sgd.c

Go to the documentation of this file.

#include "config.h"
 
#include <assert.h>
#include <cgraph/cgraph.h>
#include <math.h>
#include <neatogen/dijkstra.h>
#include <neatogen/neato.h>
#include <neatogen/neatoprocs.h>
#include <neatogen/randomkit.h>
#include <neatogen/sgd.h>
#include <stdlib.h>
#include <util/alloc.h>
#include <util/bitarray.h>
#include <util/gv_math.h>
#include <util/unreachable.h>
 
static double calculate_stress(double *pos, term_sgd *terms, size_t n_terms) {
  double stress = 0;
  for (size_t ij = 0; ij < n_terms; ij++) {
    const double dx = pos[2 * terms[ij].i] - pos[2 * terms[ij].j];
    const double dy = pos[2 * terms[ij].i + 1] - pos[2 * terms[ij].j + 1];
    const double r = hypot(dx, dy) - terms[ij].d;
    stress += terms[ij].w * (r * r);
  }
  return stress;
}
// it is much faster to shuffle term rather than pointers to term, even though
// the swap is more expensive
static void fisheryates_shuffle(term_sgd *terms, size_t n_terms,
                                rk_state *rstate) {
  for (size_t i = n_terms - 1; n_terms > 0 && i >= 1; i--) {
    const unsigned long j = rk_interval(i, rstate);
 
    SWAP(&terms[i], &terms[j]);
  }
}
 
// graph_sgd data structure exists only to make dijkstras faster
static graph_sgd *extract_adjacency(graph_t *G, int model) {
  size_t n_nodes = 0, n_edges = 0;
  for (node_t *np = agfstnode(G); np; np = agnxtnode(G, np)) {
    assert(ND_id(np) >= 0 && (size_t)ND_id(np) == n_nodes);
    n_nodes++;
    for (edge_t *ep = agfstedge(G, np); ep; ep = agnxtedge(G, ep, np)) {
      if (agtail(ep) != aghead(ep)) { // ignore self-loops and double edges
        n_edges++;
      }
    }
  }
  graph_sgd *graph = gv_alloc(sizeof(graph_sgd));
  graph->sources = gv_calloc(n_nodes + 1, sizeof(size_t));
  graph->pinneds = bitarray_new(n_nodes);
  graph->targets = gv_calloc(n_edges, sizeof(size_t));
  graph->weights = gv_calloc(n_edges, sizeof(float));
 
  graph->n = n_nodes;
  graph->sources[graph->n] = n_edges; // to make looping nice
 
  n_nodes = 0, n_edges = 0;
  for (node_t *np = agfstnode(G); np; np = agnxtnode(G, np)) {
    graph->sources[n_nodes] = n_edges;
    bitarray_set(&graph->pinneds, n_nodes, isFixed(np));
    for (edge_t *ep = agfstedge(G, np); ep; ep = agnxtedge(G, ep, np)) {
      if (agtail(ep) == aghead(ep)) { // ignore self-loops and double edges
        continue;
      }
      node_t *target = (agtail(ep) == np)
                           ? aghead(ep)
                           : agtail(ep); // in case edge is reversed
      graph->targets[n_edges] = (size_t)ND_id(target);
      graph->weights[n_edges] = ED_dist(ep);
      assert(graph->weights[n_edges] > 0);
      n_edges++;
    }
    n_nodes++;
  }
  assert(n_nodes == graph->n);
  assert(n_edges == graph->sources[graph->n]);
  graph->sources[n_nodes] = n_edges;
 
  if (model == MODEL_SHORTPATH) {
    // do nothing
  } else if (model == MODEL_SUBSET) {
    // i,j,k refer to actual node indices, while x,y refer to edge indices in
    // graph->targets initialise to no neighbours
    bitarray_t neighbours_i = bitarray_new(graph->n);
    bitarray_t neighbours_j = bitarray_new(graph->n);
    for (size_t i = 0; i < graph->n; i++) {
      int deg_i = 0;
      for (size_t x = graph->sources[i]; x < graph->sources[i + 1]; x++) {
        size_t j = graph->targets[x];
        if (!bitarray_get(neighbours_i, j)) {   // ignore multiedges
          bitarray_set(&neighbours_i, j, true); // set up sort of hashset
          deg_i++;
        }
      }
      for (size_t x = graph->sources[i]; x < graph->sources[i + 1]; x++) {
        size_t j = graph->targets[x];
        int intersect = 0;
        int deg_j = 0;
        for (size_t y = graph->sources[j]; y < graph->sources[j + 1]; y++) {
          size_t k = graph->targets[y];
          if (!bitarray_get(neighbours_j, k)) {   // ignore multiedges
            bitarray_set(&neighbours_j, k, true); // set up sort of hashset
            deg_j++;
            if (bitarray_get(neighbours_i, k)) {
              intersect++;
            }
          }
        }
        graph->weights[x] = deg_i + deg_j - (2 * intersect);
        assert(graph->weights[x] > 0);
        for (size_t y = graph->sources[j]; y < graph->sources[j + 1]; y++) {
          size_t k = graph->targets[y];
          bitarray_set(&neighbours_j, k, false); // reset sort of hashset
        }
      }
      for (size_t x = graph->sources[i]; x < graph->sources[i + 1]; x++) {
        size_t j = graph->targets[x];
        bitarray_set(&neighbours_i, j, false); // reset sort of hashset
      }
    }
    bitarray_reset(&neighbours_i);
    bitarray_reset(&neighbours_j);
  } else {
    // TODO: model == MODEL_MDS and MODEL_CIRCUIT
    UNREACHABLE(); // mds and circuit model not supported
  }
  return graph;
}
static void free_adjacency(graph_sgd *graph) {
  free(graph->sources);
  bitarray_reset(&graph->pinneds);
  free(graph->targets);
  free(graph->weights);
  free(graph);
}
 
