def __init__(self, child, groupby, aggregates, window):
QueryOperator.__init__(self)
self.child = child
self.can_query_stream_cache = self.can_query_stream_impl()
self.groupby = groupby
self.aggregates = aggregates
self.window = window
self.aggregator = Aggregator(self.aggregates, self.groupby,
self.window)
def __init__(self, child, groupby, aggregates, window):
QueryOperator.__init__(self)
self.child = child
self.can_query_stream_cache = self.can_query_stream_impl()
self.groupby = groupby
self.aggregates = aggregates
self.window = window
self.aggregator = Aggregator(self.aggregates, self.groupby, self.window)
class GroupBy(QueryOperator):
"""
Groups results from child by some set of fields, and runs the aggregate
function(s) over them, emitting results and resetting buckets every
window seconds.
"""
def __init__(self, child, groupby, aggregates, window):
QueryOperator.__init__(self)
self.child = child
self.can_query_stream_cache = self.can_query_stream_impl()
self.groupby = groupby
self.aggregates = aggregates
self.window = window
self.aggregator = Aggregator(self.aggregates, self.groupby,
self.window)
def filter(self, updates, return_passes, return_fails):
if return_passes:
(passes, fails) = self.child.filter(updates, return_passes,
return_fails)
new_emissions = []
new_emissions.extend(self.aggregator.update(passes))
return (new_emissions, None)
else:
return (None, None)
def filter_params(self):
return self.child.filter_params()
def can_query_stream(self):
return self.can_query_stream_cache
def can_query_stream_impl(self):
return self.child.can_query_stream()
def assign_descriptor(self, tuple_descriptor):
self.tuple_descriptor = tuple_descriptor
self.aggregator.tuple_descriptor = tuple_descriptor
with_aggregates = self.groupby.duplicate()
for aggregate in self.aggregates:
with_aggregates.add_descriptor(aggregate)
with_aggregates.add_descriptor(TwitterFields.created_field)
for alias, fd in tuple_descriptor.descriptors.items():
with_aggregates.add_descriptor(fd)
self.child.assign_descriptor(with_aggregates)
class GroupBy(QueryOperator):
"""
Groups results from child by some set of fields, and runs the aggregate
function(s) over them, emitting results and resetting buckets every
window seconds.
"""
def __init__(self, child, groupby, aggregates, window):
QueryOperator.__init__(self)
self.child = child
self.can_query_stream_cache = self.can_query_stream_impl()
self.groupby = groupby
self.aggregates = aggregates
self.window = window
self.aggregator = Aggregator(self.aggregates, self.groupby, self.window)
def filter(self, updates, return_passes, return_fails):
if return_passes:
(passes, fails) = self.child.filter(updates, return_passes, return_fails)
new_emissions = []
new_emissions.extend(self.aggregator.update(passes))
return (new_emissions, None)
else:
return (None, None)
def filter_params(self):
return self.child.filter_params()
def can_query_stream(self):
return self.can_query_stream_cache
def can_query_stream_impl(self):
return self.child.can_query_stream()
def assign_descriptor(self, tuple_descriptor):
self.tuple_descriptor = tuple_descriptor
self.aggregator.tuple_descriptor = tuple_descriptor
with_aggregates = self.groupby.duplicate()
for aggregate in self.aggregates:
with_aggregates.add_descriptor(aggregate)
with_aggregates.add_descriptor(TwitterFields.created_field)
for alias, fd in tuple_descriptor.descriptors.items():
with_aggregates.add_descriptor(fd)
self.child.assign_descriptor(with_aggregates)
def launch_test(test_params):
"""
This function instanciates the base classifiers and the aggregators. It
dispatches the computation on several cores of the micro-processor (if any).
It also takes care of experimental data saving, figure edition and so on.
"""
##########################################
# PARAMETERS BELOW SHOULD NOT BE CHANGED #
##########################################
rate_resolution = 0.01
if (test_params["dataset"] in ['moons', 'circles', 'blobs', 'neoblobs']):
n = test_params["n"]
if 'beta' in test_params.keys():
mybeta = test_params["beta"]
else:
mybeta = 0.5
Data = synthDataset(test_params["dataset"], n, beta=mybeta)
synth_data = True
else:
raise ValueError("Unknown dataset name.")
n_class = Data.n_class
Ns = Data.Ns
n_adver = 0
if "adversary" in test_params.keys():
for i in range(len(test_params["adversary"])):
n_adver += test_params["adversary"][i][1]
n_bad = 0
if "noisy" in test_params.keys():
for i in range(len(test_params["noisy"])):
n_bad += test_params["noisy"][i][1]
if 'names' in test_params.keys():
names = test_params["names"]
else:
names = []
for i in range(Ns):
names.append("Tree")
if (test_params["tnorm"] in ['Aczel-Alsina']):
lambda_range = np.logspace(0, 1, 101)
lambda_default = 5.0
elif (test_params["tnorm"] in ['convex']):
lambda_range = np.linspace(0.0, 1.0, 101)
lambda_default = 0.5
else:
raise ValueError('Unknown tnorm name.')
if (test_params["tnorm_ada"] in ['convex']):
lambda_range_ada = np.linspace(0.0, 1.0, 101)
lambda_default_ada = 0.5
elif (test_params["tnorm_ada"] in ['Aczel-Alsina']):
lambda_range_ada = np.logspace(0, 1, 101)
lambda_default_ada = 1.0
else:
raise ValueError('Unknown tnorm name.')
