def umbrella_sampling(self, tau, class_prob):
us = usample.UmbrellaSampler(self.gumbel_softmax_fn,
lpfargs=[tau, class_prob],
debug=True,
burn_acor=20)
temps = np.linspace(1, 10, 8)
us.add_umbrellas(temps,
ic=[0],
numwalkers=8,
sampler=emcee.EnsembleSampler)
pos_us, _, _ = us.run(1000, freq=1000, repex=10, grstop=0.01)
quant_splits = np.linspace(0, 1, len(class_prob) + 1)[1:]
quant_splits[:-1] = quant_splits[:-1] + np.random.uniform(
low=-1, high=1, size=(1, len(quant_splits) - 1)) * 0.001
quant_val = np.quantile(pos_us[:, 0], quant_splits)
sampled_classes = np.digitize(pos_us[:, 0], quant_val, right=True)
return sampled_classes
lpf = -0.5 * np.dot(pp, np.dot(icov, pp))
return lpf
#
# Now create the umbrella sampler object
#
means = np.array([0.5, -0.25])
icov = np.array([[1, 0.5], [0.5, 1]])
us = usample.UmbrellaSampler(log_prob_fn,
lpfargs=[means, icov],
mpi=True,
debug=True,
burn_acor=20)
#
# Now add some umbrellas.
# First, define some temperatures to run with.
#
temps = np.linspace(1, 10, 8)
#
# Then add an umbrella at each temperature. Use four walkers, and give some initial conditions
# Can be added individually, or in bulk:
#
def log_prob_fn(p ):
x=p[0]
y=p[1]
# Two Gaussian distributions, centered on +/- 1
lpf = np.log( np.exp(-16*(x-1)**2) + np.exp(-16*(x+1)**2 ))
# And an independent Gaussian in another direction
lpf -= 0.5*y*y
return lpf
#
# Now create the umbrella sampler object
us = usample.UmbrellaSampler( log_prob_fn , mpi=True, debug=True, burn_acor=20 )
#
# Build umbrellas along a line between two points: the two peaks of the distribution.
# This line is a 1D "CV" or Collective Variable, embedded in the higher dimensions.
cvfn = ["line", [np.array([-1.0,0]), np.array([1.0,0])] ]
# The start of the line is "0", and the end is "1".
# We define the centers of the biasing distribution. Here we define four umbrellas:
centers = np.linspace( 0 , 1 , 4 )
# We may also make use of sampling using different temperatures within a window.
# Lets try sampling at a temperature of 1 and 5.
temps = [1,5]
# Note that this gives a total of 4x2=8 umbrellas overall.
(ndim, np.size(xmod)))
print("active indices:", np.arange(np.size(xmod))[ia])
print("initial values:", x0)
print("initial scatter:", step)
print(anames)
nwalkers = 10
pp = np.vstack([
x0 + np.random.normal(scale=step, size=len(x0))
for i in range(nwalkers)
])
us = usample.UmbrellaSampler(sn_like_walk,
lpfargs=args,
lpfkwargs=kwargs,
mpi=True,
debug=True,
burn_acor=20)
if (us.is_master()):
t0 = clock()
#
# Now add some umbrellas.
# First, define some temperatures to run with.
#
temps = np.linspace(1, 10, 5.)
#
# Then add an umbrella at each temperature. Use numwalkers walkers, and give some initial conditions