class Inference(object):
'''
Represents a super class for all inference methods.
Also provides some convenience methods for collecting statistics
about the inference process and nicely outputting results.
:param mrf: the MRF inference is being applied to.
:param queries: a query or list of queries, can be either instances of
:class:`pracmln.logic.common.Logic` or string representations of them,
or predicate names that get expanded to all of their ground atoms.
If `ALL`, all ground atoms are subject to inference.
Additional keyword parameters:
:param cw: (bool) if `True`, the closed-world assumption will be applied
to all but the query atoms.
'''
def __init__(self, mrf, queries=ALL, **params):
self.mrf = mrf
self.mln = mrf.mln
self._params = edict(params)
if not queries:
self.queries = [
self.mln.logic.gnd_lit(ga, negated=False, mln=self.mln)
for ga in self.mrf.gndatoms
if self.mrf.evidence[ga.idx] is None
]
else:
# check for single/multiple query and expand
if type(queries) is not list:
queries = [queries]
self.queries = self._expand_queries(queries)
# fill in the missing truth values of variables that have only one remaining value
for variable in self.mrf.variables:
if variable.valuecount(self.mrf.evidence_dicti(
)) == 1: # the var is fully determined by the evidence
for _, value in variable.itervalues(self.mrf.evidence):
break
self.mrf.set_evidence(variable.value2dict(value), erase=False)
# apply the closed world assumptions to the explicitly specified predicates
if self.cwpreds:
for pred in self.cwpreds:
if isinstance(self.mln.predicate(pred),
SoftFunctionalPredicate):
if self.verbose:
logger.warning(
'Closed world assumption will be applied to soft functional predicate %s'
% pred)
elif isinstance(self.mln.predicate(pred), FunctionalPredicate):
raise Exception(
'Closed world assumption is inapplicable to functional predicate %s'
% pred)
for gndatom in self.mrf.gndatoms:
if gndatom.predname != pred: continue
if self.mrf.evidence[gndatom.idx] is None:
self.mrf.evidence[gndatom.idx] = 0
# apply the closed world assumption to all remaining ground atoms that are not in the queries
if self.closedworld:
qpreds = set()
for q in self.queries:
qpreds.update(q.prednames())
for gndatom in self.mrf.gndatoms:
if isinstance(self.mln.predicate(gndatom.predname), FunctionalPredicate) \
or isinstance(self.mln.predicate(gndatom.predname), SoftFunctionalPredicate):
continue
if gndatom.predname not in qpreds and self.mrf.evidence[
gndatom.idx] is None:
self.mrf.evidence[gndatom.idx] = 0
for var in self.mrf.variables:
if isinstance(var, FuzzyVariable):
var.consistent(self.mrf.evidence, strict=True)
self._watch = StopWatch()
@property
def verbose(self):
return self._params.get('verbose', False)
@property
def results(self):
if self._results is None:
raise Exception('No results available. Run the inference first.')
else:
return self._results
@property
def elapsedtime(self):
return self._watch['inference'].elapsedtime
@property
def multicore(self):
return self._params.get('multicore')
@property
def resultdb(self):
if '_resultdb' in self.__dict__:
return self._resultdb
db = Database(self.mrf.mln)
for atom in sorted(self.results, key=str):
db[str(atom)] = self.results[atom]
return db
@property
def closedworld(self):
return self._params.get('cw', False)
@property
def cwpreds(self):
return self._params.get('cw_preds', [])
def _expand_queries(self, queries):
'''
Expands the list of queries where necessary, e.g. queries that are
just predicate names are expanded to the corresponding list of atoms.
'''
equeries = []
for query in queries:
if type(query) == str:
prevLen = len(equeries)
if '(' in query: # a fully or partially grounded formula
f = self.mln.logic.parse_formula(query)
for gf in f.itergroundings(self.mrf):
equeries.append(gf)
else: # just a predicate name
if query not in self.mln.prednames:
raise NoSuchPredicateError(
'Unsupported query: %s is not among the admissible predicates.'
% (query))
continue
for gndatom in self.mln.predicate(query).groundatoms(
self.mln, self.mrf.domains):
equeries.append(
self.mln.logic.gnd_lit(self.mrf.gndatom(gndatom),
negated=False,
mln=self.mln))
if len(equeries) - prevLen == 0:
raise Exception(
"String query '%s' could not be expanded." % query)
elif isinstance(query, Logic.Formula):
equeries.append(query)
else:
raise Exception("Received query of unsupported type '%s'" %
str(type(query)))
return equeries
def _run(self):
raise Exception('%s does not implement _run()' %
self.__class__.__name__)
def run(self):
'''
Starts the inference process.
