def minimize(self):
nreps = self.nrep
nbins = self.nebins
visitsT = (self.visits1d)
#print "min vis", np.min(visitsT)
self.logP = np.where(visitsT != 0, np.log(visitsT), 0)
#print "minlogp", np.min(self.logP)
self.reduced_energy = self.binenergy[np.newaxis, :] / (
self.Tlist[:, np.newaxis] * self.k_B)
self.whampot = WhamPotential(self.logP, self.reduced_energy)
X = np.random.rand(nreps + nbins)
E = self.whampot.getEnergy(X)
#print "energy", E
#print "quenching"
try:
from pygmin.optimize import mylbfgs as quench
ret = quench(X, self.whampot, iprint=-1, maxstep=1e4)
except ImportError:
from pygmin.optimize import lbfgs_scipy as quench
ret = quench(X, self.whampot)
#print "quench energy", ret.energy
X = ret.coords
self.logn_E = X[nreps:]
self.w_i_final = X[:nreps]
def minimize(self):
#shape(visits2d) is now (nqbins, nebins, nreps)
#we need it to be (nreps, nqbins*nebins)
#first reorder indices
nreps = self.nrep
nebins = self.nebins
nqbins = self.nqbins
nbins = self.nebins * self.nqbins
reduced_energy = np.zeros([nreps, nebins, nqbins])
for j in range(self.nqbins):
reduced_energy[:, :, j] = self.binenergy[np.newaxis, :] / (
self.Tlist[:, np.newaxis] * self.k_B)
visits = self.visits2d
visits = np.reshape(visits, [nreps, nbins])
reduced_energy = np.reshape(reduced_energy, [nreps, nbins])
self.logP = np.where(visits != 0, np.log(visits.astype(float)), 0)
from wham_potential import WhamPotential
whampot = WhamPotential(visits, reduced_energy)
nvar = nbins + nreps
if False:
X = np.random.rand(nvar)
else:
# estimate an initial guess for the offsets and density of states
# so the minimizer converges more rapidly
offsets_estimate, log_dos_estimate = wham_utils.estimate_dos(
visits, reduced_energy)
X = np.concatenate((offsets_estimate, log_dos_estimate))
assert X.size == nvar
E0, grad = whampot.getEnergyGradient(X)
rms0 = np.linalg.norm(grad) / np.sqrt(grad.size)
from wham_utils import lbfgs_scipy
ret = lbfgs_scipy(X, whampot)
# print "initial energy", whampot.getEnergy(X)
# try:
# from pele.optimize import mylbfgs as quench
# ret = quench(X, whampot, iprint=10, maxstep = 1e4)
# except ImportError:
# from pele.optimize import lbfgs_scipy as quench
# ret = quench(X, whampot)
if self.verbose:
print "chi^2 went from %g (rms %g) to %g (rms %g) in %d iterations" % (
E0, rms0, ret.energy, ret.rms, ret.nfev)
#self.logn_Eq = zeros([nebins,nqbins], float64)
X = ret.coords
self.logn_Eq = X[nreps:]
self.w_i_final = X[:nreps]
self.logn_Eq = np.reshape(self.logn_Eq, [nebins, nqbins])
self.logn_Eq = np.where(
self.visits2d.sum(0) == 0, self.LOGMIN, self.logn_Eq)
def minimize(self):
#shape(visits2d) is now (nqbins, nebins, nreps)
#we need it to be (nreps, nqbins*nebins)
#first reorder indices
nreps = self.nrep
nebins = self.nebins
nqbins = self.nqbins
nbins = self.nebins * self.nqbins
#visits = np.zeros([nreps, nebins, nqbins], np.integer)
reduced_energy = np.zeros([nreps, nebins, nqbins])
# for k in range(self.nrep):
# for j in range(self.nqbins):
# for i in range(self.nebins):
# #visits[k,i,j] = self.visits2d[i,j,k]
# reduced_energy[k,i,j] = self.binenergy[i] / (self.Tlist[k]*self.k_B)
for j in range(self.nqbins):
reduced_energy[:,:,j] = self.binenergy[np.newaxis,:] / (self.Tlist[:,np.newaxis]*self.k_B)
visits = self.visits2d
visits = np.reshape( visits, [nreps, nbins ])
reduced_energy = np.reshape( reduced_energy, [nreps, nbins])
self.logP = np.where( visits != 0, np.log( visits.astype(float) ), 0 )
from wham_potential import WhamPotential
whampot = WhamPotential( self.logP, reduced_energy )
nvar = nbins + nreps
X = np.random.rand(nvar)
print "initial energy", whampot.getEnergy(X)
try:
from pygmin.optimize import mylbfgs as quench
ret = quench(X, whampot, iprint=10, maxstep = 1e4)
except ImportError:
from pygmin.optimize import lbfgs_scipy as quench
ret = quench(X, whampot)
print "quenched energy", ret.energy
global_min = False
if global_min:
from pygmin.basinhopping import BasinHopping
from pygmin.takestep.displace import RandomDisplacement
takestep = RandomDisplacement(stepsize=10.)
