// DrawGraph creates a file filename and draws an SPN spn in Graphviz dot.
func DrawGraph(filename string, s spn.SPN) {
file, err := os.Create(filename)
if err != nil {
fmt.Printf("Error. Could not create file [%s].\n", filename)
panic(err)
}
defer file.Close()
fmt.Fprintf(file, "graph {\n")
// If the SPN is itself an univariate distribution, create a graph with a single node.
if s.Type() == "leaf" {
fmt.Fprintf(file, "X1 [label=<X<sub>1</sub>>,shape=circle];\n")
fmt.Fprintf(file, "}")
file.Close()
return
}
// Else, BFS the SPN and write nodes to filename.
nvars, nsums, nprods := 0, 0, 0
queue := common.Queue{}
queue.Enqueue(&BFSPair{Spn: s, Pname: "", Weight: -1.0})
for !queue.Empty() {
currpair := queue.Dequeue().(*BFSPair)
curr, pname, pw := currpair.Spn, currpair.Pname, currpair.Weight
ch := curr.Ch()
nch := len(ch)
name := "N"
currt := curr.Type()
// In case it is a sum node. Else product node.
if currt == "sum" {
name = fmt.Sprintf("S%d", nsums)
fmt.Fprintf(file, "%s [label=\"+\",shape=circle];\n", name)
nsums++
} else if currt == "product" {
name = fmt.Sprintf("P%d", nprods)
fmt.Fprintf(file, "%s [label=<×>,shape=circle];\n", name)
nprods++
}
// If pname is empty, then it is the root node. Else, link parent node to current node.
if pname != "" {
if pw >= 0 {
fmt.Fprintf(file, "%s -- %s [label=\"%.3f\"];\n", pname, name, pw)
} else {
fmt.Fprintf(file, "%s -- %s\n", pname, name)
}
}
var w []float64
if curr.Type() == "sum" {
w = (curr.(*spn.Sum).Weights())
}
// For each children, run the BFS.
for i := 0; i < nch; i++ {
c := ch[i]
// If leaf, then simply write to the graphviz dot file. Else, recurse the BFS.
if c.Type() == "leaf" {
cname := fmt.Sprintf("X%d", nvars)
fmt.Fprintf(file, "%s [label=<X<sub>%d</sub>>,shape=circle];\n", cname, c.Sc()[0])
nvars++
if currt == "sum" {
fmt.Fprintf(file, "%s -- %s [label=\"%.3f\"]\n", name, cname, w[i])
} else {
fmt.Fprintf(file, "%s -- %s\n", name, cname)
}
} else {
tw := -1.0
if w != nil {
tw = w[i]
}
queue.Enqueue(&BFSPair{Spn: c, Pname: name, Weight: tw})
}
}
}
fmt.Fprintf(file, "}")
}
// DrawGraphTools creates a file filename and draws an SPN spn in graph-tools. The resulting file
// is a python source code that outputs a PNG image of the graph.
func DrawGraphTools(filename string, s spn.SPN) {
file, err := os.Create(filename)
if err != nil {
fmt.Printf("Error. Could not create file [%s].\n", filename)
panic(err)
}
defer file.Close()
outname := utils.StringConcat(filename[0:len(filename)-len(filepath.Ext(filename))], ".png")
fmt.Fprintf(file, "from graph_tool.all import *\n\n")
fmt.Fprintf(file, "g = Graph(directed=True)\n")
fmt.Fprintf(file, "vcolors = g.new_vertex_property(\"string\")\n")
fmt.Fprintf(file, "vnames = g.new_vertex_property(\"string\")\n")
fmt.Fprintf(file, "enames = g.new_edge_property(\"string\")\n\n")
fmt.Fprintf(file, "def add_node(name, type):\n\tv=g.add_vertex()\n\tvnames[v]=name\n\t"+
"vcolors[v]=type\n\treturn v\n\n")
fmt.Fprintf(file, "def add_edge(o, t, name):\n\te=g.add_edge(o, t)\n\tenames[e]=name\n\treturn e\n\n")
fmt.Fprintf(file, "def add_edge_nameless(o, t):\n\te=g.add_edge(o, t)\n\treturn e\n\n\n")
// If the SPN is itself an univariate distribution, create a graph with a single node.