r_range = np.logspace(-1, 2, 201)
alpha_range = np.linspace(0, 1, 101)
regul_range = np.logspace(-2, 2, 101)
rho_range = np.logspace(-2, 2, 101)
# Models
clf = []
predictors = []
for i in range(Ns):
if names[i] == 'Tree':
clf.append(DecisionTreeClassifier(max_depth=2))
if names[i] == 'Reg Log':
clf.append(LogisticRegression(penalty='l2', C=1.0, solver='lbfgs'))
if names[i] == 'NBC':
clf.append(GaussianNB())
if names[i] == 'QDA':
clf.append(QuadraticDiscriminantAnalysis())
if names[i] == 'SVM_lin':
clf.append(SVC(kernel="linear", C=0.025))
if names[i] == 'SVM_nlin':
clf.append(SVC(gamma=2, C=1))
if names[i] == 'GP':
clf.append(GaussianProcessClassifier(1.0 * RBF(1.0)))
if names[i] == 'RF':
clf.append(
RandomForestClassifier(max_depth=5,
n_estimators=10,
max_features=1))
if names[i] == 'MLP':
clf.append(MLPClassifier(alpha=1))
if names[i] == 'Ada':
clf.append(AdaBoostClassifier())
if names[i] == 'kNN':
clf.append(KNeighborsClassifier(n_neighbors=5))
predictors.append(
Predictor('Base Clf ' + str(i + 1), 'base_clf', clf[i],
rate_resolution, 'k'))
adv = []
for i in range(len(test_params["adversary"])):
for j in range(test_params["adversary"][i][1]):
adv.append(
Adversary(clf[test_params["adversary"][i][0]], n_class,
test_params["adversary"][i][2]))
predictors.append(
Predictor('Advers. Clf ' + str(i + 1), 'base_clf', adv[i],
rate_resolution, 'k'))
Ns += 1
bad = []
for i in range(len(test_params["noisy"])):
for j in range(test_params["noisy"][i][1]):
bad.append(
NoisyClf(clf[test_params["noisy"][i][0]], n_class,
test_params["noisy"][i][2]))
predictors.append(
Predictor('Noisy Clf ' + str(i + 1), 'base_clf', bad[i],
rate_resolution, 'k'))
Ns += 1
for i in range(test_params["clone"]):
predictors.append(predictors[0].clone())
predictors[-1].name = 'Clone ' + str(i + 1)
clf.append(predictors[-1].machine)
Ns += 1
for i in range(test_params["random"]):
clf.append(DummyClassifier(strategy='uniform'))
predictors.append(
Predictor('Random Clf ' + str(i + 1), 'base_clf', clf[-1],
rate_resolution, 'k'))
Ns += 1
for i in range(test_params["dummy"]):
clf.append(DummyClassifier(strategy='most_frequent'))
predictors.append(
Predictor('Constant Clf ' + str(i + 1), 'base_clf', clf[-1],
rate_resolution, 'k'))
Ns += 1
selec = Aggregator('selection', Ns, n_class)
predictors.append(
Predictor('Selected Clf.', 'agg', selec, rate_resolution, 'b'))
wvote = Aggregator('weighted_vote',
Ns,
n_class,
params={
"r_range": r_range,
"expo": False
})
predictors.append(
Predictor('Weighted Vote Ens.', 'agg', wvote, rate_resolution,
'orange'))
expow = Aggregator('weighted_vote',
Ns,
n_class,
params={"r_range": r_range})
predictors.append(
Predictor('Expo. Weighted Vote Ens.', 'agg', expow, rate_resolution,
'brown'))
naive = Aggregator('naive', Ns, n_class, params={"method": 'indep'})
predictors.append(
Predictor('Naive Bayes.', 'agg', naive, rate_resolution, '--m'))
spocc = Aggregator(
'spocc', Ns, n_class, {
"tnorm": test_params["tnorm"],
"hyper": lambda_default,
"hyper_range": lambda_range
})
predictors.append(
Predictor('SPOCC (' + test_params["tnorm"] + ')', 'agg', spocc,
rate_resolution, 'm'))
adaspocc = Aggregator(
'adaspocc', Ns, n_class, {
"tnorm": test_params["tnorm_ada"],
"hyper": lambda_default_ada,
"hyper_range": lambda_range_ada,
"alpha_range": alpha_range,
"rho": 1,
"rho_range": rho_range
})
predictors.append(
Predictor('adaSPOCC (' + test_params["tnorm_ada"] + ')', 'agg',
adaspocc, rate_resolution, 'm'))
stack = Aggregator('stacked_logreg', Ns, n_class,
{"regul_range": regul_range})
predictors.append(
Predictor('Stacked Log. Reg.', 'agg', stack, rate_resolution, 'r'))
clf_ctl = LogisticRegression(penalty='l2', C=1.0)
predictors.append(
Predictor('Centralized Clf.', 'global', clf_ctl, rate_resolution,
'--g'))
if (test_params["dataset"] not in ['20newsgroup', 'mnist', 'drive']):
bayes = Aggregator('bayes', Ns, n_class)
predictors.append(
Predictor('Bayes Agg.', 'agg', bayes, rate_resolution, ':g'))
if synth_data:
optim = Oracle(test_params["dataset"])
predictors.append(
Predictor('Optimal Clf.', 'oracle', optim, rate_resolution, 'g'))
main_loop_partial = partial(main_loop,
n=n,
Data=Data,
mode=test_params["mode"],
Ns=Ns,
predictors=predictors,
sent=test_params["sent"],
iter_max=test_params["iter_max"])
results = main_loop_partial(test_params["rand_state"])
accuracy, conf_matrix, failed_loc = results
for i in range(len(predictors)):
print(predictors[i].name + " has accuracy " + str(accuracy[i]))