void sgd(graph_t *G, /* input graph */
         int model /* distance model */) {
  if (model == MODEL_CIRCUIT) {
    agwarningf("circuit model not yet supported in Gmode=sgd, reverting to "
               "shortpath model\n");
    model = MODEL_SHORTPATH;
  }
  if (model == MODEL_MDS) {
    agwarningf("mds model not yet supported in Gmode=sgd, reverting to "
               "shortpath model\n");
    model = MODEL_SHORTPATH;
  }
  const size_t n = agnnodes_z(G);
 
  if (Verbose) {
    fprintf(stderr, "calculating shortest paths and setting up stress terms:");
    start_timer();
  }
  // calculate how many terms will be needed as fixed nodes can be ignored
  size_t n_fixed = 0;
  size_t n_terms = 0;
  for (size_t i = 0; i < n; i++) {
    if (!isFixed(GD_neato_nlist(G)[i])) {
      n_fixed++;
      n_terms += n - n_fixed;
    }
  }
  term_sgd *terms = gv_calloc(n_terms, sizeof(term_sgd));
  // calculate term values through shortest paths
  size_t offset = 0;
  graph_sgd *graph = extract_adjacency(G, model);
  for (size_t i = 0; i < n; i++) {
    if (!isFixed(GD_neato_nlist(G)[i])) {
      offset += dijkstra_sgd(graph, i, terms + offset);
    }
  }
  assert(offset == n_terms);
  free_adjacency(graph);
  if (Verbose) {
    fprintf(stderr, " %.2f sec\n", elapsed_sec());
  }
 
  // initialise annealing schedule
  float w_min = terms[0].w, w_max = terms[0].w;
  for (size_t ij = 1; ij < n_terms; ij++) {
    w_min = fminf(w_min, terms[ij].w);
    w_max = fmaxf(w_max, terms[ij].w);
  }
  // note: Epsilon is different from MODE_KK and MODE_MAJOR as it is a minimum
  // step size rather than energy threshold
  //       MaxIter is also different as it is a fixed number of iterations
  //       rather than a maximum
  const double eta_max = 1.0 / w_min;
  const double eta_min = Epsilon / w_max;
  const double lambda = log(eta_max / eta_min) / (MaxIter - 1);
 
  // initialise starting positions (from neatoprocs)
  initial_positions(G, (int)n);
  // copy initial positions and state into temporary space for speed
  double *const pos = gv_calloc(2 * n, sizeof(double));
  bool *unfixed = gv_calloc(n, sizeof(bool));
  for (size_t i = 0; i < n; i++) {
    node_t *node = GD_neato_nlist(G)[i];
    pos[2 * i] = ND_pos(node)[0];
    pos[2 * i + 1] = ND_pos(node)[1];
    unfixed[i] = !isFixed(node);
  }
 
  // perform optimisation
  if (Verbose) {
    fprintf(stderr, "solving model:");
    start_timer();
  }
  rk_state rstate;
  rk_seed(0, &rstate); // TODO: get seed from graph
  for (int t = 0; t < MaxIter; t++) {
    fisheryates_shuffle(terms, n_terms, &rstate);
    const double eta = eta_max * exp(-lambda * t);
    for (size_t ij = 0; ij < n_terms; ij++) {
      // cap step size
      const double mu = fmin(eta * terms[ij].w, 1);
 
      const double dx = pos[2 * terms[ij].i] - pos[2 * terms[ij].j];
      const double dy = pos[2 * terms[ij].i + 1] - pos[2 * terms[ij].j + 1];
      const double mag = hypot(dx, dy);
 
      const double r = (mu * (mag - terms[ij].d)) / (2 * mag);
      const double r_x = r * dx;
      const double r_y = r * dy;
 
      if (unfixed[terms[ij].i]) {
        pos[2 * terms[ij].i] -= r_x;
        pos[2 * terms[ij].i + 1] -= r_y;
      }
      if (unfixed[terms[ij].j]) {
        pos[2 * terms[ij].j] += r_x;
        pos[2 * terms[ij].j + 1] += r_y;
      }
    }
    if (Verbose) {
      fprintf(stderr, " %.3f", calculate_stress(pos, terms, n_terms));
    }
  }
  if (Verbose) {
    fprintf(stderr, "\nfinished in %.2f sec\n", elapsed_sec());
  }
  free(terms);
 
  // copy temporary positions back into graph_t
  for (size_t i = 0; i < n; i++) {
    node_t *node = GD_neato_nlist(G)[i];
    ND_pos(node)[0] = pos[2 * i];
    ND_pos(node)[1] = pos[2 * i + 1];
  }
  free(pos);
  free(unfixed);
}