| #!/usr/local/bin/perl |
| use Inline C; |
| use Inline C => Config => CLEAN_AFTER_BUILD => 0, ccflags => $Config{ccflags}." -O4"; |
| #use Inline C => Config => BUILD_NOISY => 1, CFLAGS => $Config{cflags}." -O4 -mtune k9"; |
| #use Inline C => Config => CLEAN_AFTER_BUILD => 0, ccflags => $Config{ccflags}." -Ofast -march k8 -mtune k8 "; |
| use Mojolicious::Lite; |
| use Mojo::JSON qw(decode_json encode_json to_json); |
| use Encode qw(decode encode); |
| use Getopt::Std; |
| use Mojo::Server::Daemon; |
| use Cwd; |
| app->static->paths->[0] = getcwd; |
| |
| plugin 'Log::Access'; |
| plugin "RequestBase"; |
| |
| our $opt_i = 0; # latin1-input? |
| our $opt_l = undef; |
| our $opt_p = 5676; |
| our $opt_m; |
| our $opt_M; |
| our $opt_n = ''; |
| our $opt_d; |
| our $opt_G; |
| |
| my %marked; |
| my $training_args=""; |
| my $mergedEnd=0; |
| my %cache; |
| |
| getopts('d:Gil:p:m:n:M:'); |
| |
| if($opt_M) { |
| open my $handle, '<:encoding(UTF-8)', $opt_M |
| or die "Can't open '$opt_M' for reading: $!"; |
| while(<$handle>) { |
| foreach my $mw (split /\s+/) { |
| $marked{$mw}=1 |
| } |
| } |
| close($handle); |
| } |
| |
| # -cbow 1 -size 200 -window 8 -negative 25 -hs 0 -sample 1e-4 -threads 40 -binary 1 -iter 15 |
| if(!$ARGV[0]) { |
| init_net("vectors15.bin", $opt_n, ($opt_i? 1 : 0)); |
| } else { |
| init_net($ARGV[0], $opt_n, ($opt_i? 1 : 0)); |
| if(open(FILE, "$ARGV[0].args")) { |
| $training_args = <FILE>; |
| } |
| close(FILE); |
| } |
| |
| if($opt_m) { |
| $mergedEnd = mergeVectors($opt_m); |
| } |
| |
| |
| if($opt_d) { # -d: dump vecs and exit |
| dump_vecs($opt_d); |
| exit; |
| } |
| |
| my $daemon = Mojo::Server::Daemon->new( |
| app => app, |
| listen => ['http://'.($opt_l ? $opt_l : '*').":$opt_p"] |
| ); |
| |
| if($opt_G) { |
| print "Filtering garbage\n"; |
| filter_garbage(); |
| } |
| |
| get '/derekovecs/js/*' => sub { |
| my $c = shift; |
| my $url = $c->req->url; |
| $url =~ s@/derekovecs@@g; |
| $c->app->log->info("GET: " . $url); |
| $c->reply->static($url); |
| }; |
| |
| get '/' => sub { |
| my $c = shift; |
| $c->app->log->info("get: ".$c->req->url->to_abs); |
| my $word=$c->param('word'); |
| my $no_nbs=$c->param('n') || 100; |
| my $no_iterations=$c->param('N') || 2000; |
| my $perplexity=$c->param('perplexity') || 20; |
| my $epsilon=$c->param('epsilon') || 5; |
| my $som=$c->param('som') || 0; |
| my $searchBaseVocabFirst=$c->param('sbf') || 0; |
| my $sort=$c->param('sort') || 0; |
| my $json=$c->param('json') || 0; |
| my $res; |
| my @lists; |
| my @collocations; |
| if(defined($word) && $word !~ /^\s*$/) { |
| $c->inactivity_timeout(300); |
| $word =~ s/\s+/ /g; |
| for my $w (split(' *\| *', $word)) { |
| if ($cache{$w}) { |
| $c->app->log->info("Getting $w results from cache"); |
| $res = $cache{$w} |
| } else { |
| $c->app->log->info('Looking for neighbours of '.$w); |
| if($opt_i) { |
| $res = get_neighbours(encode("iso-8859-1", $w), $no_nbs, $sort, $searchBaseVocabFirst); |
| } else { |
