def main(args):
sensor, impact = args
print(sensor, impact)
rbase, ibase = get_signals(sensor, 0)
nbaseline, nbin = rbase.shape
rsig, isig = get_signals(sensor, impact)
t = np.linspace(0, 1, window_size + 1)
# Define the classification pipeline
pipe = Pipeline([('aug', utils.FeatureAugment()),
('norm', StandardScaler()),
('clf', utils.MultiOCSVM(eta=0.2, kernel='rbf', nu=0.01, gamma=0.2))])
det = []
# perform monte carlo
for i in range(n_mc):
temperatures = utils.gp_simple(t, sigma=150, mean=10, kernel='laplacian')
temperatures[temperatures < 0] = 0
temperatures[temperatures > nbaseline - 1] = nbaseline - 1
res = 0
for j in range(nbin):
# training
rpart = rbase.iloc[:, j](temperatures[:-1], ext=3).values
ipart = ibase.iloc[:, j](temperatures[:-1], ext=3).values
pipe.fit(rpart + 1j * ipart)
# prediction
rpart = rsig.iloc[:, j](temperatures[-1], ext=3)
ipart = isig.iloc[:, j](temperatures[-1], ext=3)
res += pipe.predict(np.array([rpart + 1j * ipart]))[0]
det.append((res >= zeta * nbin).astype(int))
return (sensor, impact), 1 - np.mean(det)
""" A schematic of the classification approach, showing the classifier along with the defect
signal."""
from packages import utkit, scihdf, utils
from matplotlib import pyplot
from os.path import join
import numpy as np
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
np.random.seed(10)
baselines = utils.make_range_signals(15, [0.9, 1.1], [-np.pi / 5, np.pi / 5])
damage = utils.make_random_signals(2, 0.02, np.pi / 10,
0.9)[0] + 0.2 * np.exp(1j * np.pi / 5)
pipe = Pipeline([('aug', utils.FeatureAugment()), ('norm', StandardScaler()),
('clf',
utils.MultiOCSVM(eta=0.8, kernel='rbf', nu=0.01,
gamma=0.4))])
pipe.fit(baselines)
pyplot.style.use(join('..', 'plot_styles.mplstyle'))
utils.plot_svm_frontiers(pipe, damage, figsize=(2, 2))
ax = pyplot.gca()
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_xlabel('Real', labelpad=-4)
ax.set_ylabel('Imaginary', labelpad=-4)
pyplot.savefig('demo_classification.svg')
# pyplot.show()
rho = 0.
# Total number of defects to test
n = 500
# Amplitude of the defect signals
defect_amp = 2
nbaselines = 20
eta, gamma = 0.8, 0.2
angs = np.linspace(-np.pi, np.pi, 20)
# Make the SVM
pipe = Pipeline([('aug', utils.FeatureAugment()),
('norm', StandardScaler()),
('clf', utils.MultiOCSVM(eta=eta, kernel='rbf', nu=0.01, gamma=gamma))])
fig, axarr = pyplot.subplots(4, 5)
for ax, (amp, phase) in zip(*(axarr.ravel(), itertools.product(amp_sigma, phi_sigma))):
# for amp, phase in itertools.product(amp_sigma, phi_sigma):
print(amp, phase/np.pi)
# Generate the training data
p = 2
baseline = utils.make_range_signals(nbaselines,
amp_range=[amp_mean - p * amp, amp_mean + p * amp],
phase_range=[phi_mean - p * phase, phi_mean + p * phase],
corr=1)
# Train the SVM
pipe.fit(baseline)
det, fa = [], []
for ang in angs:
feat = feat.applymap(complex)
idx, out = [], []
for sid in conf['guided_waves']['experiment']['sensor_id']:
for k in range(1, 7):
baseline = feat.loc[pd.IndexSlice[sid, 0, :], :].values[::k, :]
nbaseline, nbin = baseline.shape
print(nbaseline)
pipes = []
for i in range(baseline.shape[1]):
pipes.append(
Pipeline([('aug', utils.FeatureAugment()),
('norm', StandardScaler()),
('clf',
utils.MultiOCSVM(eta=0.55,
kernel='rbf',
nu=0.01,
gamma=0.05))]))
pipes[-1].fit(baseline[:, i])
for i in range(10):
res = 0
for j in range(nbin):
defect = feat.loc[pd.IndexSlice[sid, i, :], :].values[:, j]
res += pipes[j].predict(defect)
res = (res >= 0.8 * nbin).astype(int)
out.append(1 - np.mean(res))
idx.append([sid, nbaseline, i])
# make the dataframe and sort its index
mux = pd.MultiIndex.from_arrays(list(zip(*idx)),