def example1():
""" Optimize AllRange workload using PIdentity template and report the expected error """
print('Example 1')
W = workload.AllRange(256)
pid = templates.PIdentity(16, 256)
res = pid.optimize(W)
err = error.rootmse(W, pid.strategy())
err2 = error.rootmse(W, workload.Identity(256))
print(err, err2)
def calculate_workload_error2(wk, strategy, eps):
# Deprecated
expected_error = hdmm_error.rootmse(wk, strategy, eps=eps)
per_query_error = hdmm_error.per_query_error_sampling(wk,
strategy,
100000,
eps,
normalize=True)
if np.var(per_query_error) < 1e-2:
pdf = False
pdf_x = False
else:
density = np.histogram(per_query_error, bins='auto', density=True)
pdf = density[0].tolist()
pdf_x = density[1][0:-1].tolist()
res = {
'pdf': pdf,
'pdf_x': pdf_x,
'method': 'HDMM',
'expected_error': expected_error,
'num_query': int(wk.shape[0]),
}
return res
ax = sns.scatterplot(x=mean, y=std)
ax.set(xlabel='Average Error',
ylabel='Stdev Error',
title=f'Avg vs Stdev of Error over re-optimizations, SF1')
plt.savefig('error_variation_std.png')
plt.show()
if __name__ == '__main__':
full = [
'P1', 'P3', 'P4', 'P5', 'P8', 'P9', 'P10', 'P11', 'P12', 'P12A_I',
'PCT12', 'PCT12A_O'
]
w_sf1_full = build_workload(full)
print('Query count, full', w_sf1_full.shape[0])
stats = {}
trials = 10 #150
strategies = [opt_strategy(w_sf1_full) for i in range(trials)]
errors = [
np.sqrt(error.per_query_error(w_sf1_full, a)) for a in strategies
]
stats['rootmse '] = [error.rootmse(w_sf1_full, a) for a in strategies]
stats['max_query_err'] = [np.max(e) for e in errors]
stats['min_query_err'] = [np.min(e) for e in errors]
stats['mean_query_err'] = [np.mean(e) for e in errors]
# mean_vs_max_error(w_sf1_full)
mean_vs_max_error(w_sf1_full, trials=100)
temp = workload.Marginals.approximate(A_manual)
A_manual_opt = templates.Marginals(domain, seed=1003)
A_manual_opt._params = temp.weights
A_manual_opt.optimize(W)
A_manual_opt._params = np.clip(A_manual_opt._params, 0, float('inf'))
print('Num queries:', W.shape[0])
print('Sensitivity:', W.sensitivity())
print('Marg', '\t\t', f'{error.rootmse(W, A_marg.strategy()):10.3f}')
print('Ident_full', '\t', f'{error.rootmse(W, A_identity_full.strategy()):10.3f}')
print('Identity', '\t', f'{error.rootmse(W, A_identity):10.3f}')
print('Workload', '\t', f'{error.rootmse(W, A_wkld):10.3f}')
print('Manual', '\t', f'{error.rootmse(W, A_manual):10.3f}')
print('Manual Opt', '\t', f'{error.rootmse(W, A_manual_opt.strategy()):10.3f}')
summarize_strategy(W, A_marg, domain)
print('')
print('')
for m in W.matrices:
print(m.base.key, m.base.shape[0], '\t', error.rootmse(m, A_marg.strategy()))
for m in W.matrices:
print(m.base.key, m.base.shape[0], '\t', error.rootmse(m, A_manual))
for m in W.matrices:
print(m.base.key, m.base.shape[0], '\t', error.rootmse(m, A_manual_opt.strategy()))
A_scen2 = templates.Marginals(domain)
A_scen2.optimize(W_scen2)
# Here we analyze the error of scenario 2, which is what the user really wants
# But we consider the strategy from scenario1 and compare with strategy derived from scenario2
print('Num queries:', W_scen2.shape[0])
print('Sensitivity:', W_scen2.sensitivity())
print('Per query RMSE, A_scen1', '\t\t',
f'{error.rootmse(W_scen1, A_scen1.strategy()):10.3f}')
print('Per query RMSE, A_scen2', '\t\t',
f'{error.rootmse(W_scen1, A_scen2.strategy()):10.3f}')
print('')
summarize_strategy(W_scen1, A_scen1, domain)
print('')
summarize_strategy(W_scen1, A_scen2, domain)
print('')
for m in W_scen1.matrices:
print(
m.base.key,
util.marginal_index_repr(m.base.key, dim, " "),
m.base.shape[0],
'\t',
error.rootmse(m, A_scen1.strategy()),
error.rootmse(m, A_scen2.strategy()),
)
return A
def marginal_strategy(workload=None):
template = templates.Marginals((2, 2, 63, 8))
A = template.restart_optimize(workload, 25)[0]
return A
if __name__ == '__main__':
W = pl94_workload(with_full_id=True)
print(W.shape)
A = opt_p_identity(workload=W)
err = error.rootmse(W, A)
print('KroneckerPIdentity', err)
#W = Marginals.approximate(W)
A = marginal_strategy(workload=W)
marg_err = error.rootmse(W, A)
#marg_err = np.sqrt( error.expected_error(W, A) / (3*3*64*9) )
print('Marginals', marg_err)
robert_err = error.rootmse(W, manual_strategy())
print('Robert', robert_err)
print(A.weights / A.weights.sum())
rmse = np.sqrt(np.mean(errors) / W.shape[0])
print(rmse)
return rmse
if __name__ == '__main__':
#x = synthetic_data(N=1000000)
engine = 'wnnls'
W = pl94_workload()
A1 = opt_p_identity(W)
A2 = marginal_strategy(W)
A3 = manual_strategy()
err1 = error.rootmse(W, A1)
err2 = error.rootmse(W, A2)
err3 = error.rootmse(W, A3)
trials = 25
Ns = [10**k for k in range(1, 8)]
errs1, errs2, errs3 = [], [], []
for N in Ns:
x = synthetic_data(N)
errs1.append(run_test(W, A1, x, trials=trials, engine=engine))
errs2.append(run_test(W, A2, x, trials=trials, engine=engine))
errs3.append(run_test(W, A3, x, trials=trials, engine=engine))
plt.plot(Ns, [err1] * len(Ns), 'b', label='KronPIdentity+LS')
plt.plot(Ns, [err2] * len(Ns), 'r', label='Marginals+LS')