| $res = get_neighbours($w, $no_nbs, $sort, $searchBaseVocabFirst); |
| } |
| $cache{$w} = $res; |
| } |
| push(@lists, $res->{paradigmatic}); |
| } |
| } |
| $word =~ s/ *\| */ | /g; |
| if($json) { |
| return $c->render(json => {word => $word, list => \@lists, collocators=>$res->{syntagmatic}}); |
| } else { |
| $c->render(template=>"index", word=>$word, no_nbs=>$no_nbs, no_iterations => $no_iterations, epsilon=> $epsilon, perplexity=> $perplexity, show_som=>$som, searchBaseVocabFirst=>$searchBaseVocabFirst, sort=>$sort, training_args=>$training_args, mergedEnd=> $mergedEnd, marked=>\%marked, lists=> \@lists, collocators=> $res->{syntagmatic}); |
| } |
| }; |
| |
| $daemon->run; # app->start; |
| |
| exit; |
| |
| __END__ |
| |
| __C__ |
| #include <stdio.h> |
| #include <string.h> |
| #include <math.h> |
| #include <malloc.h> |
| #include <stdlib.h> //strlen |
| #include <sys/mman.h> |
| #include <pthread.h> |
| |
| #define max_size 2000 |
| #define max_w 50 |
| #define MAX_NEIGHBOURS 1000 |
| #define MAX_WORDS -1 |
| #define MAX_THREADS 100 |
| #define MAX_CC 50 |
| #define EXP_TABLE_SIZE 1000 |
| #define MAX_EXP 6 |
| #define MIN_RESP 0.50 |
| |
| //the thread function |
| void *connection_handler(void *); |
| |
| typedef struct { |
| long long *index; |
| float *dist; |
| float *norm; |
| long long *pos; |
| int length; |
| } knn; |
| |
| typedef struct { |
| long long wordi[MAX_NEIGHBOURS]; |
| char sep[MAX_NEIGHBOURS]; |
| int length; |
| } wordlist; |
| |
| typedef struct { |
| wordlist *wl; |
| char *token; |
| int N; |
| long from; |
| unsigned long upto; |
| float *target_sums; |
| } knnpars; |
| |
| float *M, *M2=0L, *syn1neg_window, *expTable; |
| char *vocab; |
| char *garbage = NULL; |
| |
| long long words, size, merged_end; |
| long long merge_words = 0; |
| int num_threads=20; |
| int latin_enc=0; |
| int window; |
| |
| int init_net(char *file_name, char *net_name, int latin) { |
| FILE *f, *binvecs, *binwords; |
| int binwords_fd, binvecs_fd, net_fd, i; |
| long long a, b, c, d, cn; |
| float len; |
| double val; |
| |
| char binvecs_fname[256], binwords_fname[256]; |
| strcpy(binwords_fname, file_name); |
| strcat(binwords_fname, ".words"); |
| strcpy(binvecs_fname, file_name); |
| strcat(binvecs_fname, ".vecs"); |
| |
| latin_enc = latin; |
| f = fopen(file_name, "rb"); |
| if (f == NULL) { |
| printf("Input file %s not found\n", file_name); |
| return -1; |
| } |
| fscanf(f, "%lld", &words); |
| if(MAX_WORDS > 0 && words > MAX_WORDS) words = MAX_WORDS; |
| fscanf(f, "%lld", &size); |
| if( (binvecs_fd = open(binvecs_fname, O_RDONLY)) < 0 || (binwords_fd = open(binwords_fname, O_RDONLY)) < 0) { |
| printf("Converting %s to memory mappable structures\n", file_name); |
| vocab = (char *)malloc((long long)words * max_w * sizeof(char)); |
| M = (float *)malloc((long long)words * (long long)size * sizeof(float)); |
| if (M == NULL) { |
| printf("Cannot allocate memory: %lld MB %lld %lld\n", (long long)words * size * sizeof(float) / 1048576, words, size); |
