def k2Validate(G, start, parents): good = True (valid, levels) = G.isBfsTree(start, parents) # isBfsTree implements Graph500 tests 1 and 2 if not valid: if kdt.master(): print "isBfsTree detected failure of Graph500 test %d" % abs(ret) return False # Spec test #3: # every input edge has vertices whose levels differ by no more than 1 # Note: don't actually have input edges, will use the edges in # the resulting graph as a proxy [origI, origJ, ign] = G.toParVec() del ign li = levels[origI]; lj = levels[origJ] del origI if not ((abs(li-lj) <= 1) | ((li==-1) & (lj==-1))).all(): if kdt.master(): print "At least one graph edge has endpoints whose levels differ by more than one and is in the BFS tree" print li, lj good = False # Spec test #4: # the BFS tree spans a connected component's vertices (== all edges # either have both endpoints in the tree or not in the tree, or # source is not in tree and destination is the root) neither_in = (li == -1) & (lj == -1) both_in = (li > -1) & (lj > -1) out2root = (li == -1) & (origJ == start) del origJ if not (neither_in | both_in | out2root).all(): if kdt.master(): print "The tree does not span exactly the connected component, root=%d" % start #print levels, neither_in, both_in, out2root, (neither_in | both_in | out2root) good = False del both_in, out2root # Spec test #5: # a vertex and its parent are joined by an edge of the original graph respects = abs(li-lj) <= 1 if not (neither_in | respects).all(): if kdt.master(): print "At least one vertex and its parent are not joined by an original edge" good = False return good
def gabp(A, b, maxround, epsilon):
copy_time=0
scale_time=0
t_time=0
add_time=0
mul_divide_time=0
sum_time=0
cmp_time=0
t1 = time.time()
m = A.nvert()
pv = kdt.ParVec(0)
#Mh,MJ init to m by m all-zero matrices
Mh = kdt.DiGraph(pv,pv,pv,m)
MJ = Mh.copy()
conv = False
stencil=A.copy()
stencil.ones()
stencil.removeSelfLoops()
#stencil.reverseEdges()
#print stencil
#create an m*m identity matrix
pi = kdt.ParVec.range(0,m)
pw = kdt.ParVec(m,1)
eye = kdt.DiGraph(pi,pi,pw,m)
diagA = A*eye
[piDiagA,pjDiagA,peDiagA] = diagA.toParVec()
h = b.copy()
J = peDiagA.copy()
ha = h / J
r=1
t2 = time.time()
init_time = t2-t1
rel_norm = 40000
while r<=maxround:
if kdt.master():
print "starting GBP round %d, relnorm=%f"%(r, rel_norm)
preRes = ha
t3 = time.time()
# COPY
Mhtemp = stencil.copy()
MJtemp = stencil.copy()
t4 = time.time()
copy_time += (t4-t3)
# SCALE
Mhtemp.scale(h) # default direction: dir=kdt.DiGraph.Out, which scales rows
MJtemp.scale(J)
# print MJtemp.toParVec()
t5 = time.time()
scale_time += t5-t4
# if kdt.master():
# print "scale time: %f" % (t5-t4)
# TRANSPOSE
Mh.reverseEdges()
MJ.reverseEdges()
t6 = time.time()
t_time += t6-t5
# ADD
h_m = Mhtemp + -Mh
J_m = MJtemp + -MJ
t7 = time.time()
add_time += t7-t6
# MUL/DIVIDE
val = -A / J_m
Mh = val * h_m
MJ = val * A
t8 = time.time()
mul_divide_time += t8-t7
# SUM
Mh.removeSelfLoops()
MJ.removeSelfLoops()
h = b + Mh.sum(kdt.DiGraph.In)
