def Optimize(self,model,gradient='exct',maxnumlinesearch=50,random_starts = 2,verbose=False):
self.gradient = gradient
print('----------------------------------------------')
print(' Optimizing Hyperparameters ')
print('---------------------------------------------- \n')
print('DETAILS: ')
print('--> Performing %i non-linear CG(s) with random initialization(s)' %(random_starts+1))
print('--> Maximum number of line searches = %i' %(maxnumlinesearch))
if gradient == 'exact':
print('--> Calculations of gradients: exact gradients')
else:
print('--> Calculation of gradients: Forward Finite-Differencing')
if model.M>5000:
print('--> Gradients computed in parallel on %i cores \n' %(multiprocessing.cpu_count()))
else:
print('--> Gradients computed on 1 core \n')
# intialize optimum parameters
opthyp,optML,i = CG.minimize(self,model,maxnumlinesearch=maxnumlinesearch, maxnumfuneval=None, red=1.0, verbose=verbose)
optrankfix = self.rank_fix
print('CG run #1 done')
# Discourage very small characteristic lengths
#if opthyp[1]>1.5:
# optML[-1] = np.inf
# Rerun optimization with random intializations
for j in xrange(random_starts):
self.hyp = 0.1*np.random.randint(-40,10,size=(len(opthyp)-1,1))
self.hyp = np.vstack((self.hyp,np.random.randint(-2,6)))
if self.interpolate:
self.rank_fix = (math.exp(-self.hyp[0])**2)/1e6
else:
self.rank_fix = (math.exp(-self.hyp[0])**2)/1e6
print(self.hyp)
hyp, ML, i = CG.minimize(self,model,maxnumlinesearch=maxnumlinesearch, maxnumfuneval=None, red=1.0, verbose=verbose)
# if new optimum is better than the last, save it.
if ML[-1] != -np.inf and ML[-1] <optML[-1]: #and hyp[1]<1.5:
optML = ML
opthyp = hyp
optrankfix = self.rank_fix
print('--> CG run #%i done:' %(j+2))
print('--> Optimum Marginal Likelihood: %e \n\n' %(ML[-1]))
print('**********************************************')
print(' Global Optimum Hyperparameters: ')
print(opthyp)
print('Global Optimum Marginal Likelihood: ')
print(optML[-1])
print('********************************************** \n\n')
# Save global optimum to the Kernel object
self.hyp = opthyp
self.rank_fix = optrankfix
return
def inexactNewtonCGiter(action_of_hessian_on_vector,
eval_obj_func,
m,
g,
n,
convergence_level,
c1=1.e-4,
alpha_init=1.0):
'''
action_of_matrix_on_vector - Function that returns action of nxn matrix H on any given n vector x
eval_obj_func - Evaluate objective function, takes in m as input
m - Parameters
p - Newton direction
g - Gradient
convergence_level - 0 for linear, 1 for superlinear, 2 for quadratic
c1 - Paramteter to quantify sufficient descent
alpha_init - Initial guess for alpha, which decides the step length
'''
p = CG.linear_CG(action_of_hessian_on_vector, m, g, n, convergence_level)
alpha = line_search.backtrackingArmijoLineSearch(eval_obj_func,
m,
p,
g,
c1=1.e-4)
m_new = m + alpha * p
return m_new
def Gaussian_Kron(W,x,y,hyp,rank_fix):
# Initialize variables
N,M = W.shape
D = len(x)
sigma = math.exp(-hyp[0]/D)
s = math.exp(-hyp[-1])
# set a flag to know if the CG converged
flag=False
# iterate 5 times or until CG converged to desired accuracy
for i in xrange(5):
# initialize list for dimensional gram matrices, Eigenvalues, and Eigenvectors
K = []
E = []
