def test_unbiased_checkerboard(self):
"""
Test that checkerboard problem returns both valid solutions.
"""
sampler = GibbsSampler()
checkerboard = IsingModel(J={(0, 1): 1, (1, 2): 1, (2, 3): 1, (3, 0): 1}, h={})
result = sampler.sample(checkerboard, 10000)
solutions = [s.as_tuple for s in result]
assert (-1, 1, -1, 1) in solutions
assert (1, -1, 1, -1) in solutions
def main():
"""
Main program.
"""
data = initialize()
g = GibbsSampler(sequences=data['sequences'],
motif_width=data['width'])
g.find_motif(iterations=data['iterations'])
print_sequences(data['sequences'], data['width'])
return
class Solver:
# Calculate the log of the posterior probability of a given sentence
# with a given part-of-speech labeling. Right now just returns -999 -- fix this!
def __init__(self, train_file_name, test_file_name):
self.train_file = train_file_name
#self.trainer_builder = Trainer_builder(self.train_file).get_trainer()
#a = GibbsSampler(self.trainer_builder)
#label = a.do_gibbs('The discovery struck Nick like a blow .'.split())
# a.calculate_posterior(('at', 'the', 'same', 'instant', ',', 'nick', 'hit', 'the', 'barrel', 'and', 'threw', 'himself', 'upon', 'the', 'smaller', 'man', '.'), map(lambda item: item.upper(), label))
def posterior(self, model, sentence, label):
if model == "Simple":
simple_instance = Simplified(sentence)
return simple_instance.calc_posterior(self.trainer_builder, label)
elif model == "Complex":
return self.gibbs_instance.calculate_posterior(
sentence, map(lambda item: item.upper(), label))
elif model == "HMM":
viterbi_instance = Viterbi(sentence)
return viterbi_instance.calc_posterior(self.trainer_builder, label)
else:
print("Unknown algo!")
# Do the training!
#
def train(self, data):
self.trainer_builder = Trainer_builder(self.train_file).get_trainer()
# Functions for each algorithm. Right now this just returns nouns -- fix this!
#
def simplified(self, sentence):
simple_instance = Simplified(sentence)
return simple_instance.get_most_probable_tags(self.trainer_builder)
def complex_mcmc(self, sentence):
self.gibbs_instance = GibbsSampler(self.trainer_builder)
return self.gibbs_instance.do_gibbs(sentence)
def hmm_viterbi(self, sentence):
viterbi_instance = Viterbi(sentence)
self.hmm_tags = viterbi_instance.get_most_probable_tags(
self.trainer_builder)[0]
return self.hmm_tags
# This solve() method is called by label.py, so you should keep the interface the
# same, but you can change the code itself.
# It should return a list of part-of-speech labelings of the sentence, one
# part of speech per word.
#
def solve(self, model, sentence):
if model == "Simple":
return self.simplified(sentence)
elif model == "Complex":
return self.complex_mcmc(sentence)
elif model == "HMM":
output = self.hmm_viterbi(sentence)
return output
else:
print("Unknown algo!")
def main():
model_width = 4
model_height = 4
temperature = 1
decimals = 2
builder = RandomBuilder(model_width, model_height)
model = builder.generate(decimals=decimals)
# First produce a graph with D-Wave
sampler = DWaveSampler()
results_dwave = sample_and_histogram(sampler, model, temperature)
# Then with gibbss sampleler
sampler = GibbsSampler(n_variables=model_height * model_width)
results_gibbs = sample_and_histogram(sampler, model, temperature)
save_experiment("histogram", {
'width': model_width,
'height': model_height,
'model_decimals': decimals,
'temperature': temperature,
'dwave_samples': results_dwave.raw_data,
'gibbs_samples': results_gibbs.raw_data
})
def __init__(self, f, X, y, x_range, eval_only, extra, options):
self.f = f
self.x_range = x_range
self.options = options
self.well_defined = X.shape[0] > 0
self.solver = GibbsSampler(X, y, options)
self.eval_only = eval_only
self.opt_n = options['opt_n']
self.dx = options['dx']
if eval_only:
self.newX = extra
else:
self.n_bo = extra
self.opt_n = np.maximum(self.opt_n, self.n_bo * 2)
self.max_value = self.options['max_value']
self.n_bo_top_percent = self.options['n_bo_top_percent']
def test_biased_checkerboard(self):
"""
Test that biased checkerboard (at one qubit) returns only one solution.
