submine 0.1.1__cp312-cp312-macosx_11_0_arm64.whl

submine/backends/gspan/gspan.cpp

@@ -0,0 +1,776 @@
+/*
+    $Id: gspan.cpp,v 1.8 2004/05/21 09:27:17 taku-ku Exp $;
+ 
+   Copyright (C) 2004 Taku Kudo, All rights reserved.
+     This is free software with ABSOLUTELY NO WARRANTY.
+  
+   This program is free software; you can redistribute it and/or modify
+     it under the terms of the GNU General Public License as published by
+     the Free Software Foundation; either version 2 of the License, or
+     (at your option) any later version.
+    
+   This program is distributed in the hope that it will be useful,
+     but WITHOUT ANY WARRANTY; without even the implied warranty of
+     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+     GNU General Public License for more details.
+    
+   You should have received a copy of the GNU General Public License
+     along with this program; if not, write to the Free Software
+     Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
+     02111-1307, USA
+*/
+#include "gspan.h"
+#include <iterator>
+
+#include <cstdio>
+#include <stdexcept>
+
+#include <stdlib.h>
+#include <thread>
+#include <sstream>
+
+
+
+namespace GSPAN {
+
+gSpan::gSpan (void)
+{
+	boost = false;
+}
+
+// MATLAB/MEX support removed.
+
+std::istream &gSpan::read (std::istream &is)
+{
+	Graph g(directed);
+	while (true) {
+		g.read (is);
+		if (g.empty()) break;
+		TRANS.push_back (g);
+	}
+	return is;
+}
+
+
+/* 2-class LPBoosting gain function
+ */
+double
+gSpan::gain2 (Projected &projected, double y)
+{
+	/* Compute actual gain (eq. (10) in [Dimiriz2002]) and Problem 1 in graph
+	 * boost paper.
+	 *
+	 * For L samples x_i, calculate
+	 *   gain = \sum_{i=1}^{L} d_i y_i h(x_i)
+	 * where h is implicitly defined by this current subgraph pattern.
+	 */
+	unsigned int oid = 0xffffffff;
+	double gainm = 0.0;
+
+	Projected::iterator cur = projected.begin();
+	for ( ; cur != projected.end() ; ++cur) {
+		if (oid != cur->id) {
+			/* A positive pattern: it appears in the graph.
+			 *
+			 * h_{<t,y>} = y, if t \subseteq x,
+			 *           = -y, otherwise.
+			 *
+			 * t is implicitly defined by projected.
+			 * First, process all skipped graphs.
+			 */
+			for (unsigned int skipn = oid+1 ; skipn < cur->id ; ++skipn)
+				gainm += boostY[skipn]*boostWeights[skipn]*(-y);
+
+			/* Now process positive pattern.
+			 * This is also the gain function for 1.5 LP boosting, as we have
+			 * the constraint
+			 *   \sum_{n=1}^N \lambda_n h_j(x_{1,n})
+			 *   - \sum_{m=1}^M \mu_m h_j(x_{2,m}) <= \gamma.
+			 * In order to maximize the violation of the current ensemble, we
+			 * search for the classifier that maximizes the left side.
+			 */
+			gainm += boostY[cur->id]*boostWeights[cur->id]*y;
+		}
+		oid = cur->id;
+	}
+
+	/* All skipped graphs to the end.
+	 */
+	for (unsigned int skipn = oid+1 ; skipn < boostY.size() ; ++skipn)
+		gainm += boostY[skipn]*boostWeights[skipn]*(-y);
+
+	return (gainm);
+}
+
+
+/* 1.5-class LPBoosting gain function
+ */
+double
+gSpan::gain1d5 (Projected &projected, double y)
+{
+	/* Compute actual gain (eq. (10) in [Dimiriz2002]) and Problem 1 in graph
+	 * boost paper.
