def calculate_metric(predictions, labels, signal='pulse', window_size=360, fs=30, bpFlag=True):
if signal == 'pulse':
[b, a] = butter(1, [0.75 / fs * 2, 2.5 / fs * 2], btype='bandpass') # 2.5 -> 1.7
else:
[b, a] = butter(1, [0.08 / fs * 2, 0.5 / fs * 2], btype='bandpass')
data_len = len(predictions)
HR_pred = []
HR0_pred = []
mySNR = []
for j in range(0, data_len, window_size):
if j == 0 and (j+window_size) > data_len:
pred_window = predictions
label_window = labels
elif (j + window_size) > data_len:
break
else:
pred_window = predictions[j:j + window_size]
label_window = labels[j:j + window_size]
if signal == 'pulse':
pred_window = detrend(np.cumsum(pred_window), 100)
else:
pred_window = np.cumsum(pred_window)
label_window = np.squeeze(label_window)
if bpFlag:
pred_window = scipy.signal.filtfilt(b, a, np.double(pred_window))
pred_window = np.expand_dims(pred_window, 0)
label_window = np.expand_dims(label_window, 0)
# Predictions FFT
f_prd, pxx_pred = scipy.signal.periodogram(pred_window, fs=fs, nfft=4 * window_size, detrend=False)
if signal == 'pulse':
fmask_pred = np.argwhere((f_prd >= 0.75) & (f_prd <= 2.5)) # regular Heart beat are 0.75*60 and 2.5*60
else:
fmask_pred = np.argwhere((f_prd >= 0.08) & (f_prd <= 0.5)) # regular Heart beat are 0.75*60 and 2.5*60
pred_window = np.take(f_prd, fmask_pred)
# Labels FFT
f_label, pxx_label = scipy.signal.periodogram(label_window, fs=fs, nfft=4 * window_size, detrend=False)
if signal == 'pulse':
fmask_label = np.argwhere((f_label >= 0.75) & (f_label <= 2.5)) # regular Heart beat are 0.75*60 and 2.5*60
else:
fmask_label = np.argwhere((f_label >= 0.08) & (f_label <= 0.5)) # regular Heart beat are 0.75*60 and 2.5*60
label_window = np.take(f_label, fmask_label)
# MAE
temp_HR, temp_HR_0 = calculate_HR(pxx_pred, pred_window, fmask_pred, pxx_label, label_window, fmask_label)
temp_SNR = calculate_SNR(pxx_pred, f_prd, temp_HR_0, signal)
HR_pred.append(temp_HR)
HR0_pred.append(temp_HR_0)
mySNR.append(temp_SNR)
HR = np.array(HR_pred)
HR0 = np.array(HR0_pred)
mySNR = np.array(mySNR)
MAE = np.mean(np.abs(HR - HR0))
RMSE = np.sqrt(np.mean(np.square(HR - HR0)))
meanSNR = np.nanmean(mySNR)
return MAE, RMSE, meanSNR, HR0, HR
def main():
args = parse()
file = args.file
speech_path = os.path.join(args.speechfolder, file)
true_egg_path = os.path.join(args.eggfolder, file)
estimated_egg_path = os.path.join(args.geneggfolder, file)
speech = np.load(speech_path)
speech = minmaxnormalize(speech)
speech = smooth(speech, 21)
true_egg = np.load(true_egg_path)
true_egg = minmaxnormalize(true_egg)
if args.detrend:
_, true_egg = detrend(None, true_egg)
estimated_egg = np.load(estimated_egg_path)
estimated_egg = minmaxnormalize(estimated_egg)
# srange = slice(0, speech.shape[0])
srange = slice(10700, 15400)
speech = speech[srange]
plt.figure()
fig = plt.gcf()
plt.subplot(211)
plt.title("Speech")
plt.plot(speech, "k", label="Speech Waveform")
plt.xlabel("Sample")
plt.ylabel("Amplitude")
dft = np.fft.rfft(speech)
freqs = np.fft.rfftfreq(np.size(speech, 0), 1 / 16e3)
assert freqs.shape == dft.shape
plt.subplot(212)
plt.title("Fourier Spectra")
plt.gca().semilogx(freqs, np.abs(dft)**2, "b", label="DFT")
# plt.plot(freqs, np.abs(dft), "b", label="DFT")
plt.xlabel("Frequency")
plt.ylabel("PSD")
plt.subplots_adjust(top=0.926,
bottom=0.117,
left=0.078,
right=0.981,
hspace=0.476,
wspace=0.2)
mng = plt.get_current_fig_manager()
mng.window.showMaximized()
plt.show()
fig.savefig("images/fft.png")
def detrend_data(averaged_data):
"""
Detrend the time series, taking away long-term effects.
