def convert_X_to_GCNDataset(self, X):
"""
Same code as above, dedicated to the predict mode (no need for Y)
"""
graph_id = 0
nf = X[0]
edge = X[1]
ef = X[2]
nb_node = nf.shape[0]
graph = GCNDataset(str(graph_id))
graph.X = nf
graph.Y = -np.ones(
(nb_node, len(self.labelBinarizer.classes_)), dtype='i')
# print(edger)
A1 = sp.coo_matrix((np.ones(edge.shape[0]), (edge[:, 0], edge[:, 1])),
shape=(nb_node, nb_node))
# A2 = sp.coo_matrix((np.ones(edger.shape[0]), (edger[:, 0], edger[:, 1])), shape=(nb_node, nb_node))
graph.A = A1 # + A2
# JL: unued?? edge_normalizer = Normalizer()
# Normalize EA
E0 = np.hstack([edge, ef]) # check order
# E1 = np.hstack([edger, efr]) # check order
graph.E = E0
#graph.compute_NA()
graph.compute_NodeEdgeMat()
return graph
def convert_lX_lY_to_GCNDataset(self,
lX,
lY,
training=False,
test=False,
predict=False):
gcn_list = []
graph_id = 0
# This has state information here --> move that to DU_Model_ECN ...
lys = []
for _, ly in zip(lX, lY):
lys.extend(list(ly))
#print (lys)
if training:
self.labelBinarizer.fit(lys)
for lx, ly in zip(lX, lY):
nf = lx[0]
edge = lx[1]
ef = lx[2]
nb_node = nf.shape[0]
graph = GCNDataset(str(graph_id))
graph.X = nf
if training or test:
graph.Y = self.labelBinarizer.transform(ly)
elif predict:
graph.Y = -np.ones(
(nb_node, len(self.labelBinarizer.classes_)), dtype='i')
else:
raise Exception(
'Invalid Usage: one of train,test,predict should be true')
# We are making the adacency matrix here
# print(edger)
A1 = sp.coo_matrix(
(np.ones(edge.shape[0]), (edge[:, 0], edge[:, 1])),
shape=(nb_node, nb_node))
# A2 = sp.coo_matrix((np.ones(edger.shape[0]), (edger[:, 0], edger[:, 1])), shape=(nb_node, nb_node))
graph.A = A1 # + A2
# JL: unued?? edge_normalizer = Normalizer()
# Normalize EA
E0 = np.hstack([edge, ef]) # check order
# E1 = np.hstack([edger, efr]) # check order
graph.E = E0
#graph.compute_NA()
graph.compute_NodeEdgeMat()
gcn_list.append(graph)
graph_id += 1
return gcn_list
def get_graph_test():
#For graph att net
X = np.array([[1.0, 0.5], [0.5, 0.5], [0.0, 1.0]], dtype='float32')
E = np.array([[0, 1, 1.0], [1, 0, 1.0], [2, 1, 1.0], [1, 2, 1.0]],
dtype='float32')
Y = np.array([[1, 0], [0, 1], [0, 1]], dtype='int32')
gcn = GCNDataset('UT_test_1')
gcn.X = X
gcn.E = E
gcn.Y = Y
gcn.compute_NodeEdgeMat()
return gcn
def test_logit_convolve(self):
# 3 nodes a;b,c a<->b and c<->b a->b<c>
X = np.array([[1.0, 2.0], [6.3, 1.0], [4.3, -2.0]])
Y = np.array([[1, 0], [0, 1.0], [1.0, 0.0]])
E = np.array([
[0, 1, 1.0, 1, 0], #edge a->b
[1, 0, 1.0, 0, 1], #edge b->a
[2, 1, 1.0, 0.0, 1.0]
])
nb_node = 3
gA = GCNDataset('GLogitConvolve')
gA.X = X
gA.Y = Y
gA.E = E
gA.A = sp.coo_matrix((np.ones(E.shape[0]), (E[:, 0], E[:, 1])),
shape=(nb_node, nb_node))
gA.compute_NodeEdgeMat()
gA.compute_NA()
#Test in degree out_degree
print(gA.in_degree, gA.out_degree)
self.assertAlmostEqual(2, gA.in_degree[1])
print(gA.NA_indegree)
self.assertAlmostEqual(0.5, gA.NA_indegree[1, 0])
#self.assertAlmostEqual(2, gA.indegree[1])
#now assuming P(Y|a)=[1,0] P(Y|c)=[1,0] and current P(Y|b)=[0.5,0.5]
pY = np.array([[1, 0], [0.5, 0.5], [0.8, 0.2]])
#Node b has two edges
# Yt=[0 1;1 0]
Yt = np.array([[0.0, 1.0], [1.0, 0.0]])
pY_Yt = tf.matmul(pY, Yt, transpose_b=True)
Yt_sum = EdgeConvNet.logitconvolve_fixed(pY, Yt, gA.NA_indegree)
init = tf.global_variables_initializer()
with tf.Session() as session:
session.run(init)
Ytt = session.run(pY_Yt)
print(Ytt)
Res = session.run(Yt_sum)
print(Res)