| return -1; |
| } |
| if(strstr(file_name, ".txt")) { |
| for (b = 0; b < words; b++) { |
| a = 0; |
| while (1) { |
| vocab[b * max_w + a] = fgetc(f); |
| if (feof(f) || (vocab[b * max_w + a] == ' ')) break; |
| if ((a < max_w) && (vocab[b * max_w + a] != '\n')) a++; |
| } |
| vocab[b * max_w + a] = 0; |
| len = 0; |
| for (a = 0; a < size; a++) { |
| fscanf(f, "%lf", &val); |
| M[a + b * size] = val; |
| len += val * val; |
| } |
| len = sqrt(len); |
| for (a = 0; a < size; a++) M[a + b * size] /= len; |
| } |
| } else { |
| for (b = 0; b < words; b++) { |
| a = 0; |
| while (1) { |
| vocab[b * max_w + a] = fgetc(f); |
| if (feof(f) || (vocab[b * max_w + a] == ' ')) break; |
| if ((a < max_w) && (vocab[b * max_w + a] != '\n')) a++; |
| } |
| vocab[b * max_w + a] = 0; |
| fread(&M[b * size], sizeof(float), size, f); |
| len = 0; |
| for (a = 0; a < size; a++) len += M[a + b * size] * M[a + b * size]; |
| len = sqrt(len); |
| for (a = 0; a < size; a++) M[a + b * size] /= len; |
| } |
| } |
| if( (binvecs = fopen(binvecs_fname, "wb")) != NULL && (binwords = fopen(binwords_fname, "wb")) != NULL) { |
| fwrite(M, sizeof(float), (long long)words * (long long)size, binvecs); |
| fclose(binvecs); |
| fwrite(vocab, sizeof(char), (long long)words * max_w, binwords); |
| fclose(binwords); |
| } |
| } |
| if( (binvecs_fd = open(binvecs_fname, O_RDONLY)) >= 0 && (binwords_fd = open(binwords_fname, O_RDONLY)) >= 0) { |
| M = mmap(0, sizeof(float) * (long long)words * (long long)size, PROT_READ, MAP_SHARED, binvecs_fd, 0); |
| vocab = mmap(0, sizeof(char) * (long long)words * max_w, PROT_READ, MAP_SHARED, binwords_fd, 0); |
| if (M == MAP_FAILED || vocab == MAP_FAILED) { |
| close(binvecs_fd); |
| close(binwords_fd); |
| fprintf(stderr, "Cannot mmap %s or %s\n", binwords_fname, binvecs_fname); |
| exit(-1); |
| } |
| } else { |
| fprintf(stderr, "Cannot open %s or %s\n", binwords_fname, binvecs_fname); |
| exit(-1); |
| } |
| fclose(f); |
| |
| if(net_name && strlen(net_name) > 0) { |
| if( (net_fd = open(net_name, O_RDONLY)) >= 0) { |
| window = (lseek(net_fd, 0, SEEK_END) - sizeof(float) * words * size) / words / size / sizeof(float) / 2; |
| // lseek(net_fd, sizeof(float) * words * size, SEEK_SET); |
| // munmap(M, sizeof(float) * words * size); |
| M2 = mmap(0, sizeof(float) * words * size + sizeof(float) * 2 * window * size * words, PROT_READ, MAP_SHARED, net_fd, 0); |
| if (M2 == MAP_FAILED) { |
| close(net_fd); |
| fprintf(stderr, "Cannot mmap %s\n", net_name); |
| exit(-1); |
| } |
| syn1neg_window = M2 + words * size; |
| } else { |
| fprintf(stderr, "Cannot open %s\n", net_name); |
| exit(-1); |
| } |
| fprintf(stderr, "Successfully memmaped %s. Determined window size: %d\n", net_name, window); |
| } |
| |
| expTable = (float *) malloc((EXP_TABLE_SIZE + 1) * sizeof(float)); |
| for (i = 0; i < EXP_TABLE_SIZE; i++) { |
| expTable[i] = exp((i / (float) EXP_TABLE_SIZE * 2 - 1) * MAX_EXP); // Precompute the exp() table |
