multiclass.c

来自「卡内基梅隆大学MaCallum开发的文本分类系统」· C语言 代码 · 共 1,954 行 · 第 1/4 页

     estimate. */
  for (cisi = 0; cisi < cis_size + 2; cisi++)
    mixture[cisi] = 1.0;
  normalizer = (cis_size + 2) * 1.0;
  for (cmi = 0; cmi < cm_length; cmi++)
    {
      for (cisi = cisi2 = 0; cisi2 < cm[cmi]->cis_size; cisi2++)
	{
	  while (cisi < cisi2)
	    cisi++;
	  if (cm[cmi]->cis[cisi2] == cis[cisi])
	    {
	      mixture[cisi] += cm[cmi]->cis[cisi2];
	      normalizer += mixture[cisi];
	    }
	}
    }
  for (cisi = 0; cisi < cis_size; cisi++)
    mixture[cisi] /= normalizer;
  return;
#endif

#if 0
  /* Another (unused) option is to use a completely factored 
     representation */
  /* Calculcate normalized mixture weights just from the treenode priors,
     i.e., not using the CMIXTURE.  These may not actually get used.  */
  /* Plus one for the root, plus one for the uniform */
  double mixture_prior_sum;
  mixture_weights = alloca ((cis_size + 1 + 1) * sizeof (double));
  mixture_prior_sum = 0;
  for (cisi = 0; cisi < cis_size; cisi++)
    {
      assert (cis[cisi] >= 0);
      mixture_prior_sum += crossbow_root->children[cis[cisi]]->prior;
    }
  mixture_prior_sum += crossbow_root->prior + multiclass_uniform_prior;
  for (cisi = 0; cisi < cis_size; cisi++)
    if (cis[cisi] >= 0)
      mixture_weights[cisi] = 
	crossbow_root->children[cis[cisi]]->prior / mixture_prior_sum;
  mixture_weights[cis_size] = crossbow_root->prior / mixture_prior_sum;
  mixture_weights[cis_size+1] = multiclass_uniform_prior / mixture_prior_sum;
#endif
}

/* MIXTURE must be as large as CIS_SIZE+2 */
void
multiclass_mixture_given_doc_and_cis (crossbow_doc *doc, 
				      int *cis, int cis_size,
				      double *mixture)
{
  bow_wv *wv;
  double *cis_mixture;
  double mixture_sum;
  treenode *node;
  int cisi, wvi;
  int num_nodes;
  double *node_data_prob;
  double node_data_prob_sum;
  double *node_membership;

  wv = crossbow_wv_at_di (doc->di);
  cis_mixture = alloca (sizeof (double) * (cis_size + 2));
  multiclass_mixture_given_cis (cis, cis_size, cis_mixture);

  num_nodes = crossbow_root->children_count + 1 + 1;
  node_membership = alloca (num_nodes * sizeof (double));
  node_data_prob = alloca (num_nodes * sizeof (double));

  for (cisi = 0; cisi <= cis_size+1; cisi++)
    mixture[cisi] = 0;
  mixture_sum = 0;

  for (wvi = 0; wvi < wv->num_entries; wvi++)
    {
      /* Analagous to the per-word E-step */
      node_data_prob_sum = 0;
      for (cisi = 0; cisi <= cis_size; cisi++)
	{
	  if (cisi == cis_size)
	    node = crossbow_root;
	  else
	    node = crossbow_root->children[cis[cisi]];
	  node_data_prob[cisi] = cis_mixture[cisi] *
	    bow_treenode_pr_wi_loo_local(node,wv->entry[wvi].wi,doc->di,wvi);
	  assert (node_data_prob[cisi] >= 0);
	  node_data_prob_sum += node_data_prob[cisi];
	}
      /* For the uniform distribution */
      node_data_prob[cis_size+1] = cis_mixture[cis_size+1] *
	(1.0 / bow_num_words ());
      assert (node_data_prob[cis_size+1] >= 0);
      node_data_prob_sum += node_data_prob[cis_size+1];
      assert (node_data_prob_sum != 0);

      /* Normalize the node data probs, so they are membership
	 probabilities. */
      for (cisi = 0; cisi <= cis_size+1; cisi++)
	node_membership[cisi] = 
	  node_data_prob[cisi] / node_data_prob_sum;

      /* Analagous to the per-word M-step */
      for (cisi = 0; cisi <= cis_size+1; cisi++)
	{
	  mixture[cisi] += wv->entry[wvi].count * node_membership[cisi];
	  mixture_sum += mixture[cisi];
	}
    }

  /* Normalize the mixture to be returned */
  for (cisi = 0; cisi <= cis_size+1; cisi++)
    mixture[cisi] /= mixture_sum;
}