# Calculate and stack K, Q, and E in each dimension.
for d in xrange(D):
xd = x[d].reshape(-1,1)
K.append((sigma**2.0)*np.exp(-(np.sum(xd**2.0,1).reshape(-1,1)+np.sum(xd**2.0,1)-2*np.dot(xd,xd.T))/(2.0*(math.exp(-hyp[d+1]))**2)))
E.append(np.real(np.linalg.eigh(K[-1])[0]))
# Calculate eigenvalues of the inducing points
'''
L = E[0]
for d in xrange(1,D):
L = np.kron(L,E[d])
L = np.sort(L,kind='mergesort')
'''
L,ind = KU.largest_eigs(E,N,M)
# Approximate to eigenvalues of KSKI by a factor M/N
L = (float(N)/M)*L
# Calculate approximate log|KSKI| from L and s
complexity = np.sum(np.log(L + (s**2 + rank_fix )*np.ones((N,1))))
# Calculate alpha by Linear CG method
alpha = CG.Linear_CG(W,K,y,s,rank_fix,tolerance=1e-3,maxiter=2000)
# If the CG does not converge, increase the rank fixing term to give better conditioning
if alpha[1] != 0:
print('cg failed, reassigning rank correction term.')
rank_fix = 100*rank_fix
else:
flag = True
# if CG succeeded, return alpha. Else reiterate with a larger rank_fix term.
if flag:
break
alpha = alpha[0]
# Get negative log likelihood (objective function to be minimized)
return 0.5*(np.dot(y.T,alpha)[0][0] + complexity + N*np.log(2*math.pi)),rank_fix
def KISSGP(self):
if self.noise:
noise = self.kernel.hyp[-1]
else:
noise = math.log(1e6)
self.Kuu = self.kernel.Kuu(self.grid.x)
alpha = CG.Linear_CG(self.W,
self.Kuu,
self.y,
math.exp(-noise),
self.kernel.rank_fix,
tolerance=1e-5,
maxiter=5000)
print(alpha[1])
self.alpha = alpha[0]
return
def singleEuler(nu, u, v, Gu, Gv, dx, dy, dt, Nx, Ny, c1, c2):
'''
'''
updGhosts(u, v)
CalG(nu, u, v, dx, dy, Gu, Gv, c1)
u += c2 * dt * Gu
v += c2 * dt * Gv
b = numpy.zeros((Nx+2, Ny+2))
b[1:-1, 1:-1] = (u[2:-1, 1:-1] - u[1:-2, 1:-1]) / dx + \
(v[1:-1, 2:-1] - v[1:-1, 1:-2]) / dy
p, Nitr = CG(Nx, Ny, numpy.zeros((Nx+2, Ny+2)),
b, dx, dy, 1e-15, 'N', refP=0)
u[2:-2, 1:-1] -= (p[2:-1, 1:-1] - p[1:-2, 1:-1]) / dx
v[1:-1, 2:-2] -= (p[1:-1, 2:-1] - p[1:-1, 1:-2]) / dy
return u, v, p, Gu, Gv
def setMBS(self, mbs):
self.mbs = mbs
parity = 1
occPos = 0
empPos = 0
self.occList = []
self.nextEmptyList = []
self.emptyList = []
# Loop through each of the sps's:
for s in xrange(self.nStates):
self.parity[s] = parity # set the parity
if self.mbs & (0x1 << s): # if s is an occupied state
if self.vsList[s]: # if it is a valid state for the operator
self.occList.append(s)
self.nextEmptyList.append(empPos)
occPos = occPos + 1
parity = -parity # if occupied, change the next state parity
elif self.vsList[s]:
self.emptyList.append(s) # mark unoccupied
empPos = empPos + 1
# So, we need to find n = p electrons choose p electron states
# p electrons is nParticles - f electrons
# p states is fixed (well, based on the vsList)
# see discussion in __init__
if self.optStates > 0:
#print self.optStates, occPos
return int(CG.binomialCoeff(self.optStates, self.nParticles - occPos))
else:
return 1
def optimize():
try:
tolerance = math.pow(10, -1 *
int(entry4.get())) # Toleranzschwelle des Nutzers
except:
tolerance = 5 * math.pow(10, -1 * 8)
if numberOfFunctions < 5:
number = numberOfFunctions