+	 *
+	 * For L samples x_i, calculate
+	 *   gain = \sum_{i=1}^{L} d_i y_i h(x_i)
+	 * where h is implicitly defined by this current subgraph pattern.  In the
+	 * 1.5-class case, h(x_i) is zero if the pattern does not appear, hence we
+	 * can skip those instances silently.
+	 */
+	unsigned int oid = 0xffffffff;
+	double gain = 0.0;
+
+	Projected::iterator cur = projected.begin();
+	for ( ; cur != projected.end() ; ++cur) {
+		if (oid != cur->id)
+			gain += boostY[cur->id]*boostWeights[cur->id]*y;
+
+		oid = cur->id;
+	}
+
+	return (gain);
+}
+
+/* Gain bound for 1.5-class formulation
+ */
+double
+gSpan::gainbound1d5 (Projected &projected)
+{
+	double gain = 0.0;
+	unsigned int oid = 0xffffffff;
+
+	for (Projected::iterator cur = projected.begin() ;
+		cur != projected.end() ; ++cur)
+	{
+		if (oid != cur->id) {
+			oid = cur->id;
+			if (boostY[cur->id] <= 0.0)
+				continue;
+
+			gain += boostWeights[cur->id];
+		}
+	}
+
+	return (gain);
+}
+
+/* Gain bound for 2-class formulation
+ */
+double
+gSpan::gainbound2 (Projected &projected)
+{
+	/* Lemma 1, section 3.2 in the graph boosting paper.  We automatically
+	 * only enumerate those patterns which contain the subgraph 'projected'.
+	 *
+	 * <t^,y^> = argmax_{t \in F,y \in {+/- 1}} d_i y_i h_{<t,y>}, where
+	 * F = \unionset_{i=1}^{L} { t | t \subseteq x_i }.
+	 */
+	unsigned int oid = 0xffffffff;
+	unsigned int size = 0;
+	double gain_pos = 0.0;
+	double gain_neg = 0.0;
+
+	for (Projected::iterator cur = projected.begin() ;
+		cur != projected.end() ; ++cur)
+	{
+		if (oid != cur->id) {
+			if (boostY[cur->id] <= 0.0)
+				gain_neg += boostWeights[cur->id];
+			else
+				gain_pos += boostWeights[cur->id];
+
+			++size;
+		}
+		oid = cur->id;
+	}
+
+	/* The boostWeightSum has been precomputed as it is always the same for
+	 * one run.
+	 */
+	gain_neg = 2.0*gain_neg + boostWeightSum;
+	gain_pos = 2.0*gain_pos - boostWeightSum;
+
+	if (gain_neg >= gain_pos)
+		return (gain_neg);
+
+	return (gain_pos);
+}
+
+
+std::map<unsigned int, unsigned int>
+gSpan::support_counts (Projected &projected)
+{
+	std::map<unsigned int, unsigned int> counts;
+
+	for (Projected::iterator cur = projected.begin() ;
+		cur != projected.end() ; ++cur)
+	{
+		counts[cur->id] += 1;
+	}
+
+	return (counts);
+}
+
+
+unsigned int
+gSpan::support (Projected &projected)
+{
+	unsigned int oid = 0xffffffff;
+	unsigned int size = 0;
+
+	for (Projected::iterator cur = projected.begin(); cur != projected.end(); ++cur) {
+		if (oid != cur->id) {
+			++size;
+		}
+		oid = cur->id;
+	}
+
+	return size;
+}
+
+void gSpan::report_boosting (Projected &projected, unsigned int sup,
+	double gain, double yval)
+{
+	if (maxpat_max > maxpat_min && DFS_CODE.nodeCount () > maxpat_max)
+		return;
+	if (maxpat_min > 0 && DFS_CODE.nodeCount () < maxpat_min)
+		return;
+
+	Graph g(directed);
+	DFS_CODE.toGraph (g);
+
+	// insert individual counts
+	std::map<unsigned int, unsigned int> GYcounts;
+	unsigned int oid = 0xffffffff;
+	for (Projected::iterator cur = projected.begin(); cur != projected.end(); ++cur) {
+		if (oid != cur->id)
+			GYcounts[cur->id] = 0;
+
+		GYcounts[cur->id] += 1;
+		oid = cur->id;
+	}
+
+	report_boosting_inter (g, sup, gain, yval, GYcounts);
+}
+
+void gSpan::report_boosting_inter (Graph &g, unsigned int sup,
+	double gain, double yval, std::map<unsigned int, unsigned int>& GYcounts)
+{
+	// Pure C++ build: no MATLAB-specific graph printer.