/* MIXTURE must be as large as CIS_SIZE+2 */
void
multiclass_iterated_mixture_given_doc_and_cis (crossbow_doc *doc, 
					       int *cis, int cis_size,
					       double *mixture)
{
  bow_wv *wv;
  double *cis_mixture;
  double mixture_sum;
  treenode *node;
  int cisi, wvi;
  int num_nodes;
  double *node_data_prob;
  double node_data_prob_sum;
  double *node_membership;
  //double pp, old_pp;

  wv = crossbow_wv_at_di (doc->di);
  cis_mixture = alloca (sizeof (double) * (cis_size + 2));
  multiclass_mixture_given_cis (cis, cis_size, cis_mixture);

  num_nodes = crossbow_root->children_count + 1 + 1;
  node_membership = alloca (num_nodes * sizeof (double));
  node_data_prob = alloca (num_nodes * sizeof (double));

  for (cisi = 0; cisi <= cis_size+1; cisi++)
    mixture[cisi] = 0;
  mixture_sum = 0;

  for (wvi = 0; wvi < wv->num_entries; wvi++)
    {
      /* Analagous to the per-word E-step */
      node_data_prob_sum = 0;
      for (cisi = 0; cisi <= cis_size; cisi++)
	{
	  if (cisi == cis_size)
	    node = crossbow_root;
	  else
	    node = crossbow_root->children[cis[cisi]];
	  node_data_prob[cisi] = cis_mixture[cisi] *
	    bow_treenode_pr_wi_loo_local(node,wv->entry[wvi].wi,doc->di,wvi);
	  assert (node_data_prob[cisi] >= 0);
	  node_data_prob_sum += node_data_prob[cisi];
	}
      /* For the uniform distribution */
      node_data_prob[cis_size+1] = cis_mixture[cis_size+1] *
	(1.0 / bow_num_words ());
      assert (node_data_prob[cis_size+1] >= 0);
      node_data_prob_sum += node_data_prob[cis_size+1];
      assert (node_data_prob_sum != 0);

      /* Normalize the node data probs, so they are membership
	 probabilities. */
      for (cisi = 0; cisi <= cis_size+1; cisi++)
	node_membership[cisi] = 
	  node_data_prob[cisi] / node_data_prob_sum;

      /* Analagous to the per-word M-step */
      for (cisi = 0; cisi <= cis_size+1; cisi++)
	{
	  mixture[cisi] += wv->entry[wvi].count * node_membership[cisi];
	  mixture_sum += mixture[cisi];
	}
    }

  /* Normalize the mixture to be returned */
  for (cisi = 0; cisi <= cis_size+1; cisi++)
    mixture[cisi] /= mixture_sum;
}


/* MIXTURE must be as large as CROSSBOW_ROOT->CHILDREN_COUNT+2 */
void
multiclass_mixture_given_doc (crossbow_doc *doc, 
			      double *mixture)
{
  int mixture_count = crossbow_root->children_count + 2;
  bow_wv *wv;
  double mixture_sum;
  treenode *node;
  int mi, wvi;
  double node_membership_sum;
  double *node_membership;

  wv = crossbow_wv_at_di (doc->di);
  node_membership = alloca (mixture_count * sizeof (double));

  for (mi = 0; mi < mixture_count; mi++)
    mixture[mi] = 0;
  mixture_sum = 0;

  for (wvi = 0; wvi < wv->num_entries; wvi++)
    {
      /* Analagous to the per-word E-step */
      node_membership_sum = 0;
      for (mi = 0; mi <= mixture_count-2; mi++)
	{
	  if (mi == mixture_count-2)
	    node = crossbow_root;
	  else
	    node = crossbow_root->children[mi];
	  if (doc->tag == bow_doc_train || doc->tag == bow_doc_unlabeled)
	    node_membership[mi] = 
	      bow_treenode_pr_wi_loo_local (node,wv->entry[wvi].wi,
					    doc->di,wvi);
	  else
	    node_membership[mi] = node->words[wv->entry[wvi].wi];
	  assert (node_membership[mi] >= 0);
	  node_membership_sum += node_membership[mi];
	}
      /* For the uniform distribution */
      node_membership[mixture_count-1] = 1.0 / bow_num_words ();
      node_membership_sum += node_membership[mixture_count-1];
      assert (node_membership_sum != 0);

      /* Normalize the node data probs, so they are membership
	 probabilities. */
      for (mi = 0; mi < mixture_count; mi++)
	node_membership[mi] = node_membership[mi] / node_membership_sum;

      /* Analagous to the per-word M-step */
      for (mi = 0; mi < mixture_count; mi++)
	{
	  mixture[mi] += wv->entry[wvi].count * node_membership[mi];
	  mixture_sum += mixture[mi];
	}
    }