'''
# perform actual inference (polymorphic)
if self.verbose: print 'Inference engine: %s' % self.__class__.__name__
self._watch.tag('inference', verbose=self.verbose)
_weights_backup = list(self.mln.weights)
self._results = self._run()
self.mln.weights = _weights_backup
self._watch.finish('inference')
return self
def write(self,
stream=sys.stdout,
color=None,
sort='prob',
group=True,
reverse=True):
barwidth = 30
if tty(stream) and color is None:
color = 'yellow'
if sort not in ('alpha', 'prob'):
raise Exception('Unknown sorting: %s' % sort)
results = dict(self.results)
if group:
for var in sorted(self.mrf.variables, key=str):
res = dict([(atom, prob) for atom, prob in results.iteritems()
if atom in map(str, var.gndatoms)])
if not res: continue
if isinstance(var, MutexVariable) or isinstance(
var, SoftMutexVariable):
stream.write('%s:\n' % var)
if sort == 'prob':
res = sorted(res,
key=self.results.__getitem__,
reverse=reverse)
elif sort == 'alpha':
res = sorted(res, key=str)
for atom in res:
stream.write('%s %s\n' % (barstr(
barwidth, self.results[atom], color=color), atom))
return
# first sort wrt to probability
results = sorted(results,
key=self.results.__getitem__,
reverse=reverse)
# then wrt gnd atoms
results = sorted(results, key=str)
for q in results:
stream.write('%s %s\n' %
(barstr(barwidth, self.results[q], color=color), q))
self._watch.printSteps()
def write_elapsed_time(self, stream=sys.stdout, color=None):
if stream is sys.stdout and color is None:
color = True
elif color is None:
color = False
if color: col = 'blue'
else: col = None
total = float(self._watch['inference'].elapsedtime)
stream.write(headline('INFERENCE RUNTIME STATISTICS'))
print
self._watch.finish()
for t in sorted(self._watch.tags.values(),
key=lambda t: t.elapsedtime,
reverse=True):
stream.write(
'%s %s %s\n' %
(barstr(width=30, percent=t.elapsedtime / total,
color=col), elapsed_time_str(t.elapsedtime), t.label))
def run(self):
'''
Run the MLN learning with the given parameters.
'''
# load the MLN
if isinstance(self.mln, MLN):
mln = self.mln
else:
raise Exception('No MLN specified')
# load the training databases
if type(self.db) is list and all(
[isinstance(e, Database) for e in self.db]):
dbs = self.db
elif isinstance(self.db, Database):
dbs = [self.db]
elif isinstance(self.db, str):
db = self.db
if db is None or not db:
raise Exception('no trainig data given!')
dbpaths = [os.path.join(self.directory, 'db', db)]
dbs = []
for p in dbpaths:
dbs.extend(Database.load(mln, p, self.ignore_unknown_preds))
else:
raise Exception(
'Unexpected type of training databases: %s' % type(self.db))
if self.verbose:
print(('loaded %d database(s).' % len(dbs)))
watch = StopWatch()
if self.verbose:
confg = dict(self._config)
confg.update(eval("dict(%s)" % self.params))
if type(confg.get('db', None)) is list:
confg['db'] = '%d Databases' % len(confg['db'])
print((tabulate(
sorted(list(confg.items()), key=lambda key_v: str(key_v[0])),
headers=('Parameter:', 'Value:'))))
params = dict([(k, getattr(self, k)) for k in (
'multicore', 'verbose', 'profile', 'ignore_zero_weight_formulas')])
# for discriminative learning
if issubclass(self.method, DiscriminativeLearner):
if self.discr_preds == QUERY_PREDS: # use query preds
params['qpreds'] = self.qpreds
elif self.discr_preds == EVIDENCE_PREDS: # use evidence preds
params['epreds'] = self.epreds
# gaussian prior settings
if self.use_prior:
params['prior_mean'] = self.prior_mean
params['prior_stdev'] = self.prior_stdev
# expand the parameters
params.update(self.params)
if self.profile:
prof = Profile()
print('starting profiler...')
prof.enable()
else:
prof = None
# set the debug level
olddebug = logger.level
logger.level = eval('logs.%s' % params.get('debug', 'WARNING').upper())
mlnlearnt = None
try:
# run the learner
mlnlearnt = mln.learn(dbs, self.method, **params)
if self.verbose:
print()
print(headline('LEARNT MARKOV LOGIC NETWORK'))
print()
mlnlearnt.write()
except SystemExit:
print('Cancelled...')
finally:
if self.profile:
prof.disable()
print(headline('PROFILER STATISTICS'))
ps = pstats.Stats(prof, stream=sys.stdout).sort_stats(
'cumulative')
ps.print_stats()
# reset the debug level
logger.level = olddebug
print()
watch.finish()
watch.printSteps()
return mlnlearnt
def run(self):
watch = StopWatch()
watch.tag('inference', self.verbose)
# load the MLN
if isinstance(self.mln, MLN):
mln = self.mln
else:
raise Exception('No MLN specified')
if self.use_emln and self.emln is not None:
mlnstrio = io.StringIO()
mln.write(mlnstrio)
mlnstr = mlnstrio.getvalue()
mlnstrio.close()
emln = self.emln
mln = parse_mln(mlnstr + emln,
grammar=self.grammar,
logic=self.logic)
# load the database
if isinstance(self.db, Database):
db = self.db
elif isinstance(self.db, list) and len(self.db) == 1:
db = self.db[0]
elif isinstance(self.db, list) and len(self.db) == 0:
db = Database(mln)
elif isinstance(self.db, list):
raise Exception(
'Got {} dbs. Can only handle one for inference.'.format(
len(self.db)))
else:
raise Exception('DB of invalid format {}'.format(type(self.db)))
# expand the
# parameters
params = dict(self._config)
if 'params' in params:
params.update(eval("dict(%s)" % params['params']))
del params['params']
params['verbose'] = self.verbose
if self.verbose:
print((tabulate(sorted(list(params.items()),
key=lambda k_v: str(k_v[0])),
headers=('Parameter:', 'Value:'))))
if type(db) is list and len(db) > 1:
raise Exception('Inference can only handle one database at a time')
elif type(db) is list:
db = db[0]
params['cw_preds'] = [x for x in self.cw_preds if bool(x)]
# extract and remove all non-algorithm
for s in GUI_SETTINGS:
if s in params: del params[s]
if self.profile:
prof = Profile()
print('starting profiler...')