| expTable[i] = expTable[i] / (expTable[i] + 1); // Precompute f(x) = x / (x + 1) |
| } |
| return 0; |
| } |
| |
| long mergeVectors(char *file_name){ |
| FILE *f, *binvecs, *binwords; |
| int binwords_fd, binvecs_fd, net_fd, i; |
| long long a, b, c, d, cn; |
| float len; |
| float *merge_vecs; |
| char *merge_vocab; |
| /* long long merge_words, merge_size; */ |
| long long merge_size; |
| |
| char binvecs_fname[256], binwords_fname[256]; |
| strcpy(binwords_fname, file_name); |
| strcat(binwords_fname, ".words"); |
| strcpy(binvecs_fname, file_name); |
| strcat(binvecs_fname, ".vecs"); |
| |
| f = fopen(file_name, "rb"); |
| if (f == NULL) { |
| printf("Input file %s not found\n", file_name); |
| exit -1; |
| } |
| fscanf(f, "%lld", &merge_words); |
| fscanf(f, "%lld", &merge_size); |
| if(merge_size != size){ |
| fprintf(stderr, "vectors must have the same length\n"); |
| exit(-1); |
| } |
| if( (binvecs_fd = open(binvecs_fname, O_RDONLY)) >= 0 && (binwords_fd = open(binwords_fname, O_RDONLY)) >= 0) { |
| merge_vecs = malloc(sizeof(float) * (words + merge_words) * size); |
| merge_vocab = malloc(sizeof(char) * (words + merge_words) * max_w); |
| if (merge_vecs == NULL || merge_vocab == NULL) { |
| close(binvecs_fd); |
| close(binwords_fd); |
| fprintf(stderr, "Cannot reserve memory for %s or %s\n", binwords_fname, binvecs_fname); |
| exit(-1); |
| } |
| read(binvecs_fd, merge_vecs, merge_words * size * sizeof(float)); |
| read(binwords_fd, merge_vocab, merge_words * max_w); |
| } else { |
| fprintf(stderr, "Cannot open %s or %s\n", binwords_fname, binvecs_fname); |
| exit(-1); |
| } |
| printf("Successfully reallocated memory\nMerging...\n"); |
| fflush(stdout); |
| memcpy(merge_vecs + merge_words * size, M, words * size * sizeof(float)); |
| memcpy(merge_vocab + merge_words * max_w, vocab, words * max_w); |
| munmap(M, words * size * sizeof(float)); |
| munmap(vocab, words * max_w); |
| M = merge_vecs; |
| vocab = merge_vocab; |
| merged_end = merge_words; |
| words += merge_words; |
| fclose(f); |
| printf("merged_end: %lld, words: %lld\n", merged_end, words); |
| return((long) merged_end); |
| } |
| |
| void filter_garbage() { |
| long i; |
| unsigned char *w, previous, c; |
| garbage = malloc(words); |
| memset(garbage, 0, words); |
| for (i = 0; i < words; i++) { |
| w = vocab + i * max_w; |
| previous = 0; |
| while((c = *w++) && !garbage[i]) { |
| if( ((c <= 90 && c >= 65) && (previous >= 97 && previous <= 122)) || |
| (previous == '-' && (c & 32)) || |
| (previous == 0xc2 && (c == 0xa4 || c == 0xb6 )) |
| ) { |
| garbage[i]=1; |
| continue; |
| } |
| previous = c; |
| } |
| } |
| return; |
| } |
| |
| void *getCollocators(knnpars *pars) { |
| int N = pars->N; |
| int cc = pars->wl->wordi[0]; |
| knn *nbs = NULL; |
| long window_layer_size = size * window * 2; |
| long a, b, c, d, e, window_offset, target, max_target=0, maxmax_target; |
| float f, max_f, maxmax_f; |
| float *target_sums, *bestf, *bestn, worstbest, wpos_sum; |