"""
detrended_data = averaged_data.copy()
for moteid in range(1, 59):
for feature in ['temperature_mean', 'humidity_mean']:
if moteid not in [5, 15, 28, 57]:
detrended_data.loc[detrended_data['moteid']==moteid,feature]=\
detrend(averaged_data, moteid,feature)
return detrended_data
def extract_gci_metrics(true_egg, estimated_egg, detrend_egg=False, fs=16e3):
fnames = {}
if type(true_egg) is str:
fnames = {"true_file": true_egg}
true_egg = np.load(true_egg)
if detrend_egg:
_, true_egg = detrend(None, true_egg)
if type(estimated_egg) is str:
estimated_egg = np.load(estimated_egg)
if len(true_egg) > len(estimated_egg):
true_egg = true_egg[:len(estimated_egg)]
true_egg, estimated_egg = detect_voiced_region(true_egg, estimated_egg)
true_gci = detectgroundwaveletgci(true_egg) / fs
estimated_gci = detectgenwaveletgci(estimated_egg) / fs
metrics = corrected_naylor_metrics(true_gci, estimated_gci)
metrics.update(fnames)
return metrics
def main():
args = parse()
fname = args.file
speech_path = os.path.join(args.speechfolder, fname)
true_egg_path = os.path.join(args.eggfolder, fname)
estimated_egg_path = os.path.join(args.geneggfolder, fname)
speech = np.load(speech_path)
speech = minmaxnormalize(speech)
true_egg = np.load(true_egg_path)
true_egg = minmaxnormalize(true_egg)
if args.detrend:
_, true_egg = detrend(None, true_egg)
estimated_egg = np.load(estimated_egg_path)
estimated_egg = minmaxnormalize(estimated_egg)
# region = slice(11000, 12000)
region = slice(0, len(speech))
plt.figure()
fig = plt.gcf()
ax = plt.subplot(211)
plt.plot(speech[region], "k", label="Speech Waveform")
plt.ylabel("Amplitude")
plt.subplot(212, sharex=ax)
plt.plot(true_egg[region], "k", label="Ground Truth EGG")
plt.plot(estimated_egg[region], "b", label="Estimated EGG")
plt.xlabel("Sample Number")
plt.ylabel("Amplitude")
plt.subplots_adjust(
top=0.962, bottom=0.087, left=0.057, right=0.981, hspace=0.212, wspace=0.2
)
mng = plt.get_current_fig_manager()
mng.window.showMaximized()
plt.show()
def main():
args = parse()
fname = args.filename
eground = np.load(os.path.join(args.groundpath, fname))
if args.detrend:
_, eground = detrend(None, eground)
power = get_envelope(eground)
fig = plt.figure()
plt.title("Voicing Visualization")
plt.plot(eground / np.max(np.abs(eground)), "r", label="ground truth")
plt.plot(power, "g", label="envelope")
plt.xlabel("sample")
plt.ylabel("amplitude")
plt.legend()
fig.tight_layout()
mng = plt.get_current_fig_manager()
mng.window.showMaximized()
plt.show()
def extract_quotient_metrics(true_egg,
estimated_egg,
detrend_egg=False,
fs=16e3):
def _detect_voiced_region_as_regions(true_egg,
reconstructed_egg,
power_threshold=0.05):
def _get_signal_power(x, window):
power = np.convolve(x**2, window / window.sum(), mode="same")
return power
def _get_window(window_len=10, window="flat"):
if window == "flat": # average
w = np.ones(window_len, "d")
else:
w = eval("np." + window + "(window_len)")