"""
sampler = GibbsSampler()
biased_checkerboard = IsingModel(J={(0, 1): 1, (1, 2): 1, (2, 3): 1, (3, 0): 1}, h={0: 2})
result = sampler.sample(biased_checkerboard, 10000)
solutions = [s.as_tuple for s in result]
assert (-1, 1, -1, 1) in solutions
assert len(solutions) == 1
# Enforce the opposite checkerboard
biased_checkerboard = IsingModel(J={(0, 1): 1, (1, 2): 1, (2, 3): 1, (3, 0): 1}, h={0: -2})
result = sampler.sample(biased_checkerboard, 10000)
solutions = [s.as_tuple for s in result]
assert (1, -1, 1, -1) in solutions
assert len(solutions) == 1
class bo(object):
def __init__(self, f, X, y, x_range, eval_only, extra, options):
self.f = f
self.t = options['t']
self.x_range = x_range
self.options = options
self.well_defined = X.shape[0] > 0
self.solver = GibbsSampler(X, y, options)
self.eval_only = eval_only
self.opt_n = options['opt_n']
self.dx = options['dx']
if eval_only:
self.newX = extra
else:
self.n_bo = extra
self.opt_n = np.maximum(self.opt_n, self.n_bo * 2)
if 'ziw' in options:
self.max_value = self.options['max_value'] + np.random.normal(
) * 0.05
else:
self.max_value = self.options['max_value']
self.n_bo_top_percent = self.options['n_bo_top_percent']
def learn(self):
self.gp, self.z, self.k = self.solver.run(self.options['gibbs_iter'])
def run(self):
if self.eval_only:
ynew = [self.f(x) for x in self.newX]
return ynew
# return random inputs if X is empty
if not self.well_defined:
xnew = np.random.uniform(self.x_range[0], self.x_range[1],
(self.n_bo, self.dx))
acfnew = [self.max_value] * self.n_bo
return xnew, acfnew, self.solver.z, self.solver.k, self.max_value
# learn and optimize
self.learn()
# initialization
xnew = np.empty((self.n_bo, self.dx))
xnew[0] = np.random.uniform(self.x_range[0], self.x_range[1])
# optimize group by group
all_cat = np.unique(self.z)
for a in np.random.permutation(all_cat):
active = self.z == a
af = lambda x: MAX_VALUE(x, xnew[0], active, self.max_value, self.
gp) # lower confidential bound
xnew[:, active], acfnew = global_minimize(af, self.x_range[:,active], \
self.opt_n, self.n_bo, self.n_bo_top_percent)
return xnew, acfnew, self.z, self.k, self.max_value
parser.add_argument("--resume", help = "Specify if you want to resume sampling", action = "store_true")
parser.add_argument("-y", "--yaml", help = "Specify yaml file")
args = parser.parse_args()
yaml_file = args.yaml
with open(yaml_file) as yml:
cdict = yaml.load(yml, Loader = yaml.FullLoader)
Ldict = {key : cdict[key] for key in cdict if (key in LikelihoodFg.__init__.__code__.co_varnames)}
Lclass = LikelihoodFg(**Ldict)
Gdict = {key : cdict[key] for key in cdict if (key in GibbsSampler.__init__.__code__.co_varnames)}
Nstep = 30000
gibbs_sampler = GibbsSampler(Nstep, Lclass, args.resume, **Gdict)
gibbs_sampler.run()
def get_lag(id, kmax, accepted):
a = accepted[:, id]
a_avg = np.mean(a)
N = len(a)
denom = np.sum((a-a_avg)**2)
lag = []
for k in range(1, kmax + 1):
num = 0
for i in range(0, N - k):
num += (a[i] - a_avg) * (a[i+k]-a_avg)
lag.append(num)
lag = np.array(lag)
def complex_mcmc(self, sentence):
self.gibbs_instance = GibbsSampler(self.trainer_builder)
return self.gibbs_instance.do_gibbs(sentence)
from gibbs import GibbsSampler
from csb.io import dump, load
def create_posterior(filename):
with open(filename) as script:
exec(script)
return posterior
n_particles = (100, 200)[0]
isdfile = 'posterior_betagal_{}.py'.format(n_particles)
posterior = create_posterior(isdfile)
diameter = posterior.priors[0].forcefield.d[0, 0]
gibbs = GibbsSampler(posterior, stepsize=1e-2)
samples = isd.Ensemble(posterior.params)
# run Gibbs sampler
with isd.take_time('Gibbs sampling'):
while len(samples) < 500:
with isd.take_time('Gibbs step'):
next(gibbs)
samples.update(posterior.params)
cc = []
for image in posterior.likelihoods:
cc.append(isd.crosscorr(image.data, image.mock.get()))
print(np.round([np.mean(cc), np.min(cc), np.max(cc)], 3) * 100)
if False:
# look at structure
'''
This is an example using 3 input ensembles Adenylate Kinase pdb_chainID: 1AKE_A, 4AKE_A, 1ANK_A
'''
import numpy as np
from gibbs import load_coordinates, GibbsSampler