takestep.useAdaptiveStep()
takestep.adaptive_class.f = 2.
bh = BasinHopping(X, whampot, takestep)
bh.run(1000)
#self.logn_Eq = zeros([nebins,nqbins], float64)
X = ret.coords
self.logn_Eq = X[nreps:]
self.w_i_final = X[:nreps]
self.logn_Eq = np.reshape(self.logn_Eq, [nebins, nqbins])
self.logn_Eq = np.where( self.visits2d.sum(0) == 0, self.LOGMIN, self.logn_Eq )
def minimize(self):
#shape(visits2d) is now (nqbins, nebins, nreps)
#we need it to be (nreps, nqbins*nebins)
#first reorder indices
nreps = self.nrep
nebins = self.nebins
nqbins = self.nqbins
nbins = self.nebins * self.nqbins
#visits = np.zeros([nreps, nebins, nqbins], np.integer)
reduced_energy = np.zeros([nreps, nebins, nqbins])
# for k in range(self.nrep):
# for j in range(self.nqbins):
# for i in range(self.nebins):
# #visits[k,i,j] = self.visits2d[i,j,k]
# reduced_energy[k,i,j] = self.binenergy[i] / (self.Tlist[k]*self.k_B)
for j in range(self.nqbins):
reduced_energy[:, :, j] = self.binenergy[np.newaxis, :] / (
self.Tlist[:, np.newaxis] * self.k_B)
visits = self.visits2d
visits = np.reshape(visits, [nreps, nbins])
reduced_energy = np.reshape(reduced_energy, [nreps, nbins])
self.logP = np.where(visits != 0, np.log(visits.astype(float)), 0)
from wham_potential import WhamPotential
whampot = WhamPotential(self.logP, reduced_energy)
nvar = nbins + nreps
X = np.random.rand(nvar)
from wham_utils import lbfgs_scipy
ret = lbfgs_scipy(X, self.whampot)
# print "initial energy", whampot.getEnergy(X)
# try:
# from pele.optimize import mylbfgs as quench
# ret = quench(X, whampot, iprint=10, maxstep = 1e4)
# except ImportError:
# from pele.optimize import lbfgs_scipy as quench
# ret = quench(X, whampot)
print "quenched energy", ret.energy
#self.logn_Eq = zeros([nebins,nqbins], float64)
X = ret.coords
self.logn_Eq = X[nreps:]
self.w_i_final = X[:nreps]
self.logn_Eq = np.reshape(self.logn_Eq, [nebins, nqbins])
self.logn_Eq = np.where(
self.visits2d.sum(0) == 0, self.LOGMIN, self.logn_Eq)
def globalMinimization(self):
"""
in experimentation i've never been able to find more than
one minimum
"""
nreps = self.nrep
nbins = self.nebins
visitsT = (self.visits1d)
#print "min vis", np.min(visitsT)
self.logP = np.where(visitsT != 0, np.log(visitsT), 0)
#print "minlogp", np.min(self.logP)
self.reduced_energy = self.binenergy[np.newaxis, :] / (
self.Tlist[:, np.newaxis] * self.k_B)
self.whampot = WhamPotential(self.logP, self.reduced_energy)
X = np.random.rand(nreps + nbins)
E = self.whampot.getEnergy(X)
print "energy", E
print "quenching"
from pygmin.optimize import lbfgs_scipy as quench
ret = quench(X, self.whampot)
print "quench energy", ret.energy
from pygmin.basinhopping import BasinHopping
from pygmin.takestep.displace import RandomDisplacement
takestep = RandomDisplacement(stepsize=10)
takestep.useAdaptiveStep()
takestep.adaptive_class.f = 1.5 #i have no idea what a good stepsize should be
bh = BasinHopping(X, self.whampot, takestep)
import matplotlib.pyplot as plt
for i in range(10):
bh.run(2000)
self.logn_E = bh.coords[nreps:]
cvdata = self.calc_Cv(400)
plt.plot(cvdata[:, 0], cvdata[:, 5], '-')
plt.show()
X = bh.coords
self.logn_E = X[nreps:]
self.w_i_final = X[:nreps]