return w
true_scaler = pd.Series(np.abs(true_egg)).nlargest(100).median()
reconstructed_scaler = (pd.Series(
np.abs(reconstructed_egg)).nlargest(100).median())
true_egg = true_egg / true_scaler
reconstructed_egg = reconstructed_egg / reconstructed_scaler
window = _get_window(window_len=501, window="hanning")
power = _get_signal_power(true_egg, window)
regions = power >= power_threshold
# true_egg_voiced = true_egg[regions]
# reconstructed_egg_voiced = reconstructed_egg[regions]
return regions
fnames = {}
if type(true_egg) is str:
fnames = {"true_file": true_egg}
true_egg = np.load(true_egg)
if detrend_egg:
_, true_egg = detrend(None, true_egg)
if type(estimated_egg) is str:
estimated_egg = np.load(estimated_egg)
# true_egg = lowpass(true_egg)
# estimated_egg = lowpass(estimated_egg)
if len(true_egg) > len(estimated_egg):
true_egg = true_egg[:len(estimated_egg)]
regions = _detect_voiced_region_as_regions(true_egg, estimated_egg)
true_gci = detectgroundwaveletgci(true_egg)
estimated_gci = detectgenwaveletgci(estimated_egg)
true_egg, estimated_egg = groundeggfilter(true_egg), geneggfilter(
estimated_egg)
true_degg = np.gradient(true_egg, edge_order=2)
estimated_degg = np.gradient(estimated_egg, edge_order=2)
true_gci = positions2onehot(true_gci, regions.shape) * regions
estimated_gci = positions2onehot(estimated_gci, regions.shape) * regions
true_gci = np.nonzero(true_gci)[0]
estimated_gci = np.nonzero(estimated_gci)[0]
true_goi = []
estimated_goi = []
for i in range(true_gci.shape[0] - 1):
true_goi.append(true_gci[i] +
np.argmax(true_degg[true_gci[i] + 1:true_gci[i + 1]]) +
1)
for i in range(estimated_gci.shape[0] - 1):
estimated_goi.append(estimated_gci[i] +
np.argmax(estimated_degg[estimated_gci[i] +
1:estimated_gci[i +
1]]) + 1)
true_goi = np.array(true_goi)
estimated_goi = np.array(estimated_goi)
labelregions = label(regions)[0]
true_goi = positions2onehot(true_goi, regions.shape)
estimated_goi = positions2onehot(estimated_goi, regions.shape)
true_gci = positions2onehot(true_gci, regions.shape)
estimated_gci = positions2onehot(estimated_gci, regions.shape)
true_peaks = -true_gci + true_goi
estimated_peaks = -estimated_gci + estimated_goi
tpeaks = []
true_degg_list = []
for r in find_objects(labelregions):
tregion = true_peaks[r]
true_degg_region = true_degg[r]
tpos = np.nonzero(tregion)[0]
if len(tpos) < 2:
continue
if tregion[tpos[0]] > 0:
tpos = tpos[1:]
if tregion[tpos[-1]] > 0:
tpos = tpos[:-1]
assert len(tpos) % 2 == 1
tregion = tregion[tpos[0]:tpos[-1] + 1]
tpeaks.append(tregion)
true_degg_list.append(true_degg_region[tpos[0]:tpos[-1] + 1])
epeaks = []
estimated_degg_list = []
for r in find_objects(labelregions):
eregion = estimated_peaks[r]
estimated_degg_region = estimated_degg[r]
epos = np.nonzero(eregion)[0]
if len(epos) < 2:
continue
if eregion[epos[0]] > 0:
epos = epos[1:]
if eregion[epos[-1]] > 0:
epos = epos[:-1]
assert len(epos) % 2 == 1
eregion = eregion[epos[0]:epos[-1] + 1]