prof.enable()
# set the debug level
olddebug = logger.level
logger.level = (eval('logs.%s' %
params.get('debug', 'WARNING').upper()))
result = None
try:
mln_ = mln.materialize(db)
mrf = mln_.ground(db)
inference = self.method(mrf, self.queries, **params)
if self.verbose:
print()
print((headline('EVIDENCE VARIABLES')))
print()
mrf.print_evidence_vars()
result = inference.run()
if self.verbose:
print()
print((headline('INFERENCE RESULTS')))
print()
inference.write()
if self.verbose:
print()
inference.write_elapsed_time()
except SystemExit:
traceback.print_exc()
print('Cancelled...')
finally:
if self.profile:
prof.disable()
print((headline('PROFILER STATISTICS')))
ps = pstats.Stats(prof,
stream=sys.stdout).sort_stats('cumulative')
ps.print_stats()
# reset the debug level
logger.level = olddebug
if self.verbose:
print()
watch.finish()
watch.printSteps()
return result
class MultipleDatabaseLearner(AbstractLearner):
'''
Learns from multiple databases using an arbitrary sub-learning method for
each database, assuming independence between individual databases.
'''
def __init__(self, mln_, dbs, method, **params):
'''
:param dbs: list of :class:`mln.database.Database` objects to
be used for learning.
:param mln_: the MLN object to be used for learning
:param method: the algorithm to be used for learning. Must be a
class provided by
:class:`mln.methods.LearningMethods`.
:param **params: additional parameters handed over to the base
learners.
'''
self.dbs = dbs
self._params = edict(params)
if not mln_._materialized:
self.mln = mln_.materialize(*dbs)
else:
self.mln = mln_
self.watch = StopWatch()
self.learners = [None] * len(dbs)
self.watch.tag('setup learners', verbose=self.verbose)
if self.verbose:
bar = ProgressBar(steps=len(dbs), color='green')
if self.multicore:
pool = Pool(maxtasksperchild=1)
logger.debug(
'Setting up multi-core processing for {} cores'.format(
pool._processes))
try:
for i, learner in pool.imap(with_tracing(_setup_learner),
self._iterdbs(method)):
self.learners[i] = learner
if self.verbose:
bar.label('Database %d, %s' % ((i + 1), learner.name))
bar.inc()
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
# as MLNs and formulas have been copied to the separate processes,
# the mln pointers of the formulas now point to the MLNs in these child processes
# we have to copy the materialized weight back to our parent process
self.mln.weights = list(first(self.learners).mrf.mln.weights)
else:
for i, db in enumerate(self.dbs):
_, learner = _setup_learner(
(i, self.mln, db, method, self._params + {
'multicore': False
}))
self.learners[i] = learner
if self.verbose:
bar.label('Database %d, %s' % ((i + 1), learner.name))
bar.inc()
if self.verbose:
print 'set up', self.name
self.watch.finish('setup learners')
def _iterdbs(self, method):
for i, db in enumerate(self.dbs):
yield i, self.mln, db, method, self._params + {
'verbose': not self.multicore,
'multicore': False
}
@property
def name(self):
return "MultipleDatabaseLearner [{} x {}]".format(
len(self.learners), self.learners[0].name)
def _f(self, w):
# it turned out that it doesn't pay off to evaluate the function
# in separate processes, so we turn it off
if False: # self.multicore:
likelihood = 0
pool = Pool()
try:
for i, (f_, d_) in enumerate(
pool.imap(
with_tracing(_methodcaller('_f',
sideeffects=True)),
map(lambda l: (l, w), self.learners))):
self.learners[i].__dict__ = d_
likelihood += f_
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
return likelihood
else:
return sum(map(lambda l: l._f(w), self.learners))
def _grad(self, w):
grad = numpy.zeros(len(self.mln.formulas), numpy.float64)
if False: # self.multicore:
# it turned out that it doesn't pay off to evaluate the gradient
# in separate processes, so we turn it off
pool = Pool()
try:
for i, (grad_, d_) in enumerate(
pool.imap(
with_tracing(
_methodcaller('_grad', sideeffects=True)),
map(lambda l: (l, w), self.learners))):
self.learners[i].__dict__ = d_
grad += grad_
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for learner in self.learners:
grad += learner._grad(w)
return grad
def _hessian(self, w):
N = len(self.mln.formulas)
hessian = numpy.matrix(numpy.zeros((N, N)))
if self.multicore:
pool = Pool()
try:
for h in pool.imap(with_tracing(_methodcaller('_hessian')),
map(lambda l: (l, w), self.learners)):
hessian += h
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for learner in self.learners:
hessian += learner._hessian(w)
return hessian
def _prepare(self):
self.watch.tag('preparing optimization', verbose=self.verbose)
if self.verbose:
bar = ProgressBar(steps=len(self.dbs), color='green')
if self.multicore:
pool = Pool(maxtasksperchild=1)
try:
for i, (_, d_) in enumerate(
pool.imap(
with_tracing(
_methodcaller('_prepare', sideeffects=True)),
self.learners)):
checkmem()
self.learners[i].__dict__ = d_
if self.verbose: bar.inc()
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for learner in self.learners:
checkmem()
learner._prepare()
if self.verbose: bar.inc()
def _filter_fixweights(self, v):
'''
Removes from the vector `v` all elements at indices that correspond to
a fixed weight formula index.