+
+	/* Insertion sort based on the gain values.
+	 */
+	unsigned int insertidx;
+	for (insertidx = 0 ; insertidx < bestGraphsGain.size () ; ++insertidx) {
+		if (gain >= bestGraphsGain[insertidx])
+			break;
+	}
+
+	bestGraphs.insert (bestGraphs.begin() + insertidx, g);
+	bestGraphsY.insert (bestGraphsY.begin() + insertidx, yval);
+	bestGraphsGain.insert (bestGraphsGain.begin() + insertidx, gain);
+	bestGraphsCounts.insert (bestGraphsCounts.begin() + insertidx, GYcounts);
+
+#ifdef DEBUG
+	fprintf(stderr, "   inserted good output graph, gain %lf, sup %d, Y %lf, new size %d\n",
+		gain, sup, yval, g.size());
+#endif
+
+	/* If the maximum allowed size is exceeded, prune it down.
+	 */
+	if (bestGraphsGain.size () > boostN) {
+#ifdef DEBUG
+		fprintf(stderr, "   one element removed at end with gain %lf\n",
+			bestGraphsGain[bestGraphsGain.size()-1]);
+#endif
+
+		bestGraphs.pop_back ();
+		bestGraphsY.pop_back ();
+		bestGraphsGain.pop_back ();
+		bestGraphsCounts.pop_back ();
+	}
+
+	++ID;
+}
+
+
+/* Special report function for single node graphs.
+ */
+void gSpan::report_single (Graph &g, std::map<unsigned int, unsigned int>& ncount)
+{
+	unsigned int sup = 0;
+	for (std::map<unsigned int, unsigned int>::iterator it = ncount.begin () ;
+		it != ncount.end () ; ++it)
+	{
+		sup += (*it).second;
+	}
+
+	if (maxpat_max > maxpat_min && g.size () > maxpat_max)
+		return;
+	if (maxpat_min > 0 && g.size () < maxpat_min)
+		return;
+
+	if (enc == false) {
+		if (where == false)
+			*os << "t # " << ID << " * " << sup;
+		*os << '\n';
+
+		g.write (*os);
+		if (callback_) {
+					std::cerr << "pattern n=" << g.size()
+			<< " e0deg=" << (g.size() ? g[0].edge.size() : 0)
+			<< " sup=" << sup << "\n";
+				callback_(g, sup, nullptr);
+			}
+
+		*os << '\n';
+	} else {
+		// Encoded output mode historically relied on MATLAB glue in this fork.
+		// If you need encoded output, implement it here in pure C++.
