def __init__(self, start_joint):
self.start = start_joint
self.start.bones_out.append(self)
self.end = Joint(self)
self.end.bone_in = self
self.rotation = 0
self.length = 1
self.transform = matrix.Identity()
self.image = None
def shrink_domain(self, epsilon, delta=2.2820610e-12, bound=0):
# measure all one-way marginals and shrink domain according to noisy measurements
data = self.data
self.round1 = list(self.column_order)
self.delta = delta
sigma = moments_calibration(1.0, 1.0, epsilon, delta)
self.sigma = sigma
print('NOISE LEVEL:', sigma)
supports = {}
for i, col in enumerate(self.round1):
if self.domain[col] <= bound:
supports[col] = np.asarray(
[True for j in range(self.domain[col])])
del (self.round1[i])
# round1 measurements will be weighted to have L2 sensitivity 1
weights = np.ones(len(self.round1))
weights /= np.linalg.norm(weights) # now has L2 norm = 1
for col, wgt in zip(self.round1, weights):
##########################
### Noise-addition step ##
##########################
proj = (col, )
hist = np.asarray(data[col].value_counts())
print(hist)
noise = sigma * np.random.randn(hist.size)
y = wgt * hist + noise
#####################
## Post-processing ##
#####################
sup = y >= 3 * sigma
#If the column has fewer possible values than the bound we just leave it be.
supports[col] = sup
if sup.sum() == y.size:
y2 = y
I2 = matrix.Identity(y.size)
else:
y2 = np.append(y[sup], y[~sup].sum())
I2 = np.ones(y2.size)
I2[-1] = 1.0 / np.sqrt(y.size - y2.size + 1.0)
y2[-1] /= np.sqrt(y.size - y2.size + 1.0)
I2 = sparse.diags(I2)
self.measurements.append((I2, y2 / wgt, 1.0 / wgt, proj))
self.supports = supports
data, new_domain = transform_data(data, self.domain, supports)
self.domain = new_domain
self.data = data
return data, new_domain
def measure(self, round2, from_r=False):
# measure selected queries, which are round2 queries
print("selected queries:", round2)
if from_r:
round2 = r_to_python(round2, list(self.column_order))
self.round2 = round2 #round2 is a query list[]
# round2 measurements will be weighted to have L2 sensitivity 1
# perform round 2 measurments over compressed domain
weights = np.ones(len(self.round2))
weights /= np.linalg.norm(weights) # now has L2 norm = 1
for proj, wgt in zip(self.round2, weights):
#########################
## Noise-addition step ##
#########################
indices = itemgetter(*proj)(self.domain)
hist = datavector(self.data[list(proj)], indices)
Q = matrix.Identity(hist.size)
noise = self.sigma * np.random.randn(Q.shape[0])
y = wgt * Q.dot(hist) + noise
self.measurements.append((Q, y / wgt, 1.0 / wgt, proj))
def test_Identity(self):
N = 10
M = matrix.Identity(N)
self.assertEqual(M[0][0], 1)
self.assertEqual(M[0][1], 0)
def __init__(self, bone_in):
self.bone_in = bone_in
self.bones_out = []
self.transform = matrix.Identity()
def measure(self):
data = self.load_data()
# round1 and round2 measurements will be weighted to have L2 sensitivity 1
sigma = moments_calibration(1.0, 1.0, self.epsilon, self.delta)
print('NOISE LEVEL:', sigma)
weights = np.ones(len(self.round1))
weights[self.round1.index('INCWAGE_A')] *= 2.0
weights /= np.linalg.norm(weights) # now has L2 norm = 1
supports = {}
self.measurements = []
for col, wgt in zip(self.round1, weights):
##########################
### Noise-addition step ##
##########################
proj = (col, )
hist = data.project(proj).datavector()
noise = sigma * np.random.randn(hist.size)
y = wgt * hist + noise
#####################
## Post-processing ##
#####################
if col in [
'INCWAGE_A', 'SEA', 'METAREA', 'COUNTY', 'CITY', 'METAREAD'
]:
sup = np.ones(y.size, dtype=bool)
else:
sup = y >= 3 * sigma
supports[col] = sup
print(col, self.domain.size(col), sup.sum())
if sup.sum() == y.size:
y2 = y
I2 = matrix.Identity(y.size)
else:
y2 = np.append(y[sup], y[~sup].sum())
I2 = np.ones(y2.size)
I2[-1] = 1.0 / np.sqrt(y.size - y2.size + 1.0)
y2[-1] /= np.sqrt(y.size - y2.size + 1.0)
I2 = sparse.diags(I2)
self.measurements.append((I2, y2 / wgt, 1.0 / wgt, proj))
self.supports = supports
# perform round 2 measurments over compressed domain
data = transform_data(data, supports)
self.domain = data.domain
self.round2 = [cl for cl in self.round2 if self.domain.size(cl) < 1e6]
weights = np.ones(len(self.round2))
weights[self.round2.index(
('SEX', 'CITY', 'INCWAGE_A'))] *= self.weight3
weights[self.round2.index(('SEX', 'CITY'))] *= 2.0
weights[self.round2.index(('SEX', 'INCWAGE_A'))] *= 2.0
weights[self.round2.index(('CITY', 'INCWAGE_A'))] *= 2.0
weights /= np.linalg.norm(weights) # now has L2 norm = 1
for proj, wgt in zip(self.round2, weights):
#########################
## Noise-addition step ##
#########################
hist = data.project(proj).datavector()
if proj == ('SEX', 'CITY', 'INCWAGE_A'):
dom = self.domain.project(proj).shape
I = sparse.eye(dom[0] * dom[1])
Q = sparse.kron(I, self.Q_INCWAGE).tocsr()
elif proj == ('CITY', 'INCWAGE_A'):
I = sparse.eye(self.domain.size('CITY'))
Q = sparse.kron(I, self.Q_INCWAGE).tocsr()
else:
Q = matrix.Identity(hist.size)
noise = sigma * np.random.randn(Q.shape[0])
y = wgt * Q.dot(hist) + noise
self.measurements.append((Q, y / wgt, 1.0 / wgt, proj))