from matplotlib import pyplot as plt
input_coordinate = load_coordinates('1AKE_A', '4AKE_A', '1ANK_A')
#Run Gibb sampler using 2 priors
gibb = GibbsSampler(input_coordinate, prior=2)
gibb.run(500)
print 'Number of Domain = ', np.unique(gibb.membership).shape[0]
print 'Membership', gibb.membership
print 'Log likelihood = ', gibb.log_likelihood
plt.plot(gibb.membership)
plt.title('Membership')
plt.show()
def main():
res = []
for width in range(1, 2):
for height in range(width, 3):
model_width = width * 2
model_height = height * 2
decimals = 2
temperature = 1
runs = 2
print("{} x {}".format(model_width, model_height))
# Create the model
builder = RandomBuilder(model_width, model_height)
model = builder.generate(decimals=decimals)
# Compute partition function of this model
bruteforcer = BruteforceSampler()
Z = bruteforcer.partition_function(model, temperature)
# First produce a graph with D-Wave
sampler = DWaveSampler()
results_dwave_prob = [
sampler.sample(model, 10000, temperature) for _ in range(runs)
]
embedding = builder.embedding_two()
results_dwave_two = [
sampler.sample(model, 10000, temperature, embedding=embedding)
for _ in range(runs)
]
embedding = builder.embedding_four()
results_dwave_four = [
sampler.sample(model, 10000, temperature, embedding=embedding)
for _ in range(runs)
]
sampler = GibbsSampler(n_variables=model_width * model_height)
results_gibbs = [
sampler.sample(model, 10000, temperature) for _ in range(runs)
]
compute_kl = lambda r_list: [
r.KL_divergence(Z, temperature) for r in r_list
]
kl_gibbs = compute_kl(results_gibbs)
kl_dwave_prob = compute_kl(results_dwave_prob)
kl_dwave_two = compute_kl(results_dwave_two)
kl_dwave_four = compute_kl(results_dwave_four)
print("KL divergence: {}".format(kl_gibbs))
print("KL divergence: {}".format(kl_dwave_prob))
print("KL divergence: {}".format(kl_dwave_two))
print("KL divergence: {}".format(kl_dwave_four))
res += [{
'type': 'gibbs',
'width': model_width,
'height': model_height,
'kl': kl
} for kl in kl_gibbs]
res += [{
'type': 'dwaveP',
'width': model_width,
'height': model_height,
'kl': kl
} for kl in kl_dwave_prob]
res += [{
'type': 'dwave2',
'width': model_width,
'height': model_height,
'kl': kl
} for kl in kl_dwave_two]
res += [{
'type': 'dwave4',
'width': model_width,
'height': model_height,
'kl': kl
} for kl in kl_dwave_four]
df = pandas.DataFrame.from_dict(res)
df.to_csv("data_kl_{}.csv".format(
datetime.datetime.now().strftime("%Y%m%d_%H%M%S")))
df.to_json("data_kl_{}.json".format(
datetime.datetime.now().strftime("%Y%m%d_%H%M%S")))
save_experiment("kl", {
'result': res,
})
def main():
### Parse command line options
usage = "usage: %prog [options] datafile"
cmdline_parser = OptionParser(usage=usage)
cmdline_parser.add_option('-o',
'--output',
dest='output_filename',
metavar='FILE',
default='results.pkl',
help='Serialize results to pickle FILE')
cmdline_parser.add_option('-m',
'--model',
dest='model',
metavar='FILE',
default='model',
help='Choose model to use')
cmdline_parser.add_option('-v',
'--verbose',
dest='loglevel',
default=logging.WARNING,
action='store_const',
const=logging.DEBUG,
help='Debug logging level mode')
cmdline_parser.add_option('-i',
'--iterations',
dest='iterations',
type='int',
default=25000,
help='Number of Gibbs iterations')
cmdline_parser.add_option('-b',
'--burnin',
dest='burnin',
type='int',
default=500,
help='Number of burn-in iterations')
cmdline_parser.add_option('-s',
'--subsample',
dest='subsample',
type='int',
default=10,
help='Subsample rate')
cmdline_parser.add_option('-t',
'--start',
dest='start',
type='int',
default=None,
help='start')
cmdline_parser.add_option('-e',
'--end',
dest='end',
type='int',
default=None,
help='end')
cmdline_parser.add_option('-g',
'--visualize',
dest='visualize',
action='store_true',
help='Visualize intermediate results')
cmdline_parser.add_option('-G',
'--visualize-priors',
dest='visualize_priors',
action='store_true',
help='Visualize prior distributions')
cmdline_parser.add_option(
'-p',
'--parameter-file',
dest='parameter_filename',
help='Use known parameters in file (i.e. simulated file).')