+		throw std::runtime_error("report_single: encoded output is not supported in the pure C++ build");
+	}
+	++ID;
+}
+
+void gSpan::report(Projected &projected, unsigned int sup)
+{
+    if (maxpat_max > maxpat_min && DFS_CODE.nodeCount() > maxpat_max) return;
+    if (maxpat_min > 0 && DFS_CODE.nodeCount() < maxpat_min) return;
+
+    if (where) {
+        *os << "<pattern>\n";
+        *os << "<id>" << ID << "</id>\n";
+        *os << "<support>" << sup << "</support>\n";
+        *os << "<what>";
+    }
+
+    // ALWAYS build Graph for callback (details)
+    Graph g(directed);
+    DFS_CODE.toGraph(g);
+
+    // Always call callback if present
+    if (callback_) {
+        callback_(g, sup, &projected);
+    }
+
+    // Output format to stream still depends on enc/where
+    if (!enc) {
+        if (!where) *os << "t # " << ID << " * " << sup;
+        *os << '\n';
+        g.write(*os);
+    } else {
+        if (!where) *os << '<' << ID << ">    " << sup << " [";
+        DFS_CODE.write(*os);
+        if (!where) *os << ']';
+    }
+
+    if (where) {
+        *os << "</what>\n<where>";
+        unsigned int oid = 0xffffffff;
+        for (Projected::iterator cur = projected.begin(); cur != projected.end(); ++cur) {
+            if (oid != cur->id) {
+                if (cur != projected.begin()) *os << ' ';
+                *os << cur->id;
+            }
+            oid = cur->id;
+        }
+        *os << "</where>\n</pattern>";
+    }
+
+    *os << '\n';
+    ++ID;
+}
+
+/* Recursive subgraph mining function (similar to subprocedure 1
+ * Subgraph_Mining in [Yan2002]).
+ */
+void gSpan::project (Projected &projected)
+{
+	if (boost && (boostmax > 0 && boostseen > boostmax))
+		return;
+
+	/* Check if the pattern is frequent enough.
+	 */
+	unsigned int sup = support (projected);
+	if (sup < minsup)
+		return;
+
+	/* The minimal DFS code check is more expensive than the support check,
+	 * hence it is done now, after checking the support.
+	 */
+	if (is_min () == false) {
+		//      *os  << "NOT MIN [";  DFS_CODE.write (*os);  *os << "]" << std::endl;
+		return;
+	}
+
+	if (boost) {
+		/* Compute the gains for all classifiers we can build out of the current
+		 * subgraph pattern (which is just two).
+		 */
+		double gain_ypos = 0.0;	// Gain if y=1 in the weak learner
+		double gain_yneg = 0.0;	// Gain if y=-1 in the weak learner
+		double gainmax = 0.0;	// Maximum of the negative/positive gain
+		double yval = 1.0;	// y that achieves the maximum gain
+
+		/* Compute gains.  This differs a little for the 1.5-class and the
+		 * 2-class case.  (The gain function is replaced, as well).
+		 */
+		gain_ypos = (this->*gain) (projected, 1.0);
+		if (boostType == 1) {
+			gainmax = gain_ypos;
+		} else if (boostType == 2) {
+			gain_yneg = (this->*gain) (projected, -1.0);
+
+			if (gain_ypos >= gain_yneg) {
+				gainmax = gain_ypos;
+			} else {
+				gainmax = gain_yneg;
+				yval = -1.0;
+			}
+		}
+
+		/* Check if we have to replace the least best of our top performers.
+		 */
+		if (gainmax > boostTau) {
+			/* Update best gain so far.  The best gain is also the tau bound.
+			 */
+#ifdef DEBUG
+			fprintf(stderr, "boost gain observed with %lf, support %d\n", boostTau, sup);
+#endif
+			report_boosting (projected, sup, gainmax, yval);
+
+			/* Update lower boosting gain bound in case we already exceeded
+			 * the number of allowed patterns.
+			 */
+			if (bestGraphsGain.size () >= boostN) {
+				boostTau = bestGraphsGain[bestGraphsGain.size() - 1];
+				fprintf(stderr, "   top gains: %lf ... %lf\n",
+					bestGraphsGain[0], boostTau);
+			}
+		}
+		if (boostseen % 1000 == 0) {
+			Graph g(directed);
+			DFS_CODE.toGraph (g);
+
+			fprintf(stderr, "%7u: gain %lf  this.bound %lf  "
+				"global.bound %lf  DFS.size %zu (gsize %zu)\n", boostseen,
+				gainmax, (this->*gainbound) (projected), boostTau,
+				DFS_CODE.size(), g.size());
+		}
+
+		boostseen += 1;
+	} else {
+		// Output the frequent substructure
+		report (projected, sup);
+	}
+
+	/* In case we have a valid upper bound and our graph already exceeds it,
+	 * return.  Note: we do not check for equality as the DFS exploration may
+	 * still add edges within an existing subgraph, without increasing the
+	 * number of nodes.