else:
number = 5
chooseFunctionsAutomatic(number)
iteration = 0
x = np.zeros(dim)
p = CG(hesse_f_N(x), -1 * grad_f(x), 10 * dim, math.pow(10, -5), x)
F = f(x)
Grad = grad_f(x)
alpha = 1 # Armijo-Backtrackin-Verfahren
while f(x + (alpha * p)) > F + 0.5 * alpha * np.dot(Grad, p):
alpha = 0.5 * alpha
xold = np.copy(x) # Sicherung x
x = x + alpha * p # Update von x
F = f(x)
Grad = grad_f(x)
iteration = 1
while ((np.linalg.norm(Grad) > tolerance * (1 + (abs(F)))) and
(abs(np.linalg.norm(Grad) - np.linalg.norm(grad_f(xold)))) > math.pow(10, -7)) \
or (abs(F - f(xold)) > math.pow(10, -8) * (1 + abs(F)))\
or (np.linalg.norm(xold - x) > math.pow(10, -8) * (1 + np.linalg.norm(x))):
if np.linalg.norm(xold -
x) <= math.pow(10, -6) * (1 + np.linalg.norm(x)):
if number < numberOfFunctions:
addFunctionsAutomatic()
number += 1
else:
break
p = CG(hesse_f_N(x), -1 * Grad, 10 * dim, math.pow(10, -4), x)
alpha = 1 # Armijo-Backtracking-Verfahren
while f(x + (alpha * p)) > F + 0.5 * alpha * np.dot(Grad, p):
alpha = 0.5 * alpha
xold = np.copy(x) # Sicherung x
x = x + alpha * p # Update von x
F = f(x)
Grad = grad_f(x)
iteration += 1
print("Iterationen: " + str(iteration - 1))
print("Anzahl gewählter Funktionen: " + str(number))
print("Minimierer x der Funktionen ist: " + str(x))
print("Der Fehler im Gradienten liegt bei: " + str(np.linalg.norm(Grad)))
print("f(x) = " + str(F))
print("Norm Gradient: " + str(np.linalg.norm(Grad)))
print("Toleranz Gradient: " + str(tolerance * (1 + abs(F))))
print("Unterschied Norm Gradient: " +
str(abs(np.linalg.norm(Grad) - np.linalg.norm(Grad))))
print("Toleranz Unterschied Gradient: " + str((10**-7)))
print("Unterschied f: " + str(abs(F - f(xold))))
print("Tolleranz f: " + str((10**-8) * (1 + abs(F))))
print("Unterschied x: " + str(np.linalg.norm(xold - x)))
print("Tolleranz x: " + str(math.pow(10, -8) * (1 + np.linalg.norm(x))))
def exact_Gaussian_grad2(W,x,y,hyp,rank_fix):
# Initialize variables
N,M = W.shape
D = len(x)
P = len(hyp)
sigma = math.exp(-hyp[0]/D)
s = math.exp(-hyp[-1])
# set a flag to know if the CG converged
flag=False
# iterate 5 times or until CG converged to desired accuracy
for i in xrange(5):
# initialize list for dimensional gram matrices, Eigenvalues, Eigenvectors, and gradients
K = []
E = []
Q = []
# Calculate and stack K, Q, and E in each dimension.
for d in xrange(D):
xd = x[d].reshape(-1,1)
K.append((sigma**2.0)*np.exp(-(np.sum(xd**2.0,1).reshape(-1,1)+np.sum(xd**2.0,1) \
-2*np.dot(xd,xd.T))/(2.0*(math.exp(-hyp[d+1]))**2)))
e,q = np.linalg.eigh(K[-1])
E.append(e)
Q.append(q)
# get N largest eigenvalues
L,ind = KU.largest_eigs(E,N,M)
# Approximate to eigenvalues of KSKI by a factor M/N
L = (float(N)/M)*L
# Calculate approximate log|KSKI| from L and s
complexity = np.sum(np.log(L + (s**2 + rank_fix)*np.ones((N,1))))
# Calculate alpha by Linear CG method
alpha = CG.Linear_CG(W,K,y,s,rank_fix,tolerance=1e-3,maxiter=2000)
# If the CG does not converge, increase the rank fixing term to give better conditioning
if alpha[1] != 0:
print('cg failed, reassigning rank correction term.')
