w2v-server.pl

#!/usr/local/bin/perl
use Inline C;
#use Inline C => Config => BUILD_NOISY => 1, CFLAGS => $Config{cflags}." -O4 -mtune k9";
use Inline C => Config => CLEAN_AFTER_BUILD => 0, ccflags => $Config{ccflags}." -O4";
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;
plugin 'Log::Access';

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;

getopts('d:Gil:p:m: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 '/' => sub {
  my $c    = shift;
	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 $res;
	my @lists;
	my @collocations;
	if(defined($word) && $word !~ /^\s*$/) {
		$c->inactivity_timeout(300);
		$word =~ s/\s+/ /g;
    for my $w (split(' *\| *', $word)) {
			$c->app->log->debug('Looking for neighbours of '.$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);
      }
      push(@lists, $res->{paradigmatic});
		}
	}
	$word =~ s/ *\| */ | /g;
	$c->render(template=>"index", word=>$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;
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;

	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;
  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 & 32) == 0 && (previous & 32) == 32) || 
         previous == '-' && (c & 32) ||
         c == 'ƒ'
    ) {
        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 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;
	
	slice = words / para_threads;
  
  wl = getTargetWords(st1, search_backw);
  if(wl->length < 1)
    goto end;

	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], &para_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; 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();
  for (a = 0; a < N; 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, "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));
  }
  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:
	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);
}
__DATA__

@@ index.html.ep
<!DOCTYPE html>
<html>
<head>
	<title>DeReKo-Word-Vector-Distances</title>
	<link rel="stylesheet" href="//code.jquery.com/ui/1.11.4/themes/smoothness/jquery-ui.css">
	<script src="http://code.jquery.com/jquery-latest.min.js"></script>
  <script src="//code.jquery.com/ui/1.11.4/jquery-ui.js"></script>
	<script>
 $(function() {
    $( document ).tooltip({
			content: function() {
				return $(this).attr('title');
			}}
		)
	 })
  </script>
	<script src="//d3js.org/d3.v3.min.js" charset="utf-8"></script>
	<script src="http://klinux10/word2vec/tsne.js"></script>
	<script src="http://klinux10/word2vec/som.js"></script>
	<script src="http://klinux10/word2vec/labeler.js"></script>
<style>
body, input {
	font-family: Arial, sans-serif;
	font-size: 11pt;
}

.ui-tooltip-content {
	font-size: 9pt;
	color: #222222;
}

svg > .ui-tooltip-content {
	font-size: 8pt;
	color: #222222;
}

a.merged {
  color: green;
  fill: green;
}

#first a {
  text-decoration: none;
}

a.marked, #first a.marked {
  text-decoration: underline;
}

a.target {
  color: red;
  fill: red;
}

#collocators {
    margin-bottom: 15px;
}
  
#wrapper {
    width: 100%;
//   border: 1px solid red; 
    overflow: hidden; /* will contain if #first is longer than #second */
}
#first {
    margin-right: 20px;
    float: left;
 //    border: 1px solid green;
}
#second {
    border: 1px solid #333;
    overflow: hidden; /* if you don't want #second to wrap below #first */
}
#som2 svg {
    border: 1px solid #333;
}

#cost {
	  font-size: 8pt;
	  color: #222222;
    margin-top: 4px;
    margin-bottom: 12px;
}

#sominfo1, #sominfo {
	  font-size: 8pt;
	  color: #222222;
    margin-top: 0px;
}

#somcolor1, #somcolor2, #somcolor3 {
  display: inline-block;
  height: 10px;
  width: 10px;
}

#third {
  border: 1px solid #333;
}

</style>
<script>

var opt = {epsilon: <%= $epsilon %>, perplexity: <%= $perplexity %>},
		mapWidth = 800, // width map
		mapHeight = 800,
		jitterRadius = 7;

var T = new tsnejs.tSNE(opt); // create a tSNE instance

var Y;

var data;
var labeler;


function applyJitter() {
  svg.selectAll('.tsnet')
    .data(labels)
    .transition()
    .duration(50)
    .attr("transform", function(d, i) {
      T.Y[i][0] = (d.x - mapWidth/2 - tx)/ss/20;
      T.Y[i][1] = (d.y - mapHeight/2 - ty)/ss/20;
      return "translate(" +
        (d.x) + "," +
        (d.y) + ")";
    });
}
    
function updateEmbedding() {
  var Y = T.getSolution();
  svg.selectAll('.tsnet')
     .data(data.words)
     .attr("transform", function(d, i) { 
			 return "translate(" +
         ((Y[i][0]*20*ss + tx) + mapWidth/2) + "," +
         ((Y[i][1]*20*ss + ty) + mapHeight/2) + ")"; });
}

var svg;
var labels = [];
var anchor_array = [];
var text;

function drawEmbedding() {
  $("#embed").empty();
  var div = d3.select("#embed");
	
