thresholds = np.arange(n_levels, step=8)
thresholds += 10
encoder = MultiThresholdEncoder(thresholds=thresholds)
Xencode = encoder.transform(X)
del X
X = opu_mapping.transform(Xencode)
del Xencode
print(f"NEWMA OPU: Sketch took {time.time() - startingtime:.2f}s.")
mult = 1.5
detector = algos.NEWMA(X[0],
forget_factor=big_Lambda,
feat_func=lambda x: x.astype('float32'),
forget_factor2=small_lambda,
adapt_forget_factor=thres_ff * mult,
thresholding_quantile=0.95,
dist_func=lambda z1, z2: np.linalg.norm(z1 - z2))
detector.apply_to_data(X)
n = 0
detection_stat = np.array([i[0] for i in detector.stat_stored
])[int(10 * n):] # padding
online_th = np.array([i[1] for i in detector.stat_stored])[int(10 * n):]
computation_duration = time.time() - startingtime
print(f'NEWMA RP on OPU took {computation_duration:.2f}s.')
# ----------
# NEWMA FF on CPU
# Newma config
big_Lambda, small_lambda = algos.select_optimal_parameters(B) # forget factors chosen with heuristic in the paper
print('Chose Lambda = {:.4f} and lambda = {:.4f}'.format(big_Lambda, small_lambda))
thres_ff = small_lambda
# number of random features is set automatically with this criterion
m = int((1 / 4) / (small_lambda + big_Lambda) ** 2)
print('Number of RFs: {}'.format(m))
W, sigmasq = feat.generate_frequencies(m, d, data=data_sigma_estimate, choice_sigma=choice_sigma)
print('Start algo ', algo, ' with fixed threshold')
def feat_func(x):
return feat.fourier_feat(x, W)
detector95 = algos.NEWMA(X[0], forget_factor=big_Lambda, forget_factor2=small_lambda, feat_func=feat_func,
adapt_forget_factor=thres_ff)
detector95.apply_to_data(X)
# compute performance metrics
detection_stat95 = np.array([i[0] for i in detector95.stat_stored])[int(10 * n):] # padding
online_th95 = np.array([i[1] for i in detector95.stat_stored])[int(10 * n):]
ground_truth = ground_truth[int(10 * n):]
if args.show:
import matplotlib.pyplot as plt
plt.figure()
plt.plot(detection_stat95[:4 * n])
plt.plot(online_th95[:4 * n])
plt.plot(1.2 * np.max(detection_stat95[:4 * n]) * ground_truth[:4 * n], '--', color='k')
plt.legend([algo, 'Threshold', 'True Changes'], framealpha=1, ncol=1, handletextpad=0.1)
plt.show()
X[0],
kernel_func=lambda x, y: feat.gauss_kernel(x, y, np.sqrt(sigmasq)),
window_size=window_size,
nbr_windows=N,
adapt_forget_factor=thresh_ff,
store_values=False)
t = time.time()
ocpobj.apply_to_data(X) # actual computations
time_method = time.time() - t
print('time:', time_method)
else:
print('RFF')
W, _ = feat.generate_frequencies(m, d, data=X[:100])
updt_func = lambda x: feat.fourier_feat(x, W)
t = time.time()
ocpobj = algos.NEWMA(X[0],
forget_factor=ff,
forget_factor2=ff2,
feat_func=updt_func,
adapt_forget_factor=thresh_ff,
store_values=False)
ocpobj.apply_to_data(X) # actual computations
time_method = time.time() - t
print('Execution time: ', time_method)
running_times.append(time_method)
# save results for later plot
np.savez('test_B_runningtime_algo_{}.npz'.format(algo),
windows=window_size_list,
running_times=np.array(running_times))