rank_fix = 100*rank_fix
else:
flag = True
# if CG succeeded, return alpha. Else reiterate with a larger rank_fix term.
if flag:
break
alpha = alpha[0]
# calculate gradients of Kuu and then the gradient of the likelihood
grad = np.zeros((P,1))
# If M> 5000 use paralell gradient computation (less than 5000 parallel comunication overcomes benefits)
if M>5000:
l = [(i,K,L,W,Q,alpha,ind,rank_fix,x,hyp) for i in xrange(P)]
pool = Pool()
res = pool.imap(d_ARD, l)
i=0
for g in res:
grad[i] = g
i+=1
pool.close()
pool.terminate()
# Else compute gradients sequentially
else:
for p in xrange(P):
grad[p] = d_ARD((p,K,L,W,Q,alpha,ind,rank_fix,x,hyp))
func = 0.5*(np.dot(y.T,alpha)[0][0] + complexity + N*np.log(2*math.pi))
print(func)
# Get negative log likelihood (objective function to be minimized)
return grad,func,rank_fix
if indRows[i] != iVert :
spCoefs[i] = 0.
else :
spCoefs[i] = 1.
# Suppression des coefficients se trouvant sur la ligne iVert
# ( avec toujours 1 sur la diagonale )
for iCol in xrange(nb_verts):
if iCol != iVert :
for ptRow in xrange(begCols[iCol],begCols[iCol+1]):
if indRows[ptRow] == iVert :
spCoefs[ptRow] = 0.
# On definit ensuite la matrice comme matrice CSC :
spMatrix = sparse.csc_matrix((spCoefs, indRows, begCols))
# Visualisation second membre :
VS.view( coords, elt2verts, b, title = "second membre", visuMesh = False )
# Puis on résoud le système linaire à l'aide d'un gradient conjugué :
x0 = np.zeros(nb_verts, np.double)
for i in xrange(nb_verts):
if coords[i,3]>0:
x0[i] = g(coords[i,0],coords[i,1])
sol, iter, epsilon = CG.solve( spMatrix, x0, b, 100, 1.E-6 )
# On affiche l'erreur faite sur la solution ainsi que le nombre d'itérations
# nécessaire à la convergence
print "Erreur relative : {}".format(epsilon)
print "Iteration pour converger : {}".format(iter)
# Visualisation de la solution :
VS.view( coords, elt2verts, sol, title = "Solution", visuMesh=False )
thresholdValue = cv2.getTrackbarPos('filterThresh', 'image')
thresh = cv2.threshold(blurred, thresholdValue, 255,
cv2.THRESH_BINARY)[1]
testArray = [(lower - thrs / 2).tolist(),
(lower + thrs / 2).tolist(),
lowerBound.tolist(),
upperBound.tolist(), thresholdValue]
cnts = cv2.findContours(thresh.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)[1]
areas = 1
if cnts is not None:
areas = int(len(cnts))
CG.win()
splotch = np.zeros((1, areas), dtype=np.uint8)
if cnts is None:
CG.runGame()
# loop over the contours
#THIS IS THE BLOCK THAT DETECTS GREEN
try:
for i, c in enumerate(cnts, 0):
print(cnts)
M = cv2.moments(c)
splotch[0][i] = int(M["m00"])
try:
max1 = np.argmax(splotch)
def uncon(func, x0, max_iter, tol):
method = "BFGS_LS"
mode = 0 # solve mode
#mode = 1 # analysis mode
# get the obj function and gradient
f, g = func()
n = x0.shape[0]
if (method == "BFGS_LS"):
V0 = np.matrix(np.eye(n))
p = BFGS.BFGS(f, g, x0, V0, n, mode)
if (mode == 0):
x, J = p.optimize(tol)
elif (mode == 1):
x_list, n_iter_list, log_g_norm_list = p.optimize(tol)
elif (method == "BFGS_TR"):
B_0 = np.matrix(np.eye(n))
Delta_0 = 1.0
Delta_max = 10.0
if (mode == 0):
x, J = TR.trustRegion(Delta_0, Delta_max, tol, \
x0, B_0, f, g, mode)
elif (mode == 1):
x_list, n_iter_list, log_g_norm_list = TR.trustRegion(Delta_0, Delta_max, tol, \
x0, B_0, f, g, mode)
elif (method == "CG"):
p = CG.CG(f, g, x0, mode)
if (mode == 0):
x, J = p.optimize(tol)
elif (mode == 1):
x_list, n_iter_list, log_g_norm_list = p.optimize(tol)
if (mode == 0):
return x, J
elif (mode == 1):
return x_list, n_iter_list, log_g_norm_list