  // get min and max in each column of Y
  var Y = T.Y;
  
  svg = div.append("svg") // svg is global
					 .attr("width", mapWidth)
					 .attr("height", mapHeight);
	
  var g = svg.selectAll(".b")
						 .data(data.words)
						 .enter().append("g")
						 .attr("class", "tsnet");
	
  g.append("a")
	 .attr("xlink:href", function(word) {return "/?word="+word;})
   .attr("class", function(d, i) {
     var res="";
     if(data.marked[i]) {
			  res="marked ";
      }
			if(data.target.indexOf(" "+d+" ") >= 0) {
				return res+"target";
			} else if(data.ranks[i] < data.mergedEnd) {
			  return res+"merged";
			} else {
				return res;
			}
	 })
	 .attr("title", function(d, i) {
		 if(data.mergedEnd > 0) {
       if(data.ranks[i] >= data.mergedEnd) {	
         	return "rank: "+i +"  "+"freq. rank: "+(data.ranks[i]).toString().replace(/\B(?=(\d{3})+(?!\d))/g, ",");
       } else {
         	return "rank: "+i +"  "+"freq. rank: "+data.ranks[i].toString().replace(/\B(?=(\d{3})+(?!\d))/g, ",") + " (merged vocab)";
       }  
	   } else {
		   return "rank: "+i +"  "+"freq. rank: "+data.ranks[i].toString().replace(/\B(?=(\d{3})+(?!\d))/g, ",");
		 }
	 })
		.append("text")
   .attr("text-anchor", "top")
   .attr("font-size", 12)
   .text(function(d) { return d; });
	
  var zoomListener = d3.behavior.zoom()
											 .scaleExtent([0.1, 10])
											 .center([0,0])
											 .on("zoom", zoomHandler);
  zoomListener(svg);
}

var tx=0, ty=0;
var ss=1;
var iter_id=-1;

function zoomHandler() {
  tx = d3.event.translate[0];
  ty = d3.event.translate[1];
  ss = d3.event.scale;
  updateEmbedding();
}

var stepnum = 0;

function stopStep() {
  clearInterval(iter_id);
	text = svg.selectAll("text");
	
  // jitter function needs different data and co-ordinate representation
	labels = d3.range(data.words.length).map(function(i) {
    var x = (T.Y[i][0]*20*ss + tx) + mapWidth/2;
    var y = (T.Y[i][1]*20*ss + ty) + mapHeight/2;
		anchor_array.push({x: x, y: y, r: jitterRadius});
		return {
      x: x,
      y: y,
			name: data.words[i]
		};
	});
  
  // get the actual label bounding boxes for the jitter function
	var index = 0;
	text.each(function() {
		labels[index].width = this.getBBox().width;
		labels[index].height = this.getBBox().height;
		index += 1;
	});

	 
//	 setTimeout(updateEmbedding, 1);
//	 setTimeout(
	labeler =	 d3.labeler()
			         .label(labels)
			         .anchor(anchor_array)
			         .width(mapWidth)
			         .height(mapHeight)
			         .update(applyJitter);
  //         .start(1000);

  iter_id = setInterval(jitterStep, 1);
}

var jitter_i=0;

function jitterStep() {
  if(jitter_i++ > 100) {
    clearInterval(iter_id);
  } else {
		labeler.start2(10);
	  applyJitter();
  }
}

var last_cost=1000;

function step() {
  var i = T.iter;
  
  if(i > <%= $no_iterations %>) {
		stopStep();
  } else {
    var cost = Math.round(T.step() * 100000) / 100000; // do a few steps
    $("#cost").html("tsne iteration " + i + ", cost: " + cost.toFixed(5));
    if(i % 250 == 0 && cost >= last_cost) {
		  stopStep();
    } else {
      last_cost = cost;
      updateEmbedding();
    }
  }
}

function showMap(j) {
	data=j;
  T.iter=0;
  T.initDataRaw(data.vecs); // init embedding
  drawEmbedding(); // draw initial embedding
	
	if(iter_id >= 0) {
		clearInterval(iter_id);
	}
  //T.debugGrad();
  iter_id = setInterval(step, 1);
  if(<%= $show_som %>) {
    makeSOM(j, <%= $no_iterations %>);
  }
}
 