if s.Type() == "leaf" {
fmt.Fprintf(file, "add_node(\"X\")\n\n")
fmt.Fprintf(file, "g.vertex_properties[\"name\"]=vnames\n")
fmt.Fprintf(file, "g.vertex_properties[\"color\"]=vcolors\n")
fmt.Fprintf(file, "\ngraph_draw(g, vertex_text=g.vertex_properties[\"name\"], "+
"edge_text=enames, vertex_fill_color=g.vertex_properties[\"color\"], output=\"%s\")\n",
outname)
return
}
// Else, BFS the SPN and write nodes to filename.
nvars, nsums, nprods := 0, 0, 0
queue := common.Queue{}
queue.Enqueue(&BFSPair{Spn: s, Pname: "", Weight: -1.0})
for !queue.Empty() {
currpair := queue.Dequeue().(*BFSPair)
curr, pname, pw := currpair.Spn, currpair.Pname, currpair.Weight
ch := curr.Ch()
nch := len(ch)
name := "N"
currt := curr.Type()
// In case it is a sum node. Else product node.
if currt == "sum" {
name = fmt.Sprintf("S%d", nsums)
fmt.Fprintf(file, "%s = add_node(\"+\", \"#ff3300\")\n", name)
nsums++
} else if currt == "product" {
name = fmt.Sprintf("P%d", nprods)
fmt.Fprintf(file, "%s = add_node(\"*\", \"#669900\")\n", name)
nprods++
}
// If pname is empty, then it is the root node. Else, link parent node to current node.
if pname != "" {
if pw >= 0 {
fmt.Fprintf(file, "add_edge(%s, %s, \"%.3f\")\n", pname, name, pw)
} else {
fmt.Fprintf(file, "add_edge_nameless(%s, %s)\n", pname, name)
}
}
var w []float64
if curr.Type() == "sum" {
w = (curr.(*spn.Sum).Weights())
}
// For each children, run the BFS.
for i := 0; i < nch; i++ {
c := ch[i]
// If leaf, then simply write to the graphviz dot file. Else, recurse the BFS.
if c.Type() == "leaf" {
cname := fmt.Sprintf("X%d", nvars)
fmt.Fprintf(file, "%s = add_node(\"X_%d\", \"#0066ff\")\n", cname, c.Sc()[0])
nvars++
if currt == "sum" {
fmt.Fprintf(file, "add_edge(%s, %s, \"%.3f\")\n", name, cname, w[i])
} else {
fmt.Fprintf(file, "add_edge_nameless(%s, %s)\n", name, cname)
}
} else {
tw := -1.0
if w != nil {
tw = w[i]
}
queue.Enqueue(&BFSPair{Spn: c, Pname: name, Weight: tw})
}
}
}
fmt.Fprintf(file, "g.vertex_properties[\"name\"]=vnames\n")
fmt.Fprintf(file, "g.vertex_properties[\"color\"]=vcolors\n")
//fmt.Fprintf(file, "\ngraph_draw(g, vertex_text=g.vertex_properties[\"name\"], "+
//"edge_text=enames, vertex_fill_color=g.vertex_properties[\"color\"], "+
//"output_size=[16384, 16384], output=\"%s\", bg_color=[1, 1, 1, 1])\n", outname)
fmt.Fprintf(file, "\ngraph_draw(g, "+
"edge_text=enames, vertex_fill_color=g.vertex_properties[\"color\"], "+
"output_size=[16384, 16384], output=\"%s\", bg_color=[1, 1, 1, 1])\n", outname)
}
// Generative learning with learning rate eta.
func Generative(S spn.SPN, data []map[int]int, eta float64, mode spn.InfType) spn.SPN {
n := len(data)
conv := 1.0
last := 0.0
fmt.Println("Running generative learning.")
// Set root's partial derivative to 1.
S.Rootify("pnode")
S.SetStore(true)
for math.Abs(conv) > 0.01 {
s := 0.0
klast := 0.0
for i := 0; i < n; i++ {
// Call soft inference to store values.
k := S.Value(data[i])
s += k - klast
klast = k
fmt.Println("Computed and stored soft inference values.")
// Backpropagate through SPN.
S.RootDerive("pweight", "pnode", "soft", mode)
fmt.Println("Backpropagated through SPN computing derivatives.")
// Update weights.
S.GenUpdate(eta, "pweight")
fmt.Printf("Weight Updated according to instance %d.\n", i)
S.RResetDP("soft")
}
d := s - last
last = s
conv = d
fmt.Printf("Generative Learning diff: %.5f\n", math.Abs(conv))
}
return S
}