| long long *besti, *bestp; |
| |
| if(M2 == NULL || cc == -1) |
| return NULL; |
| |
| a = posix_memalign((void **) &target_sums, 128, words * sizeof(float)); |
| besti = malloc(N * sizeof(long long)); |
| bestp = malloc(N * sizeof(long long)); |
| bestf = malloc(N * sizeof(float)); |
| bestn = malloc(N * sizeof(float)); |
| |
| worstbest = MIN_RESP; |
| |
| for (b = 0; b < words; b++) |
| target_sums[b]=0; |
| for (b = 0; b < N; b++) { |
| besti[b] = -1; |
| bestn[b] = 1; |
| bestf[b] = worstbest; |
| } |
| |
| d = cc; |
| maxmax_f = -1; |
| maxmax_target = 0; |
| |
| for (a = pars->from; a < pars->upto; a++) { |
| if(a >= window) |
| a++; |
| wpos_sum = 0; |
| printf("window pos: %ld\n", a); |
| if (a != window) { |
| max_f = -1; |
| window_offset = a * size; |
| if (a > window) |
| window_offset -= size; |
| for(target = 0; target < words; target ++) { |
| if(target == d) |
| continue; |
| f = 0; |
| for (c = 0; c < size; c++) |
| f += M2[d* size + c] * syn1neg_window[target * window_layer_size + window_offset + c]; |
| if (f < -MAX_EXP) |
| continue; |
| else if (f > MAX_EXP) |
| continue; |
| else |
| f = expTable[(int) ((f + MAX_EXP) * (EXP_TABLE_SIZE / MAX_EXP / 2))]; |
| wpos_sum += f; |
| |
| target_sums[target] += f; |
| if(f > worstbest) { |
| for (b = 0; b < N; b++) { |
| if (f > bestf[b]) { |
| memmove(bestf + b + 1, bestf + b, (N - b -1) * sizeof(float)); |
| memmove(besti + b + 1, besti + b, (N - b -1) * sizeof(long long)); |
| memmove(bestp + b + 1, bestp + b, (N - b -1) * sizeof(long long)); |
| bestf[b] = f; |
| besti[b] = target; |
| bestp[b] = window-a; |
| break; |
| } |
| } |
| if(b == N - 1) |
| worstbest = bestf[N-1]; |
| } |
| } |
| printf("%d %.2f\n", max_target, max_f); |
| printf("%s (%.2f) ", &vocab[max_target * max_w], max_f); |
| if(max_f > maxmax_f) { |
| maxmax_f = max_f; |
| maxmax_target = max_target; |
| } |
| for (b = 0; b < N; b++) |
| if(bestp[b] == window-a) |
| bestn[b] = bestf[b] / wpos_sum; |
| } else { |
| printf("\x1b[1m%s\x1b[0m ", &vocab[d*max_w]); |
| } |
| |
| } |
| for (b = 0; b < words; b++) |
| pars->target_sums[b] += (target_sums[b] / wpos_sum ) / (window * 2); |
| free(target_sums); |
| for(b=0; b<N && besti[b] >= 0; b++) // THIS LOOP IS NEEDED (b...) |
| printf("%s %.2f %d * ", &vocab[besti[b]*max_w], bestf[b], bestp[b]); |
| printf("\n"); |
| nbs = malloc(sizeof(knn)); |
| nbs->index = besti; |
| nbs->dist = bestf; |
| nbs->norm = bestn; |
| nbs->pos = bestp; |
| nbs->length = b-1; |
| pthread_exit(nbs); |
| } |
| |
| wordlist *getTargetWords(char *st1, int search_backw) { |
| wordlist *wl = malloc(sizeof(wordlist)); |
| char st[100][max_size], sep[100]; |
| long a, b=0, c=0, cn=0; |
| int unmerged; |
| |
| while (1) { |
| st[cn][b] = st1[c]; |
| b++; |
| c++; |
| st[cn][b] = 0; |
| if (st1[c] == 0) break; |
| if (st1[c] == ' ' || st1[c] == '-') { |
| sep[cn++] = st1[c]; |
| b = 0; |
| c++; |
| } |
| } |
| cn++; |
| for (a = 0; a < cn; a++) { |
| if(search_backw) { |