J = peDiagA + MJ.sum(kdt.DiGraph.In)
t9 = time.time()
sum_time += t9-t8
Ja = 1.0/J
ha=h*Ja
ha_norm = ha.norm(2)
if (ha_norm == 0.0):
rel_norm = 0
else:
rel_norm = (ha-preRes).norm(2)/ha_norm
#rel_norm = (ha-preRes).norm(2) #Adam
t10 = time.time()
cmp_time += t10-t9
# if kdt.master():
# print "rel_norm %f after round %d"% (rel_norm,r)
if r > 2 and rel_norm<epsilon:
y = kdt.SpParVec(m)
y._spv=A._spm.SpMV_PlusTimes(ha.toSpParVec()._spv)
real_norm = (y-b).toParVec().norm(2)
if kdt.master():
after = time.time()
print "GBP Converged after %d rounds, reached rel_norm %f real_norm %f"% (r,rel_norm,real_norm)
print "run time %fs"%(after-t1)
conv = True
break
r += 1
after=time.time()
if kdt.master():
print "init time: %fs"%init_time
print "copy time: %fs"%copy_time
print "scale_time: %fs"%scale_time
print "t_time: %fs"%t_time
print "add_time: %fs"%add_time
print "m_d_time: %fs"%mul_divide_time
print "sum_time: %fs"%sum_time
print "cmp_time: %fs"%cmp_time
print "total_time: %fs"%(after-t1)
if conv==False:
y = kdt.SpParVec(m)
y._spv=A._spm.SpMV_PlusTimes(ha.toSpParVec()._spv)
real_norm = (y-b).toParVec().norm(2)
if kdt.master():
print "GBP did not converge in %d rounds, reached rel_norm %f real_norm %f"%(r-1,rel_norm,real_norm)
print "run time %fs"%(after-t1)
#print ha
#print Ja
if False:
if kdt.master():
print "writing resulting vector x to x.mtx"
X = kdt.DiGraph(kdt.ParVec.range(nvert), kdt.ParVec.ones(nvert)-1, ha, nvert)
X.save("x.mtx")
return
import getopt
import kdt
k = -1
if (len(sys.argv) > 1):
outfile = sys.argv[1]
for i in range(1, len(sys.argv)):
a = sys.argv[i]
if a == '-k':
k = int(sys.argv[i+1])
if k > 2:
import ModelProblemGen
if (kdt.master()):
print "Generating %d-by-%d model problem..."%(k,k)
A, b = ModelProblemGen.getModelProbem(k)
else:
A = kdt.DiGraph.load('thermal2/thermal2.mtx');
b = kdt.ParVec.load('thermal2/thermal2_b.mtx');
nvert = A.nvert()
def gabp(A, b, maxround, epsilon):
copy_time=0
scale_time=0
t_time=0
add_time=0
mul_divide_time=0
def k2Validate(G, start, parents): good = True [valid, levels] = G.isBfsTree(start, parents) # isBfsTree implements Graph500 tests 1 and 2 if not valid: if kdt.master(): print "isBfsTree detected failure of Graph500 test %d" % abs(ret) return False # Spec test #3: # every input edge has vertices whose levels differ by no more than 1 edgeMax = kdt.SpParVec.toSpParVec(G._spm.SpMV_SelMax(levels.toSpParVecAll()._spv)) edgeMin = -kdt.SpParVec.toSpParVec(G._spm.SpMV_SelMax((-levels).toSpParVecAll()._spv)) if ((edgeMax-edgeMin) > 1).any(): if kdt.master(): print "At least one graph edge has endpoints whose levels differ by more than one" good = False # Spec test #4: # the BFS tree spans a connected component's vertices (== all edges # either have both endpoints in the tree or not in the tree, or # source is not in tree and destination is the root) # set not-in-tree vertices' levels to -2 import pyCombBLAS as pcb levels._dpv.Apply(pcb.ifthenelse(pcb.bind2nd(pcb.equal_to(),-1), pcb.set(-2), pcb.identity())) edgeMax = kdt.SpParVec.toSpParVec(G._spm.SpMV_SelMax(levels.toSpParVecAll()._spv)) edgeMin = -kdt.SpParVec.toSpParVec(G._spm.SpMV_SelMax((-levels).toSpParVecAll()._spv)) if ((edgeMax-edgeMin) > 1).any(): if kdt.master(): print "The tree does not span exactly the connected component, root=%d" good = False # Spec test #5: # a vertex and its parent are joined by an edge of the original graph, # except for the root, which has no parent in the tree Gnv = G.nvert(); Gne = G.nedge() [Gi, Gj, ign] = G.toParVec() del ign # non-root tree vertices == NRT Vs NRTVs = (levels!=-2) & (parents!=kdt.ParVec.range(Gnv)) nNRTVs = NRTVs.nnz() TGi = kdt.ParVec.range(nNRTVs) TGj1 = kdt.ParVec.range(Gnv)[NRTVs] TGj2 = parents[NRTVs] M = max(Gne, Gnv) #FIX: really should use SpParMats here, as don't need spm and spmT tmpG1 = kdt.HyGraph(TGi, TGj1, 1, M, Gnv) tmpG2 = kdt.HyGraph(TGi, TGj2, 1, M, Gnv) tmpG1._spm += tmpG2._spm tmpG1._spmT += tmpG2._spmT del tmpG2 tmpG3 = kdt.HyGraph(Gi, Gj, 1, M, Gnv) tmpG4 = kdt.DiGraph() tmpG4._spm = tmpG1._spm.SpMM(tmpG3._spmT) #!? not tmp3._spmT ? maxIncid = tmpG4.max(kdt.DiGraph.Out)[kdt.ParVec.range(Gnv) < nNRTVs] if (maxIncid != 2).any(): if kdt.master(): print "At least one vertex and its parent are not joined by an original edge" good = False return good
del tmpG2
tmpG3 = kdt.HyGraph(Gi, Gj, 1, M, Gnv)
tmpG4 = kdt.DiGraph()
tmpG4._spm = tmpG1._spm.SpMM(tmpG3._spmT) #!? not tmp3._spmT ?
maxIncid = tmpG4.max(kdt.DiGraph.Out)[kdt.ParVec.range(Gnv) < nNRTVs]
if (maxIncid != 2).any():
if kdt.master():
print "At least one vertex and its parent are not joined by an original edge"
good = False
return good
if len(file) == 0:
raise SystemExit, "No generation of Graph500 HyGraph for now; must use file"
if kdt.master():
print "Generating a Graph500 RMAT graph with 2^%d vertices..."%(scale)
G = kdt.HyGraph()
K1elapsed = G.genGraph500Edges(scale)
#G.save("testgraph.mtx")
if kdt.master():
print "Generation took %fs."%(K1elapsed)
else:
if kdt.master():
print 'Loading %s'%(file)
G = kdt.HyGraph.load(file)
K1elapsed = 0.0
if False:
def detect(self, input, epsilon, mu, debug=False):
start = time.time()
# Load graph as a matrix
matrix = kdt.Mat.load(fname=input, element=True, par_IO=False)
if kdt.master():
print '- Matrix Loading Time: ' + str(time.time() - start)
if debug:
print 'Input Matrix: '
print matrix
# Recover symmetricity
t0 = time.time()
temp = matrix.copy()
temp.transpose()
distanceMatrix = matrix.eWiseApply(
temp,
op=(lambda e1, e2: e1 if e1 >= 0 else e2),
allowANulls=True,
allowBNulls=True,
doOp=(lambda e1, e2: e1 >= 0 or e2 >= 0),
inPlace=False,
ANull=-1,
BNull=-1)
matrix = distanceMatrix.copy()
matrix.apply(op=(lambda e: 1))
degrees = matrix.count(dir=kdt.Mat.Column, pred=(lambda e: e >= 0))
if debug and kdt.master():
print 'Degrees: '
print degrees
blankMatrix = kdt.Mat.eye(n=matrix.nrow(), m=matrix.ncol())
blankMatrix.removeMainDiagonal()
temp = matrix.copy()
