|
@@ -0,0 +1,108 @@
|
|
|
|
|
+package main
|
|
|
|
|
+
|
|
|
|
|
+import (
|
|
|
|
|
+ "math"
|
|
|
|
|
+)
|
|
|
|
|
+
|
|
|
|
|
+const (
|
|
|
|
|
+ HIDDEN int = 200
|
|
|
|
|
+)
|
|
|
|
|
+
|
|
|
|
|
+type (
|
|
|
|
|
+ Parameter struct {
|
|
|
|
|
+ Wo, Wi, B Matrix
|
|
|
|
|
+ }
|
|
|
|
|
+
|
|
|
|
|
+ Layer struct {
|
|
|
|
|
+ d int
|
|
|
|
|
+ f func(Matrix) Matrix
|
|
|
|
|
+ p Parameter
|
|
|
|
|
+ O, I, E Matrix
|
|
|
|
|
+ }
|
|
|
|
|
+)
|
|
|
|
|
+
|
|
|
|
|
+func ReLU(A Matrix) Matrix {
|
|
|
|
|
+ for i := 0; i < A.N(); i++ {
|
|
|
|
|
+ for j := 0; j < A.M(); j++ {
|
|
|
|
|
+ A[i][j] = math.Max(0, A[i][j])
|
|
|
|
|
+ }
|
|
|
|
|
+ }
|
|
|
|
|
+ return A
|
|
|
|
|
+}
|
|
|
|
|
+
|
|
|
|
|
+func Softmax(A Matrix) Matrix {
|
|
|
|
|
+ for i := 0; i < A.N(); i++ {
|
|
|
|
|
+ max, sum := 0., 0.
|
|
|
|
|
+ for j := 0; j < A.M(); j++ {
|
|
|
|
|
+ max = math.Max(max, A[i][j])
|
|
|
|
|
+ }
|
|
|
|
|
+ for j := 0; j < A.M(); j++ {
|
|
|
|
|
+ A[i][j] = math.Exp(A[i][j] - max)
|
|
|
|
|
+ sum += A[i][j]
|
|
|
|
|
+ }
|
|
|
|
|
+ for j := 0; j < A.M(); j++ {
|
|
|
|
|
+ A[i][j] /= sum
|
|
|
|
|
+ }
|
|
|
|
|
+ }
|
|
|
|
|
+ return A
|
|
|
|
|
+}
|
|
|
|
|
+
|
|
|
|
|
+func GetEmbedding(G Graph, u, k int, l []Layer) Vector {
|
|
|
|
|
+ if k == 0 {
|
|
|
|
|
+ return G.X[u]
|
|
|
|
|
+ }
|
|
|
|
|
+ l[k].E = Multiply(Matrix{GetEmbedding(G, u, k-1, l)}, l[k].p.B)
|
|
|
|
|
+ l[k-1].O, l[k-1].I = MakeMatrix(1, l[k-1].d), MakeMatrix(1, l[k-1].d)
|
|
|
|
|
+ Do, Di := 0, 0
|
|
|
|
|
+ for v, w := range G.A[u] {
|
|
|
|
|
+ if w == 1 {
|
|
|
|
|
+ l[k-1].O.Add(Matrix{GetEmbedding(G, v, k-1, l)})
|
|
|
|
|
+ Do++
|
|
|
|
|
+ } else {
|
|
|
|
|
+ l[k-1].I.Add(Matrix{GetEmbedding(G, v, k-1, l)})
|
|
|
|
|
+ Di++
|
|
|
|
|
+ }
|
|
|
|
|
+ }
|
|
|
|
|
+ if Do > 0 {
|
|
|
|
|
+ l[k-1].O.Divide(float64(Do))
|
|
|
|
|
+ l[k].E.Add(Multiply(l[k-1].O, l[k].p.Wo))
|
|
|
|
|
+ }
|
|
|
|
|
+ if Di > 0 {
|
|
|
|
|
+ l[k-1].I.Divide(float64(Di))
|
|
|
|
|
+ l[k].E.Add(Multiply(l[k-1].I, l[k].p.Wi))
|
|
|
|
|
+ }
|
|
|
|
|
+ return l[k].f(l[k].E)[0]
|
|
|
|
|
+}
|
|
|
|
|
+
|
|
|
|
|
+func Train(G Graph) []Layer {
|
|
|
|
|
+ p1 := Parameter{MakeRandomMatrix(1433, HIDDEN), MakeRandomMatrix(1433, HIDDEN), MakeRandomMatrix(1433, HIDDEN)}
|
|
|
|
|
+ p2 := Parameter{MakeRandomMatrix(HIDDEN, 7), MakeRandomMatrix(HIDDEN, 7), MakeRandomMatrix(HIDDEN, 7)}
|
|
|
|
|
+ l := []Layer{{d: 1433}, {d: HIDDEN, f: ReLU, p: p1}, {d: 7, f: Softmax, p: p2}}
|
|
|
|
|
+ for u, X := range G.X {
|
|
|
|
|
+ GetEmbedding(G, u, 2, l)
|
|
|
|
|
+ delta := MakeMatrix(1, 7)
|
|
|
|
|
+ delta[0][nodeLabel[u]] = 1
|
|
|
|
|
+ delta.Sub(l[2].E)
|
|
|
|
|
+ DWo2, DWi2, DB2 := Multiply(l[1].O.Transpose(), delta), Multiply(l[1].I.Transpose(), delta), Multiply(l[1].E.Transpose(), delta)
|
|
|
|
|
+ DWo2.Divide(10)
|
|
|
|
|
+ DWi2.Divide(10)
|
|
|
|
|
+ DB2.Divide(10)
|
|
|
|
|
+ delta = Multiply(delta, l[2].p.B.Transpose())
|
|
|
|
|
+ for i := 0; i < HIDDEN; i++ {
|
|
|
|
|
+ if l[1].E[0][i] == 0 {
|
|
|
|
|
+ delta[0][i] = 0
|
|
|
|
|
+ }
|
|
|
|
|
+ }
|
|
|
|
|
+ DWo1, DWi1, DB1 := Multiply(l[0].O.Transpose(), delta), Multiply(l[0].I.Transpose(), delta), Multiply(Matrix{X}.Transpose(), delta)
|
|
|
|
|
+ DWo1.Divide(10)
|
|
|
|
|
+ DWi1.Divide(10)
|
|
|
|
|
+ DB1.Divide(10)
|
|
|
|
|
+ l[2].p.Wo.Add(DWo2)
|
|
|
|
|
+ l[2].p.Wi.Add(DWi2)
|
|
|
|
|
+ l[2].p.B.Add(DB2)
|
|
|
|
|
+ l[1].p.Wo.Add(DWo1)
|
|
|
|
|
+ l[1].p.Wi.Add(DWi1)
|
|
|
|
|
+ l[1].p.B.Add(DB1)
|
|
|
|
|
+ }
|
|
|
|
|
+ return l
|
|
|
|
|
+}
|