  /* Normalize the mixture to be returned */
  assert (mixture_sum);
  for (mi = 0; mi < mixture_count; mi++)
    {
      mixture[mi] /= mixture_sum;
      //assert (mixture[mi] > 0);
    }
}

/* Return the most likely mixture over mixture components, assuming
   that we are already committed to including the classes in CIS, and
   that we probabilistically remove the words that they account for.
   MIXTURE must be as large as CROSSBOW_ROOT->CHILDREN_COUNT+2 */
void
multiclass_mixture_given_doc_and_partial_cis (crossbow_doc *doc, 
					      const int *cis, int cis_size,
					      const int *exclude_cis, 
					      int exclude_cis_size,
					      double *mixture)
{
  int mixture_count = crossbow_root->children_count + 2;
  bow_wv *wv;
  double mixture_sum;
  treenode *node;
  int mi, wvi, cisi;
  double node_membership_sum;
  double *node_membership;
  double *node_word_prob;
  double average_word_prob_cis, incr;

  wv = crossbow_wv_at_di (doc->di);
  node_membership = alloca (mixture_count * sizeof (double));
  node_word_prob = alloca (mixture_count * sizeof (double));

  for (mi = 0; mi < mixture_count; mi++)
    mixture[mi] = 0;
  mixture_sum = 0;

  for (wvi = 0; wvi < wv->num_entries; wvi++)
    {
      /* Analagous to the per-word E-step */
      node_membership_sum = 0;
      for (mi = 0; mi <= mixture_count-2; mi++)
	{
	  if (mi == mixture_count-2)
	    node = crossbow_root;
	  else
	    node = crossbow_root->children[mi];
	  node_word_prob[mi] = 
	    bow_treenode_pr_wi_loo_local (node,wv->entry[wvi].wi,doc->di,wvi);
	  node_membership[mi] = node_word_prob[mi];
	  assert (node_membership[mi] >= 0);
	}
      /* For the uniform distribution */
      node_membership[mixture_count-1] = 1.0 / bow_num_words ();

      /* Calculate the average word probability of the classes
	 explicitly included with CIS, and the always-included root
	 and uniform distribution.  Zero the mixture probabilities for
	 those mixtures. */
      average_word_prob_cis = 0;
      for (cisi = 0; cisi < cis_size; cisi++)
	{
	  average_word_prob_cis += node_membership[cis[cisi]];
	  node_membership[cis[cisi]] = 0;
	}
      average_word_prob_cis += node_membership[mixture_count-2];
      node_membership[mixture_count-2] = 0;
      average_word_prob_cis += node_membership[mixture_count-1];
      node_membership[mixture_count-1] = 0;
      average_word_prob_cis /= cis_size + 2;

      /* Zero the probabilities of the classes explicitly excluded */
      for (cisi = 0; cisi < exclude_cis_size; cisi++)
	node_membership[exclude_cis[cisi]] = 0;

      /* Subtract the average  */
      for (mi = 0; mi < mixture_count; mi++)
	{
	  node_membership[mi] -= average_word_prob_cis;
	  if (node_membership[mi] < 0) 
	    node_membership[mi] = 0;
	  node_membership_sum += node_membership[mi];
	}

#if 1
      /* If any of the NODE_MEMBERSHIP's are non-zero, normalize the
	 node data probs, so they are membership probabilities. */
      if (node_membership_sum != 0)
	for (mi = 0; mi < mixture_count; mi++)
	  node_membership[mi] = node_membership[mi] / node_membership_sum;
#endif

      /* Analagous to the per-word M-step */
      for (mi = 0; mi < mixture_count; mi++)
	{
	  if (node_membership[mi] == 0)
	    continue;
	  incr= (wv->entry[wvi].count * node_membership[mi]
		 * log (node_word_prob[mi]/average_word_prob_cis));
	  assert (incr >= 0);
	  mixture[mi] += incr;
	  mixture_sum += mixture[mi];
	}
    }

  /* Normalize the mixture to be returned */
  for (mi = 0; mi < mixture_count; mi++)
    mixture[mi] /= mixture_sum;
}


/* Return the perplexity */
double
multiclass_em_one_iteration ()
{
  int di;
  crossbow_doc *doc;
  bow_wv *wv;
  treenode *node;
  int cisi, wvi;
  int num_nodes;
  double *node_word_prob, log_prob_of_data2;
  double node_membership_sum, word_prob, log_prob_of_data, deposit;
  int num_data_words = 0;	/* the number of word occurrences */
  double *node_membership;
  cmixture *m;
  int cis_size;
  double *mixture_all;