options, args = cmdline_parser.parse_args()
logging.basicConfig(level=options.loglevel,
format='%(asctime)s %(message)s')
data_filename = args[0]
gibbs_iters = options.iterations
burnin = options.burnin
subsample = options.subsample
model_module = __import__(options.model)
if options.parameter_filename is not None:
known_params = pickle.load(open(options.parameter_filename, 'rb'))
# Build model and load data
data = load_as_frame(data_filename, start=options.start, end=options.end)
model = model_module.define_model(data, known_params)
# Setup gibbs sampler
sampler = GibbsSampler(model=model,
iterations=gibbs_iters,
burnin=burnin,
subsample=subsample)
if options.visualize:
sampler.add_visualizer(model_module.visualize_gibbs)
if options.visualize_priors:
model_module.visualize_priors(model.priors)
# Begin sampling
start_time = time.time()
sampler.run()
gibbs_time = time.time() - start_time
print "Gibbs sampler ran %d.1 minutes" % (gibbs_time / 60.)
# Write out results
results = {}
results['options'] = options
results['variable_names'] = model.variable_names
results['known_params'] = model.known_params
results['hyper_params'] = model.hyper_params
results['filename'] = data_filename
results['data'] = data
results['gibbs_results'] = sampler.results()
pickle.dump(results, open(options.output_filename, 'wb'))
def main():
### Parse command line options
usage = "usage: %prog [options] datafile"
cmdline_parser = OptionParser(usage=usage)
cmdline_parser.add_option(
'-o', '--output', dest='output_filename', metavar='FILE',
default='results.pkl',
help='Serialize results to pickle FILE')
cmdline_parser.add_option(
'-m', '--model', dest='model', metavar='FILE',
default='model',
help='Choose model to use')
cmdline_parser.add_option(
'-v', '--verbose', dest='loglevel', default=logging.WARNING,
action='store_const', const=logging.DEBUG,
help='Debug logging level mode')
cmdline_parser.add_option(
'-i', '--iterations', dest='iterations',
type='int', default=25000,
help='Number of Gibbs iterations')
cmdline_parser.add_option(
'-b', '--burnin', dest='burnin',
type='int', default=500,
help='Number of burn-in iterations')
cmdline_parser.add_option(
'-s', '--subsample', dest='subsample',
type='int', default=10,
help='Subsample rate')
cmdline_parser.add_option(
'-t', '--start', dest='start',
type='int', default=None,
help='start')
cmdline_parser.add_option(
'-e', '--end', dest='end',
type='int', default=None,
help='end')
cmdline_parser.add_option(
'-g', '--visualize', dest='visualize',
action='store_true',
help='Visualize intermediate results')
cmdline_parser.add_option(
'-G', '--visualize-priors', dest='visualize_priors',
action='store_true',
help='Visualize prior distributions')
cmdline_parser.add_option(
'-p', '--parameter-file', dest='parameter_filename',
help='Use known parameters in file (i.e. simulated file).')
options, args = cmdline_parser.parse_args()
logging.basicConfig(level=options.loglevel, format='%(asctime)s %(message)s')
data_filename = args[0]
gibbs_iters = options.iterations
burnin = options.burnin
subsample = options.subsample
model_module = __import__(options.model)
if options.parameter_filename is not None:
known_params = pickle.load(open(options.parameter_filename, 'rb'))
# Build model and load data
data = load_as_frame(data_filename, start=options.start, end=options.end)
model = model_module.define_model(data, known_params)
# Setup gibbs sampler
sampler = GibbsSampler(
model=model,
iterations=gibbs_iters,
burnin=burnin,
subsample=subsample)
if options.visualize:
sampler.add_visualizer(model_module.visualize_gibbs)
if options.visualize_priors:
model_module.visualize_priors(model.priors)
# Begin sampling
start_time = time.time()
sampler.run()
gibbs_time = time.time() - start_time
print "Gibbs sampler ran %d.1 minutes" % (gibbs_time/60.)
# Write out results
results = {}
results['options'] = options
results['variable_names'] = model.variable_names
results['known_params'] = model.known_params
results['hyper_params'] = model.hyper_params
results['filename'] = data_filename
results['data'] = data
results['gibbs_results'] = sampler.results()
pickle.dump(results, open(options.output_filename, 'wb'))