epeaks.append(eregion)
estimated_degg_list.append(estimated_degg_region[epos[0]:epos[-1] + 1])
true_peaks_list = [np.nonzero(t)[0] for t in tpeaks]
estimated_peaks_list = [np.nonzero(e)[0] for e in epeaks]
metrics = {
"CQ_true": 0,
"OQ_true": 0,
"SQ_true": 0,
"CQ_estimated": 0,
"OQ_estimated": 0,
"SQ_estimated": 0,
}
count = 0
sq_count = 0
for tr, tr_degg in zip(true_peaks_list, true_degg_list):
for i in range(1, tr.shape[0], 2):
count += 1
metrics["CQ_true"] += (tr[i + 1] - tr[i]) / (tr[i + 1] - tr[i - 1])
metrics["OQ_true"] += (tr[i] - tr[i - 1]) / (tr[i + 1] - tr[i - 1])
temp = extract_speed_quotient(tr_degg, tr)
metrics["SQ_true"] += temp[0]
sq_count += temp[1]
metrics["SQ_true"] /= sq_count
metrics["CQ_true"] /= count
metrics["OQ_true"] /= count
count = 0
sq_count = 0
for er, er_degg in zip(estimated_peaks_list, estimated_degg_list):
for i in range(1, er.shape[0], 2):
count += 1
metrics["CQ_estimated"] += (er[i + 1] - er[i]) / (er[i + 1] -
er[i - 1])
metrics["OQ_estimated"] += (er[i] - er[i - 1]) / (er[i + 1] -
er[i - 1])
temp = extract_speed_quotient(er_degg, er)
metrics["SQ_estimated"] += temp[0]
sq_count += temp[1]
metrics["SQ_estimated"] /= sq_count
metrics["CQ_estimated"] /= count
metrics["OQ_estimated"] /= count
metrics.update(fnames)
return metrics
env = PctChange(env)
env = EWMA(env, alpha=alpha)
return env
env_name = "buy_sell_hold_pct_ewma"
register_env(env_name, lambda config: buy_sell_hold_pct_ewma(config))
if __name__ == "__main__":
ray.init()
open_prices_detrended = (quandl.get(
'WIKI/MSFT',
start_date="2014-01-01",
end_date="2017-01-01",
api_key=my_secrets.quandl_api_key).assign(
Open=lambda df: detrend(df['Open'])).assign(
Open=lambda df: df['Open'] - df['Open'].min() + 1)['Open'])
tune.run_experiments({
"PPO_Detrended": {
"run": "PPO",
"stop": {
"time_total_s": 60 * 10,
},
"checkpoint_at_end": True,
"checkpoint_freq": 20,
"config": {
"env": env_name,
"num_workers": 2, # parallelism
"lr": grid_search([5e-4, 5e-5]), # try different lrs
"train_batch_size": grid_search([4_000]),
def pattern_detector(
matrices,
kernel,
pattern_type="loops",
precision=4.0,
area=8,
undetermined_percentage=1.,
labels=None,
):
if isinstance(matrices, np.ndarray):
matrix_list = [matrices]
else:
matrix_list = matrices
if labels is None:
labels = range(len(matrix_list))
elif isinstance(labels, str):
labels = [labels]
pattern_windows = [] # list containing all pannel of detected patterns
pattern_sums = np.zeros(
(area * 2 + 1, area * 2 + 1)
) # sum of all detected patterns
agglomerated_pattern = np.zeros(
(area * 2 + 1, area * 2 + 1)
) # median of all detected patterns
detected_patterns = []
n_patterns = 0
for matrix, name in zip(matrix_list, labels):
detrended, threshold_vector = utils.detrend(matrix)
matrix_indices = np.where(matrix.sum(axis=0) > threshold_vector)
n = matrix.shape[0]
res2 = utils.corrcoef2d(
detrended, kernel, centered_p=False
) # !! Here the pattern match !!