| for (b = words - 1; b >= 0; b--) if (!strcmp(&vocab[b * max_w], st[a])) break; |
| } else { |
| for (b = 0; b < words; b++) if (!strcmp(&vocab[b * max_w], st[a])) break; |
| } |
| if (b == words) b = -1; |
| wl->wordi[a] = b; |
| fprintf(stderr, "Word: \"%s\" Position in vocabulary: %lld\n", st[a], wl->wordi[a]); |
| if (b == -1) { |
| fprintf(stderr, "Out of dictionary word!\n"); |
| cn--; |
| break; |
| } |
| } |
| wl->length=cn; |
| return(wl); |
| } |
| |
| void *_get_neighbours(knnpars *pars) { |
| char *st1 = pars->token; |
| int N = pars->N; |
| long from = pars -> from; |
| unsigned long upto = pars -> upto; |
| char file_name[max_size], st[100][max_size], *sep; |
| float dist, len, *bestd, vec[max_size]; |
| long long a, b, c, d, cn, *bi, *besti; |
| char ch; |
| knn *nbs = NULL; |
| wordlist *wl = pars->wl; |
| |
| besti = malloc(N * sizeof(long long)); |
| bestd = malloc(N * sizeof(float)); |
| |
| float worstbest=-1; |
| |
| for (a = 0; a < N; a++) bestd[a] = 0; |
| a = 0; |
| bi = wl->wordi; |
| cn = wl->length; |
| sep = wl->sep; |
| b = bi[0]; |
| c = 0; |
| if (b == -1) { |
| N = 0; |
| goto end; |
| } |
| for (a = 0; a < size; a++) vec[a] = 0; |
| for (b = 0; b < cn; b++) { |
| if (bi[b] == -1) continue; |
| if(b>0 && sep[b-1] == '-') |
| for (a = 0; a < size; a++) vec[a] -= M[a + bi[b] * size]; |
| else |
| for (a = 0; a < size; a++) vec[a] += M[a + bi[b] * size]; |
| } |
| len = 0; |
| for (a = 0; a < size; a++) len += vec[a] * vec[a]; |
| len = sqrt(len); |
| for (a = 0; a < size; a++) vec[a] /= len; |
| for (a = 0; a < N; a++) bestd[a] = -1; |
| for (c = from; c < upto; c++) { |
| if(garbage && garbage[c]) continue; |
| a = 0; |
| // do not skip taget word |
| // for (b = 0; b < cn; b++) if (bi[b] == c) a = 1; |
| // if (a == 1) continue; |
| dist = 0; |
| for (a = 0; a < size; a++) dist += vec[a] * M[a + c * size]; |
| if(dist > worstbest) { |
| for (a = 0; a < N; a++) { |
| if (dist > bestd[a]) { |
| memmove(bestd + a + 1, bestd + a, (N - a -1) * sizeof(float)); |
| memmove(besti + a + 1, besti + a, (N - a -1) * sizeof(long long)); |
| bestd[a] = dist; |
| besti[a] = c; |
| break; |
| } |
| } |
| worstbest = bestd[N-1]; |
| } |
| } |
| |
| nbs = malloc(sizeof(knn)); |
| nbs->index = besti; |
| nbs->dist = bestd; |
| nbs->length = N; |
| end: |
| pthread_exit(nbs); |
| } |
| |
| |
| SV *get_neighbours(char *st1, int N, int sort_by, int search_backw) { |
| HV *result = newHV(); |
| float *target_sums, bestd[MAX_NEIGHBOURS], bestn[MAX_NEIGHBOURS], bests[MAX_NEIGHBOURS], vec[max_size]; |
| long long old_words; |
| long besti[MAX_NEIGHBOURS], bestp[MAX_NEIGHBOURS], a, b, c, d, slice; |
| knn *para_nbs[MAX_THREADS]; |
| knn *syn_nbs[MAX_THREADS]; |
| knnpars pars[MAX_THREADS]; |
| pthread_t *pt = (pthread_t *)malloc((num_threads+1) * sizeof(pthread_t)); |
| wordlist *wl; |
| int syn_threads = (M2? window * 2 : 0); |
| int para_threads = num_threads - syn_threads; |
| |
| if(N>MAX_NEIGHBOURS) N=MAX_NEIGHBOURS; |
| |
