y;
for(int ii=0;ii<length;ii++){
for(int j=ii+1;j<length;j++){
for(int lr=0;lr<2;lr++){
int index = get_index2(length,ii,j,lr);
//important ...
double temp = 0;
if(CONF_if_y_calss){
for(int c=0;c<CONF_y_class_size;c++)
temp += (*assign_y++)*c;
tmp_scores[index] = temp;
}
else
tmp_scores[index] = *assign_y++;
}
}
}
//- decode and write
vector<int> *ret = decodeProjective(length,tmp_scores);
for(int i2=1;i2<length;i2++){ //ignore root
if((*ret)[i2] != (*(x->heads))[i2])
miss_count ++;
}
delete x->heads;
x->heads = ret;
writer->write(x);
delete x;
delete []tmp_scores;
#ifdef INSANE_DEBUG
fprintf(x_file,"Sentence %d:\n",sen_num);
for(int i=0;i<num_pair;i++){
for(int j=0;j<IND_CONF_x_dim_final;j++)
fprintf(x_file,"%d ",(int)mach_x[i*IND_CONF_x_dim_final+j]);
fprintf(x_file,"\n");
}
#endif
delete []mach_x;
delete []mach_y;
delete []word_index;
x = reader->getNext();
}
#ifdef INSANE_DEBUG
fclose(x_file);
#endif
reader->finishReading();
writer->finishWriting();
delete reader;
delete writer;
cout << "Finished testing in " << (clock()/1000-TIME_start) << "ms" << endl;
//conclude and evaluate
cout << "Testing data description:\n"
<< "Sentences: " << sen_num << '\n'
<< "Tokens: " << token_num << '\n'
<< "OOV token: " << oov_num << '\n'
<< "Miss token: " << miss_count << endl;
string t;
DependencyEvaluator::evaluate(CONF_gold_file,CONF_output_file,t,false);
}