temp.scale(degrees, op=(lambda e1, e2: e2 + 1), dir=kdt.Mat.Row)
temp.scale(degrees,
op=(lambda e1, e2: math.sqrt(e1 * (e2 + 1))),
dir=kdt.Mat.Column)
distanceMatrix.eWiseApply(temp,
op=(lambda e1, e2: e1 / e2),
allowANulls=False,
allowBNulls=True,
inPlace=True)
epsilonMatrix = blankMatrix.eWiseApply(
distanceMatrix,
op=(lambda e1, e2: e2),
allowANulls=True,
allowBNulls=False,
doOp=(lambda e1, e2: e2 >= epsilon),
inPlace=False)
if kdt.master():
print '- E-distance Computation Time: ' + str(time.time() - t0)
if debug:
print 'Epsilon Matrix: '
print epsilonMatrix
t0 = time.time()
cores = epsilonMatrix.count(dir=kdt.Mat.Row)
cores.apply(op=(lambda e: 1 if e >= mu else 0))
if kdt.master():
print '- Computing Core Time: ' + str(time.time() - t0)
if debug and kdt.master():
print 'Cores: '
print cores
t0 = time.time()
temp.scale(cores, dir=kdt.Mat.Row)
temp.scale(cores, dir=kdt.Mat.Column)
coreMatrix = temp.eWiseApply(temp,
op=(lambda e1, e2: 1),
allowANulls=False,
allowBNulls=False,
doOp=(lambda e1, e2: e1 > 0 and e2 > 0),
allowIntersect=True)
coreMatrix.eWiseApply(epsilonMatrix,
op=(lambda e1, e2: e1),
allowANulls=False,
allowBNulls=False,
inPlace=True)
if kdt.master():
print '- Computing Core Matrix: ' + str(time.time() - t0)
if debug and kdt.master():
print 'Core Matrix: '
print coreMatrix
t0 = time.time()
sourceIndex, targetIndex, valueIndex = coreMatrix.toVec()
graph = kdt.DiGraph(sourceIndex, targetIndex, valueIndex,
matrix.nrow())
components = graph.connComp()
if kdt.master():
print '- Computing Component Time: ' + str(time.time() - t0)
if debug and kdt.master():
print 'Components: '
print components
t0 = time.time()
frequencies = components.hist()
frequencies.eWiseApply(cores,
op=(lambda e1, e2: 0),
allowANulls=False,
allowBNulls=False,
doOp=(lambda e1, e2: e2 == 1),
inPlace=True)
if kdt.master():
print '- Computing Frequency Time: ' + str(time.time() - t0)
if debug and kdt.master():
print 'Frequencies: '
print frequencies
t0 = time.time()
solitaries = frequencies.eWiseApply(
cores,
op=(lambda e1, e2: 1),
allowANulls=False,
allowBNulls=False,
doOp=(lambda e1, e2: e1 == 1 and e2 != 1),
inPlace=False)
if kdt.master():
print '- Computing Solitary Time: ' + str(time.time() - t0)
if debug and kdt.master():
print 'Solitaries: '
print solitaries
t0 = time.time()
# Set all solitary nodes as outliers
components.eWiseApply(solitaries,
op=(lambda e1, e2: -1),
allowANulls=False,
allowBNulls=False,
doOp=(lambda e1, e2: e2 == 1),
inPlace=True)
# Find borders
temp = epsilonMatrix.copy()
temp.apply(op=(lambda e: 1))
temp.scale(cores, dir=kdt.Mat.Column)
temp.scale(components, dir=kdt.Mat.Column)
coreMatrix = blankMatrix.eWiseApply(temp,
op=(lambda e1, e2: e2),
allowANulls=True,
allowBNulls=False,
doOp=(lambda e1, e2: e2 > 0),
inPlace=False)
borders = coreMatrix.max(dir=kdt.Mat.Row, init=-1)
if kdt.master():
print '- Computing Border Time: ' + str(time.time() - t0)
if debug and kdt.master():
print 'Borders: '
print borders
t0 = time.time()
