// Graph converts an HMM Net object to a graph where states are nodes. Like
// an HMM net, the graph contains a single entry and exit non-emmiting state.
func (m *Net) Graph() *graph.Graph {
g := graph.New()
// Create nodes.
for i := 0; i < m.ns; i++ {
var entry, exit bool
if i == 0 {
entry = true
}
if i == m.ns-1 {
exit = true
}
key := m.Name + "-" + strconv.FormatInt(int64(i), 10)
nv := nodeValue{scorer: m.B[i], entry: entry, exit: exit}
g.Set(key, nv)
}
// Connect nodes.
for i := 0; i < m.ns; i++ {
fromKey := m.Name + "-" + strconv.FormatInt(int64(i), 10)
for j := i; j < m.ns; j++ {
if m.A.At(i, j) > math.Inf(-1) {
toKey := m.Name + "-" + strconv.FormatInt(int64(j), 10)
g.Connect(fromKey, toKey, m.A.At(i, j))
}
}
}
return g
}
// SearchGraph creates a graph based on a set of HMM networks.
func (set *Set) SearchGraph() *graph.Graph {
glog.Infof("building search graph using %d hmms", set.size())
g := graph.New()
for _, m := range set.Nets {
glog.Infof("adding model [%s]", m.Name)
err := g.Add(m.Graph())
if err != nil {
glog.Fatalf("failed to build search graph with error: ", err)
}
}
// Log prob of model in search graph.
logp := -math.Log(float64(set.size()))
// Start node. Only node with no predescessors.
start := g.Set("start", nodeValue{})
// End node. Only no with no successors.
end := g.Set("end", nodeValue{})
// Backoff node connects all networks exit states to all entry states.
backoff := g.Set("backoff", nodeValue{})
nodes := g.GetAll()
for _, n := range nodes {
val := n.Value().(nodeValue)
if val.entry {
// backoff to entry
backoff.Connect(n, logp)
}
if val.exit {
// exit to backoff
n.Connect(backoff, 0)
}
}
// backoff to end
backoff.Connect(end, 0)
// start to backoff
start.Connect(backoff, 0)
// Normalize all transition weights.
// TODO
glog.Infof("search graph created - num nodes: %d", g.Len())
glog.V(3).Info("search graph:\n", g)
return g
}