'''
if len(v) != len(self.mln.formulas):
raise Exception('Vector must have same length as formula weights')
return [
v[i] for i in range(len(self.mln.formulas))
if not self.mln.fixweights[i] and self.mln.weights[i] != HARD
]
def _add_fixweights(self, w):
i = 0
w_ = []
for f in self.mln.formulas:
if self.mln.fixweights[f.idx] or f.weight == HARD:
w_.append(self._w[f.idx])
else:
w_.append(w[i])
i += 1
return w_
def run(self, **params):
if 'scipy' not in sys.modules:
raise Exception("Scipy was not imported! Install numpy and scipy "
"if you want to use weight learning.")
runs = 0
self._w = [0] * len(self.mln.formulas)
while runs < self.maxrepeat:
self._prepare()
# initial parameter vector: all zeros or weights from formulas
for f in self.mln.formulas:
if self.mln.fixweights[
f.idx] or self.use_init_weights or f.ishard:
self._w[f.idx] = f.weight
self._optimize(**self._params)
self._cleanup()
runs += 1
if not any([l.repeat() for l in self.learners]): break
return self.weights
def run(self):
'''
Run the MLN learning with the given parameters.
'''
# load the MLN
if isinstance(self.mln, MLN):
mln = self.mln
else:
raise Exception('No MLN specified')
# load the training databases
if type(self.db) is list and all(
map(lambda e: isinstance(e, Database), self.db)):
dbs = self.db
elif isinstance(self.db, Database):
dbs = [self.db]
elif isinstance(self.db, basestring):
db = self.db
if db is None or not db:
raise Exception('no trainig data given!')
dbpaths = [os.path.join(self.directory, 'db', db)]
dbs = []
for p in dbpaths:
dbs.extend(Database.load(mln, p, self.ignore_unknown_preds))
else:
raise Exception(
'Unexpected type of training databases: %s' % type(self.db))
if self.verbose:
print 'loaded %d database(s).' % len(dbs)
watch = StopWatch()
if self.verbose:
confg = dict(self._config)
confg.update(eval("dict(%s)" % self.params))
if type(confg.get('db', None)) is list:
confg['db'] = '%d Databases' % len(confg['db'])
print tabulate(
sorted(list(confg.viewitems()), key=lambda (key, v): str(key)),
headers=('Parameter:', 'Value:'))
params = dict([(k, getattr(self, k)) for k in (
'multicore', 'verbose', 'profile', 'ignore_zero_weight_formulas')])
# for discriminative learning
if issubclass(self.method, DiscriminativeLearner):
if self.discr_preds == QUERY_PREDS: # use query preds
params['qpreds'] = self.qpreds
elif self.discr_preds == EVIDENCE_PREDS: # use evidence preds
params['epreds'] = self.epreds
# gaussian prior settings
if self.use_prior:
params['prior_mean'] = self.prior_mean
params['prior_stdev'] = self.prior_stdev
# expand the parameters
params.update(self.params)
if self.profile:
prof = Profile()
print 'starting profiler...'
prof.enable()
else:
prof = None
# set the debug level
olddebug = praclog.level()
praclog.level(
eval('logging.%s' % params.get('debug', 'WARNING').upper()))
mlnlearnt = None
try:
# run the learner
mlnlearnt = mln.learn(dbs, self.method, **params)
if self.verbose:
print
print headline('LEARNT MARKOV LOGIC NETWORK')
print
mlnlearnt.write()
except SystemExit:
print 'Cancelled...'
finally:
if self.profile:
prof.disable()
print headline('PROFILER STATISTICS')
ps = pstats.Stats(prof, stream=sys.stdout).sort_stats(
'cumulative')
ps.print_stats()
# reset the debug level
praclog.level(olddebug)
print
watch.finish()
watch.printSteps()
return mlnlearnt
class MultipleDatabaseLearner(AbstractLearner):
'''
Learns from multiple databases using an arbitrary sub-learning method for
each database, assuming independence between individual databases.
'''
def __init__(self, mln_, dbs, method, **params):
'''
:param dbs: list of :class:`mln.database.Database` objects to
be used for learning.
:param mln_: the MLN object to be used for learning
:param method: the algorithm to be used for learning. Must be a
class provided by
:class:`mln.methods.LearningMethods`.
:param **params: additional parameters handed over to the base
learners.