| |
| wl = getTargetWords(st1, search_backw); |
| if(wl->length < 1) |
| goto end; |
| |
| old_words = words; |
| if(merge_words > 0) |
| words = merge_words * 1.25; /* HACK */ |
| slice = words / para_threads; |
| |
| a = posix_memalign((void **) &target_sums, 128, words * sizeof(float)); |
| for(a = 0; a < words; a++) |
| target_sums[a] = 0; |
| |
| printf("Starting %d threads\n", para_threads); |
| fflush(stdout); |
| for(a=0; a < para_threads; a++) { |
| pars[a].token = st1; |
| pars[a].wl = wl; |
| pars[a].N = N; |
| pars[a].from = a*slice; |
| pars[a].upto = ((a+1)*slice > words? words:(a+1)*slice); |
| pthread_create(&pt[a], NULL, _get_neighbours, (void *) &pars[a]); |
| } |
| if(M2) { |
| for(a=0; a < syn_threads; a++) { |
| pars[a + para_threads].target_sums = target_sums; |
| pars[a + para_threads].wl = wl; |
| pars[a + para_threads].N = N; |
| pars[a + para_threads].from = a; |
| pars[a + para_threads].upto = a+1; |
| pthread_create(&pt[a + para_threads], NULL, getCollocators, (void *) &pars[a + para_threads]); |
| } |
| } |
| printf("Waiting for para threads to join\n"); |
| fflush(stdout); |
| for (a = 0; a < para_threads; a++) pthread_join(pt[a], ¶_nbs[a]); |
| printf("Para threads joint\n"); |
| fflush(stdout); |
| |
| /* if(!syn_nbs[0]) */ |
| /* goto end; */ |
| |
| for(b=0; b < N; b++) { |
| besti[b] = para_nbs[0]->index[b]; |
| bestd[b] = para_nbs[0]->dist[b]; |
| } |
| |
| for(a=1; a < para_threads; a++) { |
| for(b=0; b < para_nbs[a]->length && para_nbs[a]->index[b] >= 0; b++) { |
| for(c=0; c < N * para_threads; c++) { |
| if(para_nbs[a]->dist[b] > bestd[c]) { |
| for(d=N-1; d>c; d--) { |
| bestd[d] = bestd[d-1]; |
| besti[d] = besti[d-1]; |
| } |
| besti[c] = para_nbs[a]->index[b]; |
| bestd[c] = para_nbs[a]->dist[b]; |
| break; |
| } |
| } |
| } |
| } |
| |
| AV* array = newAV(); |
| int i; |
| int l1_words=0, l2_words=0; |
| for (a = 0, i = 0; i < N && a < 600; a++) { |
| long long c = besti[a]; |
| if(merge_words > 0) { |
| if(c >= merge_words) { |
| if(l1_words > N / 2) |
| continue; |
| else |
| l1_words++; |
| } else { |
| if(l2_words > N / 2) |
| continue; |
| else |
| l2_words++; |
| } |
| } |
| fflush(stdout); |
| printf("%s l1:%d l2:%d i:%d a:%ld\n", &vocab[c * max_w], l1_words, l2_words, i, a); |
| |
| HV* hash = newHV(); |
| SV* word = newSVpvf(&vocab[c * max_w], 0); |
| if(latin_enc == 0) SvUTF8_on(word); |
| fflush(stdout); |
| hv_store(hash, "word", strlen("word"), word , 0); |
| hv_store(hash, "dist", strlen("dist"), newSVnv(bestd[a]), 0); |
| hv_store(hash, "rank", strlen("rank"), newSVuv(besti[a]), 0); |
| AV *vector = newAV(); |
| for (b = 0; b < size; b++) { |
| av_push(vector, newSVnv(M[b + besti[a] * size])); |
| } |
| hv_store(hash, "vector", strlen("vector"), newRV_noinc((SV*)vector), 0); |
| av_push(array, newRV_noinc((SV*)hash)); |
| i++; |
| } |
| hv_store(result, "paradigmatic", strlen("paradigmatic"), newRV_noinc((SV*)array), 0); |
| |
| for(b=0; b < MAX_NEIGHBOURS; b++) { |
| besti[b] = -1L; |
