def _strategy(self):
return implicit.krons(*[S._strategy() for S in self.strategies])
def _inverse(self):
return implicit.krons(*[S._inverse() for S in self.strategies])
def squared_error(self, noise):
WtW = implicit.krons(*[w.WtW for w in self.workloads])
return noise.dot(WtW.dot(noise))
def evaluate(self, x):
W = implicit.krons(*[w.W for w in self.workloads])
return W.dot(x)
for eps in [0.1, 1.0]:
if geography:
dims = [[0], [1], [2], [4], [5]]
axes = 3
p = [1, 1, 8, 10, 1]
else:
dims = [[0], [1], [2], [4]]
axes = (3, 5)
p = [1, 1, 8, 10]
x = X.sum(axis=axes).flatten()
N = x.sum()
sf1 = CensusSF1(geography=True).project_and_merge(dims)
W = implicit.stack(*[
implicit.krons(*[S.W for S in K.workloads]) for K in sf1.workloads
])
A_sf1 = optimize.restart_union_kron(sf1, 50, p)
A1 = implicit.krons(*[np.linalg.pinv(A) for A in A_sf1])
noise = A1.dot(
np.random.laplace(loc=0, scale=1.0 / eps, size=A1.shape[1]))
err = W.dot(noise)
ans = W.dot(x)
result[(geography, eps)] = (ans, err)
rel = np.abs(err) / np.maximum(W.dot(x),
100) #np.maximum(W.dot(x), 0.001*N)
print geography, '%.1f, %.4f, %.4f' % (eps, np.median(rel),
np.percentile(rel, 95))
# tmp = np.arange(ans.max())
# plt.plot(ans, ans+err, '.')
try:
time2d, time3d, time4d, timesum, timedc = pickle.load(
open('scalability_running.pkl', 'rb'))
print 'checkpt try'
except:
time2d = []
time3d = []
time4d = []
timesum = []
timedc = []
for n in dom2d:
A = sparse.csr_matrix(strategies[n])
T = np.ones((1, n))
B1 = implicit.krons(A, T)
B2 = implicit.krons(T, A)
B = implicit.stack(B1, B2)
# B1 = sparse.kron(A, T, format='csr')
# B2 = sparse.kron(T, A, format='csr')
# B = sparse.vstack([B1, B2], format='csr')
timesum.append(run_mechanism(B, implicit.sparse_inverse(B)))
print n, timesum[-1]
# embed()
# 2d domains
for n in dom2d:
print n
A = strategies[n]
X = np.load('%s/census.npy' % base)
for geography in [True]: #[False, True]:
if geography:
x = X.flatten()
else:
x = X.sum(axis=5).flatten()
for compact in [False]: # [False, True]:
if compact:
sf1 = CensusSF1(geography)
else:
sf1 = CensusSF1Big(geography, reallybig=False)
t0 = time.time()
W = implicit.stack(*[
implicit.krons(*[S.W for S in K.workloads]) for K in sf1.workloads
])
marg = approximation.marginals_approx(sf1)
q = sf1.queries
split = len(sf1.workloads) / 2
sf1a = workload.Concat(sf1.workloads[:split])
sf1b = workload.Concat(sf1.workloads[split:])
strategy = optimize.restart_union_kron(sf1, 25, [1, 1, 8, 1, 10, 1])
theta, phi = optimize.restart_marginals(marg, 25)
A1 = optimize.restart_union_kron(sf1a, 25, [1, 1, 8, 1, 10, 1])
A2 = optimize.restart_union_kron(sf1b, 25, [1, 1, 8, 1, 10, 1])
for A in strategy:
A[A < 1e-3] = 0.0