</script>
</head>
<body>
	<form action="<%=url_for('/')->to_abs%>" method="GET">
		word(s): 
    <input type="text" name="word" size="20"  value="<%= $word %>" title="When looking for multiple words use spaces as separators to search around the average vector and | as separator to get the neighbours for each word."> 
		% if($mergedEnd > 0) {
     backw. <input type="checkbox" name="sbf" value="1" <%= ($searchBaseVocabFirst ? "checked" : "") %> title="If checkecked base vocabulary will be searched first. Otherwise merged vocabulray will be searched first.">
    % }
		max. neighbours: <input type="text" size="8" name="n" value="<%= $no_nbs %>">
		max. iterations: <input type="text" name="N" size="8" value="<%= $no_iterations %>">
    SOM <input type="checkbox" name="som" value="1" <%= ($show_som ? "checked" : "") %>>
		% if($collocators) {
    <span>  </span>sort collocators by
    <select name="sort">
      <option value="0"  <%= ($sort!=1? "selected":"") %>>responsiveness</option>
      <option value="1" <%= ($sort==1? "selected":"") %>>mean p</option>
    </select>
		% }
		<span>  </span><input type="submit" value="Show">
	</form>
	<br>
	% if($lists) {
	<div id="wrapper">
		<table id="first">
			<tr>
				<th align="right">#</th><th align="right">cos</th><th align="left">paradigmatic</th>
        % if($collocators) {
				<th title="Position in winodw around target word. Absolute value can be too low because of sub-sampling frequent words.">@</th><th align="right" title="&#34;Responsivenes&#34; of the collocator at the relative position @. Approximation of the probability that the combination of the target word and the collocator at the relative position @ come from the corpus.">resp.</th><th title="Probability of the collocator at window location @."align="right">p(c<sub><small>@</small></sub>)</th><th align="right">Σp(c<sub><small>@</small></sub>)/|w|</th><th align="left">syntagmatic</th>
				% }
			</tr>
			% my $j=0; my @words; my @vecs; my @ranks; my @marked; for my $list (@$lists) {
			% my $i=0; while($list) {
      % my $item = (@$list)[$i];
      % my $c = ($collocators? (@$collocators)[$i] : 0);
      % last if(!$c && !$item);
			<tr>
				<td align="right">
  				<%= ++$i %>.
				</td>
        % if($item) {
			  % if(!grep{$_ eq $item->{word}} @words) {
        %   push @vecs, $item->{vector};
			  %   push @words, $item->{word};
			  %   push @ranks, $item->{rank};
			  %   push @marked, ($marked->{$item->{word}}? 1 : 0);
        % }
				<td align="right">
  				<%= sprintf("%.3f", $item->{dist}) %>
				</td>
				<td>
          % my $class = ($marked->{$item->{word}}? "marked " : ""); 
          % my $r = $item->{rank}; 
          % if($r < $mergedEnd) {
          %   $class .= "merged";
          %   $r .= " (merged vocab)";
          % } elsif($mergedEnd!=0 && $r > $mergedEnd) {
          %   $r -= $mergedEnd;
          % }
  				<a class="<%= $class %>" title="freq. rank: <%= $r %>" href="/?word=<%= $item->{word} %>"><%= $item->{word} %></a>
				</td>
        % } else {
          <td colspan="2"/>
        % }
        % if($c) {
				<td align="right">
          <%= $c->{pos} %>:
				</td>
				<td align="right">
  				<%= sprintf("%.3f", $c->{dist}) %>
				</td>
				<td align="right">
  				<%= sprintf("%.3e", $c->{norm}) %>
				</td>
				<td align="right">
  				<%= sprintf("%.3e", $c->{sum}) %>
				</td>
				<td align="left">
  				<a href="/?word=<%= $c->{word} %>">
            <%= $c->{word} %>
				</td>
        % } else {
        <td colspan="5"/>
        % }
			</tr>
			% }
			% }
		</table>
		<script>
		 % use Mojo::ByteStream 'b';
		 $(window).load(function() {
			 showMap(<%= b(Mojo::JSON::to_json({target => " $word ", mergedEnd=> $mergedEnd, words => \@words, vecs => \@vecs, ranks => \@ranks, marked => \@marked})); %>);
		 });
    </script>
		% }
		<div id="second" style="width:800px; height:800px; font-family: arial;">
			<div id="embed">
			</div>
		</div>
    <div id="cost"></div>
		% if($show_som) {
		<div id="som2">
		</div>
    <div id="sominfo1"><span id="somcolor1">   </span> <span id="somword1"> </span> <span id="somcolor2">   </span> <span id="somword2"> </span> <span id="somcolor3">   </span></div>
    <div id="sominfo">SOM iteration <span id="iterations">0</span></div>
    % }
  </div>
  % if($training_args) {
    <p>
      Word vector model trained with <a href="https://code.google.com/p/word2vec/">word2vec</a> using the following parameters: <pre><%= $training_args %></pre>
    </p>
  % }
  </body>
</html>