| bestd[b] = 0; |
| bestn[b] = 0; |
| bestp[b] = 0; |
| bests[b] = 0; |
| } |
| |
| if (M2) { |
| printf("Waiting for syn threads to join\n"); |
| fflush(stdout); |
| for (a = 0; a < syn_threads; a++) pthread_join(pt[a+para_threads], &syn_nbs[a]); |
| printf("syn threads joint\n"); |
| fflush(stdout); |
| |
| for(b=0; b < syn_nbs[0]->length; b++) { |
| besti[b] = syn_nbs[0]->index[b]; |
| bestd[b] = syn_nbs[0]->dist[b]; |
| bestn[b] = syn_nbs[0]->norm[b]; |
| bestp[b] = syn_nbs[0]->pos[b]; |
| bests[b] = target_sums[syn_nbs[0]->index[b]]; |
| } |
| |
| if(sort_by != 1) { // sort by responsiveness |
| for(a=1; a < syn_threads; a++) { |
| for(b=0; b < syn_nbs[a]->length; b++) { |
| for(c=0; c < MAX_NEIGHBOURS; c++) { |
| if(syn_nbs[a]->dist[b] > bestd[c]) { |
| for(d=MAX_NEIGHBOURS-1; d>c; d--) { |
| bestd[d] = bestd[d-1]; |
| besti[d] = besti[d-1]; |
| bestn[d] = bestn[d-1]; |
| bestp[d] = bestp[d-1]; |
| } |
| besti[c] = syn_nbs[a]->index[b]; |
| bestd[c] = syn_nbs[a]->dist[b]; |
| bestn[c] = syn_nbs[a]->norm[b]; |
| bestp[c] = syn_nbs[a]->pos[b]; |
| break; |
| } |
| } |
| } |
| } |
| } else { // sort by mean p |
| for(a=1; a < syn_threads; a++) { |
| for(b=0; b < syn_nbs[a]->length; b++) { |
| for(c=0; c < MAX_NEIGHBOURS; c++) { |
| if(target_sums[syn_nbs[a]->index[b]] > bests[c]) { |
| for(d=MAX_NEIGHBOURS-1; d>c; d--) { |
| bestd[d] = bestd[d-1]; |
| besti[d] = besti[d-1]; |
| bestn[d] = bestn[d-1]; |
| bestp[d] = bestp[d-1]; |
| bests[d] = bests[d-1]; |
| } |
| besti[c] = syn_nbs[a]->index[b]; |
| bestd[c] = syn_nbs[a]->dist[b]; |
| bestn[c] = syn_nbs[a]->norm[b]; |
| bestp[c] = syn_nbs[a]->pos[b]; |
| bests[c] = target_sums[syn_nbs[a]->index[b]]; |
| break; |
| } |
| } |
| } |
| } |
| } |
| array = newAV(); |
| for (a = 0; a < MAX_NEIGHBOURS && besti[a] >= 0; a++) { |
| HV* hash = newHV(); |
| SV* word = newSVpvf(&vocab[besti[a] * max_w], 0); |
| if(latin_enc == 0) SvUTF8_on(word); |
| hv_store(hash, "word", strlen("word"), word , 0); |
| hv_store(hash, "dist", strlen("dist"), newSVnv(bestd[a]), 0); |
| hv_store(hash, "norm", strlen("norm"), newSVnv(bestn[a]), 0); |
| hv_store(hash, "sum", strlen("sum"), newSVnv(target_sums[besti[a]]), 0); |
| hv_store(hash, "pos", strlen("pos"), newSVnv(bestp[a]), 0); |
| av_push(array, newRV_noinc((SV*)hash)); |
| } |
| hv_store(result, "syntagmatic", strlen("syntagmatic"), newRV_noinc((SV*)array), 0); |
| } |
| end: |
| words = old_words; |
| return newRV_noinc((SV*)result); |
| } |
| |
| int dump_vecs(char *fname) { |
| long i, j; |
| FILE *f; |
| /* if(words>200000) */ |
| /* words=200000; */ |
| |
| if((f=fopen(fname, "w")) == NULL) { |
| fprintf(stderr, "cannot open %s for writing\n", fname); |
| return(-1); |
| } |
| fprintf(f, "%lld %lld\n", words, size); |
| for (i=0; i < words; i++) { |
| fprintf(f, "%s ", &vocab[i * max_w]); |
| for(j=0; j < size - 1; j++) |
| fprintf(f, "%f ", M[i*size + j]); |
| fprintf(f, "%f\n", M[i*size + j]); |
| } |
| fclose(f); |
| return(0); |
| } |
| |