def genFermionStateTable(nFermions):
plist = Plist.AtomicOrbitalPlist()
plist.readPlist()
nStates = plist.aoList.nStates
debugFlag = True
# In real situations, more than 32 states will overflow memory
# But in small situations, this is allowed
if nStates > 32:
print '*** WARNING: *** nStates is larger than 32!! Continuing ... '
# The number of fermions cannot be more than the number of states
if nFermions > nStates:
raise Exception('Number of particles exceeds the number of states! '+str(nFermions)+' '+str(nStates))
return
# The dimension is the total number of many-body states
dim = CG.binomialCoeff(nStates, nFermions)
print '<------------------------ State Table -------------------->'
print 'Fermions = ', nFermions, ' ... States = ', nStates, '... Dimension = ', dim
# Create an empty python dictionary
stateTableDict = dict()
stateTable = []
# initialize the first state
# the low nFermions bits are set to 1
state = Bit.BitStr(nStates)
for ferm in range(nFermions):
state.set(ferm)
# cursor is the position of the highest bit in this group of ones
cursor = nFermions - 1
# ones is the total number of ones in this group of ones
ones = nFermions
# --------------------------------------------------------------------
# Loop through each many-body state
# --------------------------------------------------------------------
for mbState in range(dim):
stateTableDict[state.value] = mbState # set the key & value
stateTable.append(state.value)
if dim == 1:
break
# move the high bit up
state.clear(cursor)
cursor = cursor
state.set(cursor+1)
ones = ones - 1 # decrement the number of ones in the group
if ones > 0:
# --------------------------------------------------------------------
# move all of the lower ones down
# --------------------------------------------------------------------
for sbState in range(cursor):
if sbState < ones:
state.set(sbState)
else:
state.clear(sbState)
# --------------------------------------------------------------------
cursor = ones - 1 # move the cursor to end of group
else: # no more ones in the group
# --------------------------------------------------------------------
# find the next one to move
for sbState in range(cursor+1,nStates):
if not state.get(sbState):
cursor = sbState - 1
break
else:
ones = ones + 1
# --------------------------------------------------------------------
# end for sbState in range(cursor+1,nStates+1):
# --------------------------------------------------------------------
f = open('./stateTable.pkl', 'w')
pickle.dump(stateTable, f)
f.close()
f = open('./stateTableDict.pkl','w')
pickle.dump(stateTableDict, f)
f.close()
del stateTable
del stateTableDict
# --------------------------------------------------------------------
# Dump out the table for debugging
# --------------------------------------------------------------------
if debugFlag:
# Read in the state table pickles
f = open('stateTable.pkl', 'r')
dbgStateTable = pickle.load(f)
f.close()
f = open('stateTableDict.pkl', 'r')
dbgStateTableDict = pickle.load(f)
f.close()
for rStateIndex, rState in enumerate(dbgStateTable):
state = Bit.BitStr(nStates, rState)
print 'State %d = %s' % (rStateIndex, state.display())
if rStateIndex != dbgStateTableDict[rState]:
raise Exception('Mismatch between state index and dictionary!'+str(rStateIndex))
f = open('./mbDimension.pkl','w')
pickle.dump(dim, f)
f.close()
f = open('./nParticles.pkl','w')
pickle.dump(nFermions, f)
f.close()
print '<------------------------ End State Table -------------------->'
print