res2[np.isnan(res2)] = 0.0
n2 = res2.shape[0]
res_rescaled = np.zeros(np.shape(matrix))
res_rescaled[
np.ix_(
range(int(area), n2 + int(area)),
range(int(area), n2 + int(area)),
)
] = res2
VECT_VALUES = np.reshape(res_rescaled, (1, n ** 2))
VECT_VALUES = VECT_VALUES[0]
thr = np.median(VECT_VALUES) + precision * np.std(VECT_VALUES)
indices_max = np.where(res_rescaled > thr)
indices_max = np.array(indices_max)
res_rescaled = np.triu(res_rescaled)
res_rescaled[(res_rescaled) < 0] = 0
pattern_peak = utils.picker(res_rescaled, thr)
if pattern_peak != "NA":
if pattern_type == "loops":
# Assume all loops are not found too far-off in the matrix
mask = (
np.array(abs(pattern_peak[:, 0] - pattern_peak[:, 1]))
< 5000
)
pattern_peak = pattern_peak[mask, :]
mask = (
np.array(abs(pattern_peak[:, 0] - pattern_peak[:, 1])) > 2
)
pattern_peak = pattern_peak[mask, :]
elif pattern_type == "borders":
# Borders are always on the diagonal
mask = (
np.array(abs(pattern_peak[:, 0] - pattern_peak[:, 1])) == 0
)
pattern_peak = pattern_peak[mask, :]
for l in pattern_peak:
if l[0] in matrix_indices[0] and l[1] in matrix_indices[0]:
p1 = int(l[0])
p2 = int(l[1])
if p1 > p2:
p22 = p2
p2 = p1
p1 = p22
if (
p1 - area >= 0
and p1 + area + 1 < n
and p2 - area >= 0
and p2 + area + 1 < n
):
window = detrended[
np.ix_(
range(p1 - area, p1 + area + 1),
range(p2 - area, p2 + area + 1),
)
]
if (
len(window[window == 1.])
< ((area * 2 + 1) ** 2)
* undetermined_percentage
/ 100.
): # there should not be many indetermined bins
n_patterns += 1
score = res_rescaled[l[0], l[1]]
detected_patterns.append([name, l[0], l[1], score])
pattern_sums += window
pattern_windows.append(window)
else:
detected_patterns.append([name, "NA", "NA", "NA"])
else:
detected_patterns.append([name, "NA", "NA", "NA"])
# Computation of stats on the whole set - Agglomerated procedure :
for i in range(0, area * 2 + 1):
for j in range(0, area * 2 + 1):
list_temp = []
for el in range(1, len(pattern_windows)):
list_temp.append(pattern_windows[el][i, j])
agglomerated_pattern[i, j] = np.median(list_temp)
return detected_patterns, agglomerated_pattern
def main():
args = parse()
fname = args.file
true_egg_path = os.path.join(args.eggfolder, fname)
true_egg = np.load(true_egg_path)
true_egg = minmaxnormalize(true_egg)
if args.detrend:
_, true_egg = detrend(None, true_egg)
degg = np.gradient(true_egg, edge_order=2)
# region = slice(11000, 12000)
region = slice(8000, len(true_egg))
true_egg = true_egg[region]
degg = degg[region]
plt.figure()
fig = plt.gcf()
ax1 = plt.subplot(211)
plt.plot(true_egg, "b", label="EGG Waveform")
plt.ylabel("Amplitude")
ax2 = plt.subplot(212, sharex=ax1)
plt.plot(degg, "g", label="DEGG waveform")
plt.xlabel("Sample Number")