'''
self.dbs = dbs
self._params = edict(params)
if not mln_._materialized:
self.mln = mln_.materialize(*dbs)
else:
self.mln = mln_
self.watch = StopWatch()
self.learners = [None] * len(dbs)
self.watch.tag('setup learners', verbose=self.verbose)
if self.verbose:
bar = ProgressBar(steps=len(dbs), color='green')
if self.multicore:
pool = Pool(maxtasksperchild=1)
logger.debug('Setting up multi-core processing for {} cores'.format(pool._processes))
try:
for i, learner in pool.imap(with_tracing(_setup_learner), self._iterdbs(method)):
self.learners[i] = learner
if self.verbose:
bar.label('Database %d, %s' % ((i + 1), learner.name))
bar.inc()
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for i, db in enumerate(self.dbs):
_, learner = _setup_learner((i, self.mln, db, method, self._params + {'multicore': False}))
self.learners[i] = learner
if self.verbose:
bar.label('Database %d, %s' % ((i + 1), learner.name))
bar.inc()
if self.verbose:
print 'set up', self.name
self.watch.finish('setup learners')
def _iterdbs(self, method):
for i, db in enumerate(self.dbs):
yield i, self.mln, db, method, self._params + {
'verbose': not self.multicore, 'multicore': False}
@property
def name(self):
return "MultipleDatabaseLearner [{} x {}]".format(len(self.learners), self.learners[0].name)
def _f(self, w):
# it turned out that it doesn't pay off to evaluate the function
# in separate processes, so we turn it off
if False: # self.multicore:
likelihood = 0
pool = Pool()
try:
for i, (f_, d_) in enumerate(pool.imap(with_tracing(_methodcaller('_f', sideeffects=True)), map(lambda l: (l, w), self.learners))):
self.learners[i].__dict__ = d_
likelihood += f_
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
return likelihood
else:
return sum(map(lambda l: l._f(w), self.learners))
def _grad(self, w):
grad = numpy.zeros(len(self.mln.formulas), numpy.float64)
if False: # self.multicore:
# it turned out that it doesn't pay off to evaluate the gradient
# in separate processes, so we turn it off
pool = Pool()
try:
for i, (grad_, d_) in enumerate(pool.imap(with_tracing(_methodcaller('_grad', sideeffects=True)), map(lambda l: (l, w), self.learners))):
self.learners[i].__dict__ = d_
grad += grad_
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for learner in self.learners: grad += learner._grad(w)
return grad
def _hessian(self, w):
N = len(self.mln.formulas)
hessian = numpy.matrix(numpy.zeros((N, N)))
if self.multicore:
pool = Pool()
try:
for h in pool.imap(with_tracing(_methodcaller('_hessian')), map(lambda l: (l, w), self.learners)):
hessian += h
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for learner in self.learners: hessian += learner._hessian(w)
return hessian
def _prepare(self):
self.watch.tag('preparing optimization', verbose=self.verbose)
if self.verbose:
bar = ProgressBar(steps=len(self.dbs), color='green')
if self.multicore:
pool = Pool(maxtasksperchild=1)
try:
for i, (_, d_) in enumerate(pool.imap(with_tracing(_methodcaller('_prepare', sideeffects=True)), self.learners)):
checkmem()
self.learners[i].__dict__ = d_
if self.verbose: bar.inc()
except Exception as e:
logger.error('Error in child process. Terminating pool...')
pool.close()
raise e
finally:
pool.terminate()
pool.join()
else:
for learner in self.learners:
checkmem()
learner._prepare()
if self.verbose: bar.inc()
def _filter_fixweights(self, v):
'''
Removes from the vector `v` all elements at indices that correspond to
a fixed weight formula index.
'''
if len(v) != len(self.mln.formulas):
raise Exception('Vector must have same length as formula weights')
return [v[i] for i in range(len(self.mln.formulas)) if not self.mln.fixweights[i] and self.mln.weights[i] != HARD]
def _add_fixweights(self, w):
i = 0
w_ = []
for f in self.mln.formulas:
if self.mln.fixweights[f.idx] or f.weight == HARD:
w_.append(self._w[f.idx])
else:
w_.append(w[i])
i += 1
return w_
def run(self, **params):
if 'scipy' not in sys.modules:
raise Exception("Scipy was not imported! Install numpy and scipy "
"if you want to use weight learning.")
# initial parameter vector: all zeros or weights from formulas
self._w = [0] * len(self.mln.formulas)
for f in self.mln.formulas:
if self.mln.fixweights[f.idx] or self.use_init_weights or f.ishard:
self._w[f.idx] = f.weight
runs = 0
while runs < self.maxrepeat:
self._prepare()
self._optimize(**self._params)
self._cleanup()
runs += 1
if not any([l.repeat() for l in self.learners]): break
return self.weights
def run(self):
watch = StopWatch()
watch.tag('inference', self.verbose)
# load the MLN
if isinstance(self.mln, MLN):
mln = self.mln
else:
raise Exception('No MLN specified')
if self.use_emln and self.emln is not None:
mlnstr = StringIO.StringIO()
mln.write(mlnstr)
mlnstr.close()
mlnstr = str(mlnstr)
emln = self.emln
mln = parse_mln(mlnstr + emln,
grammar=self.grammar,
logic=self.logic)
# load the database
if isinstance(self.db, Database):
db = self.db
elif isinstance(self.db, list) and len(self.db) == 1:
db = self.db[0]
elif isinstance(self.db, list):
raise Exception(
'Got {} dbs. Can only handle one for inference.'.format(
len(self.db)))
else:
raise Exception('DB of invalid format {}'.format(type(self.db)))
# expand the
# parameters
params = dict(self._config)
if 'params' in params:
params.update(eval("dict(%s)" % params['params']))
del params['params']
if self.verbose:
print tabulate(sorted(list(params.viewitems()),
key=lambda (k, v): str(k)),
headers=('Parameter:', 'Value:'))
# create the MLN and evidence database and the parse the queries
# mln = parse_mln(modelstr, searchPath=self.dir.get(), logic=self.config['logic'], grammar=self.config['grammar'])
# db = parse_db(mln, db_content, ignore_unknown_preds=params.get('ignore_unknown_preds', False))
if type(db) is list and len(db) > 1:
raise Exception('Inference can only handle one database at a time')
elif type(db) is list:
db = db[0]
# parse non-atomic params
# if type(self.queries) is not list:
# queries = parse_queries(mln, str(self.queries))
params['cw_preds'] = filter(lambda x: bool(x), self.cw_preds)
# extract and remove all non-algorithm
for s in GUI_SETTINGS:
if s in params: del params[s]
if self.profile:
prof = Profile()
print 'starting profiler...'