# Assign maximum labels to borders
components.eWiseApply(borders,
op=(lambda e1, e2: e2),
allowANulls=False,
allowBNulls=False,
doOp=(lambda e1, e2: e2 >= 0),
inPlace=True)
if kdt.master():
print '- Assigning Border Label Time ' + str(time.time() - t0)
t0 = time.time()
# Find hubs
solitaries.eWiseApply(borders,
op=(lambda e1, e2: 1 if e1 * e2 < 0 else 0),
allowANulls=False,
allowBNulls=False,
inPlace=True)
matrix.scale(solitaries,
op=(lambda e1, e2: e1 if e2 > 0 else 0),
dir=kdt.Mat.Row)
neighborMatrix = blankMatrix.eWiseApply(matrix,
op=(lambda e1, e2: e2),
allowANulls=True,
allowBNulls=False,
doOp=(lambda e1, e2: e2 > 0),
inPlace=False)
neighborMatrix.scale(components,
op=(lambda e1, e2: e2 if e1 > 0 else 0),
dir=kdt.Mat.Column)
neighborMatrix = blankMatrix.eWiseApply(neighborMatrix,
op=(lambda e1, e2: e2),
allowANulls=True,
allowBNulls=False,
doOp=(lambda e1, e2: e2 >= 0),
inPlace=False)
def check_cluster(current, previous):
if previous == -2:
return current
elif previous == -1:
return -1
else:
if current == -2:
return previous
else:
if previous != current:
return -1
else:
return current
neighbor_clusters = neighborMatrix.reduce(dir=kdt.Mat.Row,
op=check_cluster,
init=-2)
components.eWiseApply(neighbor_clusters,
op=(lambda e1, e2: -3),
allowANulls=False,
allowBNulls=False,
doOp=(lambda e1, e2: e2 == -1),
inPlace=True)
if kdt.master():
print 'Finding Hub Time: ' + str(time.time() - t0)
if kdt.master():
print 'Community Detection Time: ' + str(time.time() - start)
clusters = set()
outlier = 0
hub = 0
for cluster in components:
if cluster == -1:
outlier += 1
elif cluster == -3:
hub += 1
else:
clusters.add(cluster)
if kdt.master():
print 'Clusters: ' + str(len(clusters))
print 'Hubs: ' + str(hub)
print 'Outliers: ' + str(outlier)
'''
Created on Sep 30, 2017
@author: Seokyong Hong
'''
import os
import sys
import kdt
import time
from community.SCAN import SCAN
if __name__ == '__main__':
if len(sys.argv) != 4:
print(
"Usage: mpirun -np <#processes> python Test.py <input> <mu> <epsilon>"
)
sys.exit()
start = time.time()
input_path = sys.argv[1]
mu = int(sys.argv[2])
epsilon = float(sys.argv[3])
scan = SCAN()
scan.detect(input=input_path, epsilon=epsilon, mu=mu, debug=False)
if kdt.master():
print 'Time: ' + str(time.time() - start)
import kdt import sys import time from stats import splitthousands if (len(sys.argv) < 1): B = kdt.DiGraph() B.genGraph500Edges(10) B._spm.Apply(kdt.pyCombBLAS.set(1)) else: inmatrixfile = sys.argv[1] if (kdt.master()): print "Loading matrix from",inmatrixfile B = kdt.DiGraph.load(inmatrixfile) bedges = B._spm.getnee() expansion=3 inflation = 3 prunelimit = 0.0000001 addSelfLoops=False # nedges run if kdt.master(): print "Starting run to find number of edges..." C, nedges = B._markov(addSelfLoops=addSelfLoops, expansion=expansion, inflation=inflation, prunelimit=prunelimit, retNEdges=True) # timed run if kdt.master(): print "nedges=%d. Starting timed run..."%(nedges) before = time.time()
