def edgeEva(self, edge, adj, modelnum=20, edgenum=20):
res = 0
n = adj.shape[0]
for i in range(modelnum):
tempadj = copy.deepcopy(adj)
for j in range(edgenum):
a = random.randint(0, n - 1)
b = random.randint(0, n - 1)
tempadj[a, b] = 1
tempadj[b, a] = 1
g = gemodel_GCN(tempadj,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled))
g.train()
p1 = g.performance()
tempadj[edge[0], edge[1]] = 1
tempadj[edge[1], edge[0]] = 1
g = gemodel_GCN(tempadj,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled))
g.train()
p2 = g.performance()
if p2 >= p1:
res += 1
print('edge: {} evaluated {} better score'.format(edge, res))
return (res, (edge[0], edge[1]))
def edgeEvaNew(self, candidates, adj):
res = collections.defaultdict(list)
try:
lencan = len(candidates)
modelnum = int(2 * lencan * self.edgeevaltimes / self.poolnum /
self.evalEdgeNum) + 1
# print('eval model num: {}'.format(modelnum))
for i in range(modelnum):
addededge = []
tempadj = copy.deepcopy(adj)
for j in range(self.evalEdgeNum):
a, b = candidates[random.randint(0, lencan - 1)]
tempadj[a, b] = 1
tempadj[b, a] = 1
addededge.append((a, b))
g = gemodel_GCN(tempadj,
self.features,
self.labels,
split_t=self.split_t
) # , seed=self.seed, dropout=self.dropout)
g.train()
p1 = g.acu()
for e in addededge:
res[e].append(p1)
except BaseException as err:
print('raised exception edgeEvaNew: {}'.format(err))
return res
def edgeEvaNew(self, candidators, adj, edgenum=20):
res = collections.defaultdict(list)
try:
# n = adj.shape[0]
lencan = len(candidators)
modelnum = int(2 * self.candinum * self.knn * self.edgeevaltimes /
self.poolnum / edgenum)
# print('modelnum', modelnum)
for i in range(modelnum):
addededge = []
tempadj = copy.deepcopy(adj)
for j in range(edgenum):
a, b = candidators[random.randint(0, lencan - 1)]
tempadj[a, b] = 1
tempadj[b, a] = 1
addededge.append((a, b))
g = gemodel_GCN(tempadj,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled))
g.train()
p1 = g.performance()
for e in addededge:
res[e].append(p1)
except BaseException as err:
print('raised exception edgeEvaNew: {}'.format(err))
# print('edge: {} evaluated {} better score'.format(edge, res))
# print('return eval res:{}'.format(res))
return res
def par(self, es, prevadj):
tempAdj = copy.deepcopy(prevadj)
for i, j in es:
tempAdj[i, j] = 1
tempAdj[j, i] = 1
g = gemodel_GCN(tempAdj,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled),
seed=self.seed,
dropout=0)
g.train()
per = g.performance()
# print('{}, {} finished'.format(i, j))
return per
def subseteval2(self, topset, prevadj):
'''eval edge set performance, randomly and some edges from top set
'''
# print('topset len: {}'.format(len(topset)))
tempadj = copy.deepcopy(prevadj)
eset = set()
for i in range(self.edgeaddnum):
ran = random.randint(0, len(topset)-1)
eset.add(topset[ran])
a, b = topset[ran]
tempadj[a, b] = 1
tempadj[b, a] = 1
g = gemodel_GCN(tempadj, self.features, self.labels, split_t=self.split_t, seed=self.seed, dropout=self.dropout)
g.train()
return (eset, g.acu())
def getembs(self, adj):
print('generate several embds')
embeddings = []
if self.default_model == 'GCN':
a = utils.preprocess_graph(adj)
for i in range(self.model_num):
g = gemodel_GCN(a,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled))
g.train()
emb = g.getembeddings()
embeddings.append(emb)
else:
print('ERR: wrong default GE model for add edges')
exit(-1)
return embeddings
def subseteval(self, adj, candidators):
tempadj = copy.deepcopy(adj)
eset = set()
for i in range(self.enum2add):
ran = random.randint(0, len(candidators) - 1)
eset.add(candidators[ran])
a, b = candidators[ran]
tempadj[a, b] = 1
tempadj[b, a] = 1
# _, p = Modeltest_GCN.subprocess_GCN(tempadj, self.features, self.labels, split_t=(self.split_train, self.split_val, self.split_unlabeled), seed=1, dropout=0)
g = gemodel_GCN(tempadj,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled),
seed=1,
