def seek(X, H2, ren=renormL2, beta = 0.001, alpha=0.1, cutoff = 1e-4, maxIter = 100000, recordStep=100):
"""If recordStep = 0, then there is no recording done.
Undefined behavior if not integer."""
errors = []
i = 0
y = ren(X)
while i < maxIter:
#### First, the deep math.
E = errorMat(y, H2)
# Gradient.
g = grad(E, H2) + alpha*repulsiveGrad(X)
# Grad descent.
y = y - beta*g
# Renorm
y = ren(y)
#### Next the logistics; counters, etc.
r = norm(E)
if recordStep != 0 and i % recordStep == 0:
print y
print E
print r
print '#####'
errors.append(r)
if r < cutoff:
break
i += 1
#If we did not dip below the cutoff, we did not converge.
if r > cutoff:
print "Warning: Error greater than cutoff, did not converge."
#Return vector, the clique it represents, and the error trace.
return y, extractClique(np.fabs(np.sign(np.round(y, 4)))), errors
def biggestClique(y, G):
"""Returns the biggest clique from key y, by looking at that biggest k
elements of y until it is no longer a clique."""
l = len(y)
ind = np.argsort(y)
rev = np.argsort(ind)
R = []
for i in range(1, l):
k = np.zeros(l)
k[-i:] = 1
k = k[rev]
r = extractClique(k)
if verifyClique(r, G):
R = r
else:
break
return R
from NPL4 import * from graphReader import getGraph from cliqueProject import extractClique from cliqueProject import verifyClique G = np.array( [[0, 1, 1, 0, 0], [1, 0, 1, 0, 1], [1, 1, 0, 1, 0], [0, 0, 1, 0, 0], [0, 1, 0, 0, 0]]) x, c, errors = solve(G, renormL2, 1e-2, 1e-4, 1000000, 0) G = getGraph() H2 = graphToH(G) x = np.ones(125) y, c2, errors2 = seek(x, H2, renormCube, 1e-2, 1e-8, 100000000, 1000) R = extractClique(np.outer(c2, c2)) print R print len(R) print verifyClique(R, G)