+	 */
+	if (maxpat_max > maxpat_min && DFS_CODE.nodeCount () > maxpat_max)
+		return;
+
+	/* Compute gain bound for this pattern (projected) and only explore in
+	 * case the bound allows this subgraph to be better.  That is, if the
+	 * bound lets no supergraph be better than all reported ones so far, we
+	 * just return.
+	 */
+	if (boost) {
+		double gainb = (this->*gainbound) (projected);
+#ifdef DEBUG
+		fprintf(stderr, "   gain bound: %lf,  gain min required: %lf\n", gainb, boostTau);
+		fprintf(stderr, "        actual gain for this subgraph: %lf\n", 
+			(this->*gain) (projected, 1.0));
+#endif
+
+		if (gainb <= boostTau + 1e-8) {
+#ifdef DEBUG
+			fprintf(stderr, "      ==> return\n");
+#endif
+			return;
+		}
+	}
+
+	/* We just outputted a frequent subgraph.  As it is frequent enough, so
+	 * might be its (n+1)-extension-graphs, hence we enumerate them all.
+	 */
+	const RMPath &rmpath = DFS_CODE.buildRMPath ();
+	int minlabel = DFS_CODE[0].fromlabel;
+	int maxtoc = DFS_CODE[rmpath[0]].to;
+
+	Projected_map3 new_fwd_root;
+	Projected_map2 new_bck_root;
+	EdgeList edges;
+
+	/* Enumerate all possible one edge extensions of the current substructure.
+	 */
+	for (unsigned int n = 0; n < projected.size(); ++n) {
+
+		unsigned int id = projected[n].id;
+		PDFS *cur = &projected[n];
+		History history (TRANS[id], cur);
+
+		// XXX: do we have to change something here for directed edges?
+
+		// backward
+		for (int i = (int)rmpath.size()-1; i >= 1; --i) {
+			Edge *e = get_backward (TRANS[id], history[rmpath[i]], history[rmpath[0]], history);
+			if (e)
+				new_bck_root[DFS_CODE[rmpath[i]].from][e->elabel].push (id, e, cur);
+		}
+
+		// pure forward
+		// FIXME: here we pass a too large e->to (== history[rmpath[0]]->to
+		// into get_forward_pure, such that the assertion fails.
+		//
+		// The problem is:
+		// history[rmpath[0]]->to > TRANS[id].size()
+		if (get_forward_pure (TRANS[id], history[rmpath[0]], minlabel, history, edges))
+			for (EdgeList::iterator it = edges.begin(); it != edges.end(); ++it)
+				new_fwd_root[maxtoc][(*it)->elabel][TRANS[id][(*it)->to].label].push (id, *it, cur);
+
+		// backtracked forward
+		for (int i = 0; i < (int)rmpath.size(); ++i)
+			if (get_forward_rmpath (TRANS[id], history[rmpath[i]], minlabel, history, edges))
+				for (EdgeList::iterator it = edges.begin(); it != edges.end();  ++it)
+					new_fwd_root[DFS_CODE[rmpath[i]].from][(*it)->elabel][TRANS[id][(*it)->to].label].push (id, *it, cur);
+	}
+
+	/* Test all extended substructures.