plt.ylabel("Amplitude")
# Timepoints in region shifted coordinates
t0 = 19750 # start of EGG cycle, GOI
t4 = 19853 # end of EGG cycle
t1 = t0 + np.argmax(true_egg[t0:t4]) # t0=19750 => t1=19760 EGG peak
t2 = t1 + np.argmin(degg[t1:t4]) # t0=19750 => t1=19760 => t2=19834 GCI
t5 = 19908 # start of new egg cycle
t9 = 20010 # end of new egg cycle
t6 = t5 + np.argmax(true_egg[t5:t9]) # EGG peak
t8 = t6 + np.argmin(degg[t6:t9]) # GCI
y1max = 1.0
y2min = -0.05
# Epochs
goi_ax1_1 = (t0, true_egg[t0])
goi_ax1_2 = (t5, true_egg[t5])
goi_ax2_1 = (t0, degg[t0])
goi_ax2_2 = (t5, degg[t5])
gci_ax1_1 = (t2, true_egg[t2])
gci_ax1_2 = (t8, true_egg[t8])
gci_ax2_1 = (t2, degg[t2])
gci_ax2_2 = (t8, degg[t8])
eggpeak_ax1_1 = (t1, true_egg[t1])
eggpeak_ax1_2 = (t6, true_egg[t6])
# eggstart_ax1_1 = (t0, true_egg[t0])
# eggstart_ax1_2 = (t5, true_egg[t5])
eggend_ax1_1 = (t4, true_egg[t4])
eggend_ax1_2 = (t9, true_egg[t9])
# Epoch Markers
goi_ax1 = np.array([goi_ax1_1, goi_ax1_2])
gci_ax1 = np.array([gci_ax1_1, gci_ax1_2])
goi_ax2 = np.array([goi_ax2_1, goi_ax2_2])
gci_ax2 = np.array([gci_ax2_1, gci_ax2_2])
eggpeak_ax1 = np.array([eggpeak_ax1_1, eggpeak_ax1_2])
# eggstart_ax1 = np.array([eggstart_ax1_1, eggstart_ax1_2])
eggend_ax1 = np.array([eggend_ax1_1, eggend_ax1_2])
# GOI scatter
ax1.scatter(goi_ax1[:, 0],
goi_ax1[:, 1],
c="tab:pink",
marker="x",
s=100,
label="GOI")
ax2.scatter(goi_ax2[:, 0],
goi_ax2[:, 1],
c="tab:pink",
marker="x",
s=100,
label="GOI")
# GCI Scatter
ax1.scatter(gci_ax1[:, 0],
gci_ax1[:, 1],
c="r",
marker="*",
s=100,
label="GCI")
ax2.scatter(gci_ax2[:, 0],
gci_ax2[:, 1],
c="r",
marker="*",
s=100,
label="GCI")
# EGG Peak Scatter
ax1.scatter(
eggpeak_ax1[:, 0],
eggpeak_ax1[:, 1],
# c="g",
marker="s",
s=100,
label="EGG Peak",
facecolors="none",
edgecolors="g",
)
# EGG Start Scatter
# ax1.scatter(
# eggstart_ax1[:, 0],
# eggstart_ax1[:, 1],
# # c="g",
# marker="^",
# s=100,
# label="EGG Cycle Start",
# facecolors="none",
# edgecolors="b",
# )
# EGG End Scatter
ax1.scatter(
eggend_ax1[:, 0],
eggend_ax1[:, 1],
# c="g",
marker="v",
s=100,
label="EGG Cycle End",
facecolors="none",
edgecolors="tab:purple",
)
# Across axes vertical lines
# Annotations
yoffset = 0.1
xoffset = 3
ax1.text(
t0 + xoffset,
true_egg[t0] + yoffset,
r"$t_0$",
ha="center",
va="center",
transform=ax1.transData,
fontsize=30,
)
ax1.text(
t1 + xoffset,
true_egg[t1] + yoffset,
r"$t_1$",
ha="center",
va="center",
transform=ax1.transData,
fontsize=30,
)
ax1.text(
t2 + xoffset,
true_egg[t2] + yoffset,
r"$t_2$",
ha="center",
va="center",
transform=ax1.transData,
fontsize=30,
)
ax1.text(
t4 + xoffset,
true_egg[t4] + yoffset,
r"$t_4$",
ha="center",
va="center",
transform=ax1.transData,
fontsize=30,
)