prof.enable()
# set the debug level
olddebug = praclog.level()
praclog.level(
eval('logging.%s' % params.get('debug', 'WARNING').upper()))
result = None
try:
mln_ = mln.materialize(db)
mrf = mln_.ground(db)
inference = self.method(mrf, self.queries, **params)
if self.verbose:
print
print headline('EVIDENCE VARIABLES')
print
mrf.print_evidence_vars()
result = inference.run()
if self.verbose:
print
print headline('INFERENCE RESULTS')
print
inference.write()
if self.verbose:
print
inference.write_elapsed_time()
except SystemExit:
print 'Cancelled...'
finally:
if self.profile:
prof.disable()
print headline('PROFILER STATISTICS')
ps = pstats.Stats(prof,
stream=sys.stdout).sort_stats('cumulative')
ps.print_stats()
# reset the debug level
praclog.level(olddebug)
if self.verbose:
print
watch.finish()
watch.printSteps()
return result
class DefaultGroundingFactory:
'''
Implementation of the default grounding algorithm, which
creates ALL ground atoms and ALL ground formulas.
:param simplify: if `True`, the formula will be simplified according to the
evidence given.
:param unsatfailure: raises a :class:`mln.errors.SatisfiabilityException` if a
hard logical constraint is violated by the evidence.
'''
def __init__(self,
mrf,
simplify=False,
unsatfailure=False,
formulas=None,
cache=auto,
**params):
self.mrf = mrf
self.formulas = ifNone(formulas, list(self.mrf.formulas))
self.total_gf = 0
for f in self.formulas:
self.total_gf += f.countgroundings(self.mrf)
self.grounder = None
self._cachesize = CACHE_SIZE if cache is auto else cache
self._cache = None
self.__cacheinit = False
self.__cachecomplete = False
self._params = params
self.watch = StopWatch()
self.simplify = simplify
self.unsatfailure = unsatfailure
@property
def verbose(self):
return self._params.get('verbose', False)
@property
def multicore(self):
return self._params.get('multicore', False)
@property
def iscached(self):
return self._cache is not None and self.__cacheinit
@property
def usecache(self):
return self._cachesize is not None and self._cachesize > 0
def _cacheinit(self):
if False: #self.total_gf > self._cachesize:
logger.warning(
'Number of formula groundings (%d) exceeds cache size (%d). Caching is disabled.'
% (self.total_gf, self._cachesize))
else:
self._cache = []
self.__cacheinit = True
def itergroundings(self):
'''
Iterates over all formula groundings.
'''
self.watch.tag('grounding', verbose=self.verbose)
if self.grounder is None:
self.grounder = iter(
self._itergroundings(simplify=self.simplify,
unsatfailure=self.unsatfailure))
if self.usecache and not self.iscached:
self._cacheinit()
counter = -1
while True:
counter += 1
if self.iscached and len(self._cache) > counter:
yield self._cache[counter]
elif not self.__cachecomplete:
try:
gf = self.grounder.next()
except StopIteration:
self.__cachecomplete = True
return
else:
if self._cache is not None:
self._cache.append(gf)
yield gf
else:
return
self.watch.finish('grounding')
if self.verbose: print
def _itergroundings(self, simplify=False, unsatfailure=False):
if self.verbose:
bar = ProgressBar(width=100, color='green')
for i, formula in enumerate(self.formulas):
if self.verbose: bar.update((i + 1) / float(len(self.formulas)))
for gndformula in formula.itergroundings(self.mrf,
simplify=simplify):
if unsatfailure and gndformula.weight == HARD and gndformula(
self.mrf.evidence) != 1:
print
gndformula.print_structure(self.mrf.evidence)
raise SatisfiabilityException(
'MLN is unsatisfiable due to hard constraint violation %s (see above)'
% self.mrf.formulas[gndformula.idx])
yield gndformula
class DefaultGroundingFactory:
'''
Implementation of the default grounding algorithm, which
creates ALL ground atoms and ALL ground formulas.
:param simplify: if `True`, the formula will be simplified according to the
evidence given.
:param unsatfailure: raises a :class:`mln.errors.SatisfiabilityException` if a
hard logical constraint is violated by the evidence.
'''
def __init__(self, mrf, simplify=False, unsatfailure=False, formulas=None, cache=auto, **params):
self.mrf = mrf
self.formulas = ifNone(formulas, list(self.mrf.formulas))
self.total_gf = 0
for f in self.formulas:
self.total_gf += f.countgroundings(self.mrf)
self.grounder = None
self._cachesize = CACHE_SIZE if cache is auto else cache
self._cache = None
self.__cacheinit = False
self.__cachecomplete = False
self._params = params
self.watch = StopWatch()
self.simplify = simplify
self.unsatfailure = unsatfailure
@property
def verbose(self):
return self._params.get('verbose', False)
@property
def multicore(self):
return self._params.get('multicore', False)
@property
def iscached(self):
return self._cache is not None and self.__cacheinit
@property
def usecache(self):
return self._cachesize is not None and self._cachesize > 0
def _cacheinit(self):
if False:#self.total_gf > self._cachesize:
logger.warning('Number of formula groundings (%d) exceeds cache size (%d). Caching is disabled.' % (self.total_gf, self._cachesize))
else:
self._cache = []
self.__cacheinit = True
def itergroundings(self):
'''
Iterates over all formula groundings.