dropout=0)
g.train()
p = g.performance()
return (eset, p)
def edgeReconstruction(self, prevAdj, embds, edgenum=20):
better = worse = eq = 0
# multiprocessing.set_start_method('fork')
p = Pool()
res = []
res2 = []
enum = 30
for e in range(1000):
es = []
for f in range(enum):
a = self.deleted_edges[random.randint(
0,
len(self.deleted_edges) - 1)]
es.append((a[1], a[2]))
r = p.apply_async(self.par, args=(es, prevAdj))
res.append(r)
for e in range(1000):
es = []
for f in range(enum):
a = random.randint(0, prevAdj.shape[0] - 1)
b = random.randint(0, prevAdj.shape[0] - 1)
es.append((a, b))
r = p.apply_async(self.par, args=(es, prevAdj))
res2.append(r)
p.close()
p.join()
g = gemodel_GCN(prevAdj,
self.features,
self.labels,
split_t=(self.split_train, self.split_val,
self.split_unlabeled),
seed=self.seed,
dropout=0)
g.train()
initperformance = g.performance()
ret = []
for x in res:
ret.append(x.get())
best = 0
worst = 100
for a in ret:
best = max(best, a)
worst = min(worst, a)
if a > initperformance:
better += 1
elif a < initperformance:
worse += 1
else:
eq += 1
ret2 = []
for x in res2:
ret2.append(x.get())
better2 = worse2 = eq2 = 0
best2 = 0
worst2 = 100
for a in ret2:
best2 = max(best2, a)
worst2 = min(worst2, a)
if a > initperformance:
better2 += 1
elif a < initperformance:
worse2 += 1
else:
eq2 += 1
print(
'better2: {}, worse2: {}, eq2: {}, best2: {}, worse2: {}, init: {}'
.format(better2, worse2, eq2, best2, worst2, initperformance))
print('better: {}, worse: {}, eq: {}, best: {}, worse: {}, init: {}'.
format(better, worse, eq, best, worst, initperformance))
exit(-1)
split_train, split_val, split_unlabeled = split_t = Preprocess.splitdata(
_N, labels, seed=123, share=share)
distnum = 1000
dst = distEdge_ran(distnum)
deletesizes = [1, 0.8, 0.5, 0.2]
res = []
labelk = []
fo = '{}/{}'
for i in range(len(deletesizes)):
adj, remained, deleted = spa.delete_edges(_A_obs, k=deletesizes[i])
adj_d = dst.disturb(adj)
print('disturb adj, add {} ran edges, prev {}, after {}'.format(
len(deleted), adj.nnz, adj_d.nnz))
initadjset = edge2list.sett(adj)
disadjset = edge2list.sett(adj_d)
g = gemodel_GCN(adj_d, feas, labels, split_t=split_t, sGCN=True)
g.train()
print(g.acu())
W = g.model.W1.eval(session=g.model.session)
fl = FastLoss(adj_d, feas, labels, W, split_val, deleted_e=deleted)
bins, dens = fl.test(initset=initadjset, disset=disadjset)
res.append((bins, dens))
labelk.append(fo.format(distnum, adj.nnz / 2))
Hist.subplots(res, labelk)
# fl.test_deleted()
# fl.test_remained(remained)
adj, remained, deleted = spa.delete_edges(_A_obs, k=percent)
_N = _A_prev.shape[0]
# split_train, split_val, split_unlabeled = Preprocess.splitdata(_N, labels) #seed share as default
# split_t = (split_train, split_val, split_unlabeled)
# gcn = gemodel_GCN(_A_prev, feas, labels, split_t=split_t, seed=1, dropout=0)
# run_time(gcn.train)
# print('performance: {}, acu: {}'.format(gcn.performance(), gcn.acu()))
# gcn = gemodel_GCN(adj, feas, labels, split_t=split_t, seed=1, dropout=0)
# gcn.train()
# print('performance: {}, acu: {}'.format(gcn.performance(), gcn.acu()))
split_train, split_val, split_unlabeled = Preprocess.splitdata(_N, labels, seed=12, share=(0.052, 0.3693)) #seed share as default
split_t = (split_train, split_val, split_unlabeled)
gcn = gemodel_GCN(_A_prev, feas, labels, split_t=split_t, seed=1, dropout=0)
run_time(gcn.train)
print('performance: {}, acu: {}'.format(gcn.performance(), gcn.acu()))
gcn = gemodel_GCN(_A_prev, feas, labels, split_t=split_t, seed=1, dropout=0, sGCN=True)
run_time(gcn.train)
print('performance: {}, acu: {}'.format(gcn.performance(), gcn.acu()))
gcn = gemodel_GCN(adj, feas, labels, split_t=split_t, seed=1, dropout=0)
gcn.train()
print('performance: {}, acu: {}'.format(gcn.performance(), gcn.acu()))
# print('preprocess, delete some edges, remaind edges num(bi): {}'.format(adj.nnz))
# savefile = 'data/coradele.npz'