+	 */
+	// backward
+	for (Projected_iterator2 to = new_bck_root.begin(); to != new_bck_root.end(); ++to) {
+		for (Projected_iterator1 elabel = to->second.begin(); elabel != to->second.end(); ++elabel) {
+			DFS_CODE.push (maxtoc, to->first, -1, elabel->first, -1);
+			project (elabel->second);
+			DFS_CODE.pop();
+		}
+	}
+
+	// forward
+	for (Projected_riterator3 from = new_fwd_root.rbegin() ;
+		from != new_fwd_root.rend() ; ++from)
+	{
+		for (Projected_iterator2 elabel = from->second.begin() ;
+			elabel != from->second.end() ; ++elabel)
+		{
+			for (Projected_iterator1 tolabel = elabel->second.begin();
+					tolabel != elabel->second.end(); ++tolabel)
+			{
+				DFS_CODE.push (from->first, maxtoc+1, -1, elabel->first, tolabel->first);
+				project (tolabel->second);
+				DFS_CODE.pop ();
+			}
+		}
+	}
+
+	return;
+}
+
+void gSpan::boost_setup (unsigned int _boostN, double _boostTau,
+	unsigned int _boostmax,
+	std::vector<double>& _boostY,
+	std::vector<double>& _boostWeights,
+	int _boostType)
+{
+	boost = true;
+	boostseen = 0;
+	boostmax = _boostmax;
+	boostN = _boostN;
+	boostTau = _boostTau;
+	boostY = _boostY;
+	boostWeights = _boostWeights;
+	boostType = _boostType;
+
+	/* Setup the corresponding function pointers for the gain function adn the
+	 * gainbound function.
+	 */
+	if (boostType == 1) {
+		gain = &gSpan::gain1d5;
+		gainbound = &gSpan::gainbound1d5;
+	} else if (boostType == 2) {
+		gain = &gSpan::gain2;
+		gainbound = &gSpan::gainbound2;
+	}
+
+	/* Precompute the overall label*weight sum-balance.
+	 */
+	boostWeightSum = 0.0;
+	for (unsigned int n = 0 ; n < boostY.size() ; ++n)
+		boostWeightSum += boostWeights[n] * boostY[n];
+
+	bestGraphs.clear ();
+}
+
+// MATLAB/MEX entrypoint removed.
+
+void gSpan::run (std::istream &is, std::ostream &_os,
+		 unsigned int _minsup,
+		 unsigned int _maxpat_min, unsigned int _maxpat_max,
+		 bool _enc,
+		 bool _where,
+		 bool _directed)
+{
+	os = &_os;
+	ID = 0;
+	minsup = _minsup;
+	maxpat_min = _maxpat_min;
+	maxpat_max = _maxpat_max;
+	enc = _enc;
+	where = _where;
+	directed = _directed;
+	boost = false;
+
+	read (is);
+	run_intern ();
+}
+
+void gSpan::run_intern (void)
+{
+	/* In case 1 node subgraphs should also be mined for, do this as
+	 * preprocessing step.
+	 */
+	if (maxpat_min <= 1) {
+		/* Do single node handling, as the normal gspan DFS code based processing
+		 * cannot find subgraphs of size |subg|==1.  Hence, we find frequent node
+		 * labels explicitly.
+		 */
+		for (unsigned int id = 0; id < TRANS.size(); ++id) {
+			for (unsigned int nid = 0 ; nid < TRANS[id].size() ; ++nid) {
+				if (singleVertex[id][TRANS[id][nid].label] == 0) {
+					// number of graphs it appears in
+					singleVertexLabel[TRANS[id][nid].label] += 1;
+				}
+
+				singleVertex[id][TRANS[id][nid].label] += 1;
+			}
+		}
+		/* All minimum support node labels are frequent 'subgraphs'.
+		 * singleVertexLabel[nodelabel] gives the number of graphs it appears
+		 * in.
+		 *
+		 * 1/1.5-class case: All nodelabels that do not appear at all have a
+		 *    gain of zero, hence we do not need to consider them.
+		 *
+		 * 2-class case: Not appearing nodelabels are counted negatively.
+		 */
+		for (std::map<unsigned int, unsigned int>::iterator it =
+			singleVertexLabel.begin () ; it != singleVertexLabel.end () ; ++it)
+		{
+			if ((*it).second < minsup)
+				continue;
+
+			unsigned int frequent_label = (*it).first;
+
+			/* Found a frequent node label, report it.