plt.subplots_adjust(top=0.962,
bottom=0.08,
left=0.053,
right=0.981,
hspace=0.116,
wspace=0.2)
# ax1.spines["bottom"].set_visible(False)
# ax2.spines["top"].set_visible(False)
ax1.get_xaxis().set_visible(False)
ax1.get_yaxis().set_label_coords(-0.037, 0.5)
ax2.get_yaxis().set_label_coords(-0.037, 0.5)
ax1.set_xlim([19715, 20015])
mng = plt.get_current_fig_manager()
mng.window.showMaximized()
ax1.legend()
ax2.legend()
plt.show()
def main():
args = parse()
fname = args.filename
eground = np.load(os.path.join(args.groundpath, fname))
egen = np.load(os.path.join(args.generatedpath, fname))
if len(eground) > len(egen):
eground = eground[:len(egen)]
if args.detrend:
_, eground = detrend(None, eground)
rawground, rawgen = eground, egen
voicedground, voicedgen = detect_voiced_region(rawground, rawgen)
voicedground = voicedground / np.max(np.abs(voicedground))
voicedgen = voicedgen / np.max(np.abs(voicedgen))
eground, egen = detect_voiced_region(eground, egen)
eground, egen = groundeggfilter(eground), geneggfilter(egen)
assert len(eground) == len(voicedground)
degground = groundegg_process(eground)
deggen = genegg_process(egen)
peaksposground = detectgroundwaveletgci(eground)
peaksground = positions2onehot(peaksposground, eground.shape)
peaksposgen = detectgenwaveletgci(egen)
peaksgen = positions2onehot(peaksposgen, egen.shape)
assert len(peaksgen) == len(peaksground)
metrics = corrected_naylor_metrics(peaksposground / 16e3,
peaksposgen / 16e3)
idr = metrics["identification_rate"]
msr = metrics["miss_rate"]
far = metrics["false_alarm_rate"]
ida = metrics["identification_accuracy"]
nhits = metrics["nhits"]
nmisses = metrics["nmisses"]
nfars = metrics["nfars"]
ncycles = metrics["ncycles"]
hits = metrics["hits"]
hits, hit_distances = zip(*hits)
fs = 16e3
hits = np.array(hits).squeeze() * fs
misses = np.array(metrics["misses"]).squeeze() * fs
fars = np.array(metrics["fars"]).squeeze() * fs
hits = hits.astype(np.int)
misses = misses.astype(np.int)
fars = fars.astype(np.int)
ax = plt.subplot(411)
plt.plot(voicedground, "r", label="ground truth")
plt.plot(voicedgen, "g", label="generated egg")
ax.set_xlim([-100, len(voicedground) + 100])
plt.gca().set_ylabel("amplitude")
plt.title("EGG")
plt.legend(loc=1)
plt.subplot(412, sharex=ax)
plt.plot(eground, "r", label="ground truth")
plt.plot(egen, "g", label="generated egg")
plt.gca().set_ylabel("amplitude")
plt.title("Proc EGG")
plt.legend(loc=1)
plt.subplot(413, sharex=ax)
plt.plot(degground, "r", label="ground truth")
plt.plot(deggen, "g", label="generated egg")
plt.gca().set_ylabel("amplitude")
plt.title("Neg DEGG")
plt.legend(loc=1)
plt.subplot(414, sharex=ax)
x = np.arange(len(peaksground))
lax = plt.gca()
lax.axhline(x[0], x[-1], 0, color="k")