'''
self.watch.tag('grounding', verbose=self.verbose)
if self.grounder is None:
self.grounder = iter(self._itergroundings(simplify=self.simplify, unsatfailure=self.unsatfailure))
if self.usecache and not self.iscached:
self._cacheinit()
counter = -1
while True:
counter += 1
if self.iscached and len(self._cache) > counter:
yield self._cache[counter]
elif not self.__cachecomplete:
try:
gf = self.grounder.next()
except StopIteration:
self.__cachecomplete = True
return
else:
if self._cache is not None:
self._cache.append(gf)
yield gf
else: return
self.watch.finish('grounding')
if self.verbose: print
def _itergroundings(self, simplify=False, unsatfailure=False):
if self.verbose:
bar = ProgressBar(width=100, color='green')
for i, formula in enumerate(self.formulas):
if self.verbose: bar.update((i+1) / float(len(self.formulas)))
for gndformula in formula.itergroundings(self.mrf, simplify=simplify):
if unsatfailure and gndformula.weight == HARD and gndformula(self.mrf.evidence) != 1:
print
gndformula.print_structure(self.mrf.evidence)
raise SatisfiabilityException('MLN is unsatisfiable due to hard constraint violation %s (see above)' % self.mrf.formulas[gndformula.idx])
yield gndformula
class Inference(object):
'''
Represents a super class for all inference methods.
Also provides some convenience methods for collecting statistics
about the inference process and nicely outputting results.
:param mrf: the MRF inference is being applied to.
:param queries: a query or list of queries, can be either instances of
:class:`pracmln.logic.common.Logic` or string representations of them,
or predicate names that get expanded to all of their ground atoms.
If `ALL`, all ground atoms are subject to inference.
Additional keyword parameters:
:param cw: (bool) if `True`, the closed-world assumption will be applied
to all but the query atoms.
'''
def __init__(self, mrf, queries=ALL, **params):
self.mrf = mrf
self.mln = mrf.mln
self._params = edict(params)
if not queries:
self.queries = [self.mln.logic.gnd_lit(ga, negated=False, mln=self.mln) for ga in self.mrf.gndatoms if self.mrf.evidence[ga.idx] is None]
else:
# check for single/multiple query and expand
if type(queries) is not list:
queries = [queries]
self.queries = self._expand_queries(queries)
# fill in the missing truth values of variables that have only one remaining value
for variable in self.mrf.variables:
if variable.valuecount(self.mrf.evidence_dicti()) == 1: # the var is fully determined by the evidence
for _, value in variable.itervalues(self.mrf.evidence): break
self.mrf.set_evidence(variable.value2dict(value), erase=False)
# apply the closed world assumptions to the explicitly specified predicates
if self.cwpreds:
for pred in self.cwpreds:
if isinstance(self.mln.predicate(pred), SoftFunctionalPredicate):
if self.verbose: logger.warning('Closed world assumption will be applied to soft functional predicate %s' % pred)
elif isinstance(self.mln.predicate(pred), FunctionalPredicate):
raise Exception('Closed world assumption is inapplicable to functional predicate %s' % pred)
for gndatom in self.mrf.gndatoms:
if gndatom.predname != pred: continue
if self.mrf.evidence[gndatom.idx] is None:
self.mrf.evidence[gndatom.idx] = 0
# apply the closed world assumption to all remaining ground atoms that are not in the queries
if self.closedworld:
qpreds = set()
for q in self.queries:
qpreds.update(q.prednames())
for gndatom in self.mrf.gndatoms:
if isinstance(self.mln.predicate(gndatom.predname), FunctionalPredicate) \
or isinstance(self.mln.predicate(gndatom.predname), SoftFunctionalPredicate):
continue
if gndatom.predname not in qpreds and self.mrf.evidence[gndatom.idx] is None:
self.mrf.evidence[gndatom.idx] = 0
for var in self.mrf.variables:
if isinstance(var, FuzzyVariable):
var.consistent(self.mrf.evidence, strict=True)
self._watch = StopWatch()
@property
def verbose(self):
return self._params.get('verbose', False)
@property
def results(self):
if self._results is None:
raise Exception('No results available. Run the inference first.')
else:
return self._results
@property
def elapsedtime(self):
return self._watch['inference'].elapsedtime
@property
def multicore(self):
return self._params.get('multicore')
@property
def resultdb(self):
db = Database(self.mrf.mln)
for atom in sorted(self.results, key=str):
db[str(atom)] = self.results[atom]
return db
@property
def closedworld(self):
return self._params.get('cw', False)
@property
def cwpreds(self):
return self._params.get('cw_preds', [])
def _expand_queries(self, queries):
'''
Expands the list of queries where necessary, e.g. queries that are
just predicate names are expanded to the corresponding list of atoms.