+			 */
+			Graph g(directed);
+			g.resize (1);
+			g[0].label = frequent_label;
+
+			/* [graph_id] = count for current substructure
+			 */
+			std::vector<unsigned int> counts (TRANS.size ());
+			for (std::map<unsigned int, std::map<unsigned int, unsigned int> >::iterator it2 =
+				singleVertex.begin () ; it2 != singleVertex.end () ; ++it2)
+			{
+				counts[(*it2).first] = (*it2).second[frequent_label];
+			}
+
+			if (boost) {
+				/* Calculate gain and yval.  Here we do not use the normal
+				 * gain function as there is no Projected/DFS_CODE there yet.
+				 * Hence we need to make a distinction between the 1/1.5-class
+				 * and the 2-class case here.
+				 */
+				double gainm = 0.0;
+				double gainm_pos = 0.0;
+				double gainm_neg = 0.0;
+
+				for (unsigned int cid = 0 ; cid < counts.size () ; ++cid) {
+					if (boostType == 1) {
+						// Only consider the positive instances
+						if (counts[cid] == 0)
+							continue;
+
+						gainm += boostY[cid]*boostWeights[cid];	// *1.0 (Y)
+					} else if (boostType == 2) {
+						double addfactor = 1.0;
+
+						if (counts[cid] == 0)
+							addfactor = -1.0;	// negation: pattern does not exist
+
+						gainm_pos += addfactor*boostY[cid]*boostWeights[cid];
+						gainm_neg += -addfactor*boostY[cid]*boostWeights[cid];
+					}
+				}
+				double yval = 1.0;
+
+				if (boostType == 2) {
+					if (gainm_pos >= gainm_neg) {
+						gainm = gainm_pos;
+					} else {
+						gainm = gainm_neg;
+						yval = -1.0;
+					}
+				}
+
+//#ifdef	DEBUG
+				fprintf(stderr, "   single node graph, node label %d, gain %lf, yval %lf\n",
+					frequent_label, gainm, yval);
+//#endif
+				if (gainm > boostTau) {
+					/* Copy it into vector form
+					 */
+					std::map<unsigned int, unsigned int> gycounts;
+					for (unsigned int n = 0 ; n < counts.size () ; ++n)
+						gycounts[n] = counts[n];
+
+					report_boosting_inter (g, (*it).second, gainm, yval, gycounts);
+				}
+			} else {
+				std::map<unsigned int, unsigned int> gycounts;
+				for (unsigned int n = 0 ; n < counts.size () ; ++n)
+					gycounts[n] = counts[n];
+
+				report_single (g, gycounts);
+			}
+		}
+	}
+
+	EdgeList edges;
+	Projected_map3 root;
+
+	for (unsigned int id = 0; id < TRANS.size(); ++id) {
+		Graph &g = TRANS[id];
+		for (unsigned int from = 0; from < g.size() ; ++from) {
+			if (get_forward_root (g, g[from], edges)) {
+				for (EdgeList::iterator it = edges.begin(); it != edges.end();  ++it)
+					root[g[from].label][(*it)->elabel][g[(*it)->to].label].push (id, *it, 0);
+			}
+		}
+	}
+
+	for (Projected_iterator3 fromlabel = root.begin() ;
+		fromlabel != root.end() ; ++fromlabel)
+	{
+		for (Projected_iterator2 elabel = fromlabel->second.begin() ;
+			elabel != fromlabel->second.end() ; ++elabel)
+		{
+			for (Projected_iterator1 tolabel = elabel->second.begin();
+				tolabel != elabel->second.end(); ++tolabel)
+			{
+				/* Build the initial two-node graph.  It will be grown
+				 * recursively within project.
+				 */
+				DFS_CODE.push (0, 1, fromlabel->first, elabel->first, tolabel->first);
+				project (tolabel->second);
+				DFS_CODE.pop ();
+			}
+		}
+	}
+}
+
+}