  /* One node for each topic, plus one for all-english, plus one for uniform */
  num_nodes = crossbow_root->children_count + 1 + 1;
  node_membership = alloca (num_nodes * sizeof (double));
  node_word_prob = alloca (num_nodes * sizeof (double));
  mixture_all = alloca ((crossbow_root->children_count+2) * sizeof(double));

  log_prob_of_data = log_prob_of_data2 = 0;
  for (di = 0; di < crossbow_docs->length; di++)
    {
      doc = bow_array_entry_at_index (crossbow_docs, di);
      if (doc->tag != bow_doc_train && doc->tag != bow_doc_unlabeled)
	continue;
      /* Temporary fix */
      if (strstr (doc->filename, ".include")
	  || strstr (doc->filename, ".exclude"))
	continue;

      multiclass_mixture_given_doc (doc, mixture_all);
      bow_verbosify (bow_verbose, "%s ", doc->filename);
      for (cisi = 0; cisi < crossbow_root->children_count+2; cisi++)
	{
	  bow_verbosify (bow_verbose, "%s=%g,",
			 (cisi < crossbow_root->children_count
			  ? bow_int2str (crossbow_classnames, cisi)
			  : (cisi == crossbow_root->children_count
			     ? "root"
			     : "uniform")),
			 mixture_all[cisi]);
	}
      bow_verbosify (bow_verbose, "\n");


      /* Get the word vector for this document, and for each word,
         estimate its membership probability in each of its classes
         (and the root class), and then gather stats for the M-step */
      wv = crossbow_wv_at_di (di);
      m = cmixture_for_cis (doc->cis, doc->cis_size, 0, &cis_size);
      assert (m);
      assert (m->doc_count > 0);
      /* Zero the document-specific mixture in preparation for incrementing */
      for (cisi = 0; cisi < cis_size + 2; cisi++)
	doc->cis_mixture[cisi] = 0;
      for (wvi = 0; wvi < wv->num_entries; wvi++)
	{
	  num_data_words += wv->entry[wvi].count;

	  /* Per-word E-step */
	  node_membership_sum = 0;
	  for (cisi = 0; cisi <= doc->cis_size; cisi++)
	    {
	      if (cisi == doc->cis_size)
		node = crossbow_root;
	      else
		node = crossbow_root->children[doc->cis[cisi]];
	      node_word_prob[cisi] = 
		bow_treenode_pr_wi_loo_local (node, wv->entry[wvi].wi,
					      di, wvi);
	      node_membership[cisi] = m->m[cisi] * node_word_prob[cisi];
	      assert (node_word_prob[cisi] >= 0);
	      node_membership_sum += node_membership[cisi];
	    }
	  /* For the uniform distribution */
	  node_word_prob[doc->cis_size+1] = (1.0 / bow_num_words ());
	  node_membership[doc->cis_size+1] = m->m[doc->cis_size+1] *
	    node_word_prob[doc->cis_size+1];
	  node_membership_sum += node_membership[doc->cis_size+1];
	  assert (node_membership_sum != 0);

	  /* Normalize the node membership probs.  Also increment
             perplexity */
	  word_prob = 0;
	  for (cisi = 0; cisi <= doc->cis_size+1; cisi++)
	    {
	      node_membership[cisi] /= node_membership_sum;
	      word_prob += node_membership[cisi] * node_word_prob[cisi];
	      if (node_membership[cisi])
		log_prob_of_data2 += (node_membership[cisi]
				      * wv->entry[wvi].count
				      * log (node_word_prob[cisi]));
	    }
	  log_prob_of_data += wv->entry[wvi].count * log (word_prob);

	  /* Per-word M-step */
	  for (cisi = 0; cisi <= doc->cis_size; cisi++)
	    {
	      if (cisi == doc->cis_size)
		node = crossbow_root;
	      else
		node = crossbow_root->children[doc->cis[cisi]];
	      deposit = wv->entry[wvi].count * node_membership[cisi];
	      node->new_words[wv->entry[wvi].wi] += deposit;
	      bow_treenode_add_new_loo_for_di_wvi
		(node, deposit, di, wvi, 
		 wv->num_entries, crossbow_docs->length);

	      /* For non-combo version */
	      node->new_prior += deposit;
	      /* For combo version */
	      m->new_m[cisi] += deposit;
	      doc->cis_mixture[cisi] += deposit;
	    }
	  /* For the uniform distribution */
	  deposit = wv->entry[wvi].count * node_membership[doc->cis_size+1];
	  multiclass_uniform_new_prior += deposit;
	  m->new_m[doc->cis_size+1] += deposit;
	  doc->cis_mixture[cis_size+1] += deposit;
	}

      /* Normalize the document-specific CIS_MIXTURE, (and print it out) */
      {
	double max = -FLT_MAX;
	double cis_mixture_sum;
	for (cisi = 0; cisi < cis_size+2; cisi++)
	  if (doc->cis_mixture[cisi] > max)
	    max = doc->cis_mixture[cisi];
	cis_mixture_sum = 0;