'''
equeries = []
for query in queries:
if type(query) == str:
prevLen = len(equeries)
if '(' in query: # a fully or partially grounded formula
f = self.mln.logic.parse_formula(query)
for gf in f.itergroundings(self.mrf):
equeries.append(gf)
else: # just a predicate name
if query not in self.mln.prednames:
raise NoSuchPredicateError('Unsupported query: %s is not among the admissible predicates.' % (query))
continue
for gndatom in self.mln.predicate(query).groundatoms(self.mln, self.mrf.domains):
equeries.append(self.mln.logic.gnd_lit(self.mrf.gndatom(gndatom), negated=False, mln=self.mln))
if len(equeries) - prevLen == 0:
raise Exception("String query '%s' could not be expanded." % query)
elif isinstance(query, Logic.Formula):
equeries.append(query)
else:
raise Exception("Received query of unsupported type '%s'" % str(type(query)))
return equeries
def _run(self):
raise Exception('%s does not implement _run()' % self.__class__.__name__)
def run(self):
'''
Starts the inference process.
'''
# perform actual inference (polymorphic)
if self.verbose: print 'Inference engine: %s' % self.__class__.__name__
self._watch.tag('inference', verbose=self.verbose)
_weights_backup = list(self.mln.weights)
self._results = self._run()
self.mln.weights = _weights_backup
self._watch.finish('inference')
return self
def write(self, stream=sys.stdout, color=None, sort='prob', group=True, reverse=True):
barwidth = 30
if tty(stream) and color is None:
color = 'yellow'
if sort not in ('alpha', 'prob'):
raise Exception('Unknown sorting: %s' % sort)
results = dict(self.results)
if group:
for var in sorted(self.mrf.variables, key=str):
res = dict([(atom, prob) for atom, prob in results.iteritems() if atom in map(str, var.gndatoms)])
if not res: continue
if isinstance(var, MutexVariable) or isinstance(var, SoftMutexVariable):
stream.write('%s:\n' % var)
if sort == 'prob':
res = sorted(res, key=self.results.__getitem__, reverse=reverse)
elif sort == 'alpha':
res = sorted(res, key=str)
for atom in res:
stream.write('%s %s\n' % (barstr(barwidth, self.results[atom], color=color), atom))
return
# first sort wrt to probability
results = sorted(results, key=self.results.__getitem__, reverse=reverse)
# then wrt gnd atoms
results = sorted(results, key=str)
for q in results:
stream.write('%s %s\n' % (barstr(barwidth, self.results[q], color=color), q))
self._watch.printSteps()
def write_elapsed_time(self, stream=sys.stdout, color=None):
if stream is sys.stdout and color is None:
color = True
elif color is None:
color = False
if color: col = 'blue'
else: col = None
total = float(self._watch['inference'].elapsedtime)
stream.write(headline('INFERENCE RUNTIME STATISTICS'))
print
self._watch.finish()
for t in sorted(self._watch.tags.values(), key=lambda t: t.elapsedtime, reverse=True):
stream.write('%s %s %s\n' % (barstr(width=30, percent=t.elapsedtime / total, color=col), elapsed_time_str(t.elapsedtime), t.label))
def run(self):
watch = StopWatch()
watch.tag('inference', self.verbose)
# load the MLN
if isinstance(self.mln, MLN):
mln = self.mln
else:
raise Exception('No MLN specified')
if self.use_emln and self.emln is not None:
mlnstr = StringIO.StringIO()
mln.write(mlnstr)
mlnstr.close()
mlnstr = str(mlnstr)
emln = self.emln
mln = parse_mln(mlnstr + emln, grammar=self.grammar,
logic=self.logic)
# load the database
if isinstance(self.db, Database):
db = self.db
elif isinstance(self.db, list) and len(self.db) == 1:
db = self.db[0]
elif isinstance(self.db, list):
raise Exception(
'Got {} dbs. Can only handle one for inference.'.format(
len(self.db)))
else:
raise Exception('DB of invalid format {}'.format(type(self.db)))
# expand the
# parameters
params = dict(self._config)
if 'params' in params:
params.update(eval("dict(%s)" % params['params']))
del params['params']
if self.verbose:
print tabulate(sorted(list(params.viewitems()), key=lambda (k, v): str(k)), headers=('Parameter:', 'Value:'))
if type(db) is list and len(db) > 1:
raise Exception('Inference can only handle one database at a time')
elif type(db) is list:
db = db[0]
params['cw_preds'] = filter(lambda x: bool(x), self.cw_preds)
# extract and remove all non-algorithm
for s in GUI_SETTINGS:
if s in params: del params[s]
if self.profile:
prof = Profile()
print 'starting profiler...'
prof.enable()
# set the debug level
olddebug = praclog.level()
praclog.level(eval('logging.%s' % params.get('debug', 'WARNING').upper()))
result = None
try:
mln_ = mln.materialize(db)
mrf = mln_.ground(db)
inference = self.method(mrf, self.queries, **params)
if self.verbose:
print
print headline('EVIDENCE VARIABLES')
print
mrf.print_evidence_vars()
result = inference.run()
if self.verbose:
print
print headline('INFERENCE RESULTS')
print
inference.write()
if self.verbose:
print
inference.write_elapsed_time()
except SystemExit:
print 'Cancelled...'
finally:
if self.profile:
prof.disable()
print headline('PROFILER STATISTICS')
ps = pstats.Stats(prof, stream=sys.stdout).sort_stats('cumulative')
ps.print_stats()
# reset the debug level
praclog.level(olddebug)
if self.verbose:
print
watch.finish()
watch.printSteps()
return result