lax.vlines(x, 0, peaksground, color="r", label="ground truth", linewidth=2)
lax.vlines(x,
0,
2 * positions2onehot(hits, peaksground.shape),
color="g",
label="hits")
lax.vlines(x,
0,
2 * positions2onehot(misses, peaksground.shape),
color="b",
label="misses")
lax.vlines(x,
0,
2 * positions2onehot(fars, peaksground.shape),
color="m",
label="fars")
# plt.plot(peaksground, "r", label="ground truth", linewidth=2)
# plt.plot(2 * peaksgen, "g", label="generated egg")
plt.gca().set_xlabel("sample")
plt.gca().set_ylabel("amplitude")
plt.title("GCI")
plt.legend(loc=1)
plt.subplots_adjust(top=0.91,
bottom=0.045,
left=0.035,
right=0.99,
hspace=0.2,
wspace=0.2)
plt.suptitle(
"{} IDR: {:2.2f} MR: {:2.2f} FAR: {:2.2f} IDA {:2.3f}\n H: {} M: {} F: {} C: {}"
.format(
fname,
idr * 100,
msr * 100,
far * 100,
ida * 1000,
nhits,
nmisses,
nfars,
ncycles,
))
mng = plt.get_current_fig_manager()
mng.window.showMaximized()
plt.show()
def gsr_preprocessing(signals):
''' Preprocessing for GSR signals '''
der_signals = np.gradient(signals)
con_signals = 1.0 / signals
nor_con_signals = (con_signals -
np.mean(con_signals)) / np.std(con_signals)
mean = np.mean(signals)
der_mean = np.mean(der_signals)
neg_der_mean = np.mean(der_signals[der_signals < 0])
neg_der_pro = float(der_signals[der_signals < 0].size) / float(
der_signals.size)
local_min = 0
for i in range(signals.shape[0] - 1):
if i == 0:
continue
if signals[i - 1] > signals[i] and signals[i] < signals[i + 1]:
local_min += 1
# Using SC calculates rising time
det_nor_signals, trend = detrend(nor_con_signals)
lp_det_nor_signals = butter_lowpass_filter(det_nor_signals, 0.5, 128.)
der_lp_det_nor_signals = np.gradient(lp_det_nor_signals)
rising_time = 0
rising_cnt = 0
for i in range(der_lp_det_nor_signals.size - 1):
if der_lp_det_nor_signals[i] > 0:
rising_time += 1
if der_lp_det_nor_signals[i + 1] < 0:
rising_cnt += 1
avg_rising_time = rising_time * (1. / 128.) / rising_cnt
freqs, power = getfreqs_power(signals,
fs=128.,
nperseg=signals.size,
scaling='spectrum')
power_0_24 = []
for i in range(21):
power_0_24.append(
getBand_Power(freqs,
power,
lower=0 + (i * 0.8 / 7),
upper=0.1 + (i * 0.8 / 7)))
SCSR, _ = detrend(butter_lowpass_filter(nor_con_signals, 0.2, 128.))
SCVSR, _ = detrend(butter_lowpass_filter(nor_con_signals, 0.08, 128.))
zero_cross_SCSR = 0
zero_cross_SCVSR = 0
peaks_cnt_SCSR = 0
peaks_cnt_SCVSR = 0
peaks_value_SCSR = 0.
peaks_value_SCVSR = 0.
zc_idx_SCSR = np.array([], int) # must be int, otherwise it will be float
zc_idx_SCVSR = np.array([], int)
for i in range(nor_con_signals.size - 1):
if SCSR[i] * next((j for j in SCSR[i + 1:] if j != 0), 0) < 0:
zero_cross_SCSR += 1
zc_idx_SCSR = np.append(zc_idx_SCSR, i + 1)
if SCVSR[i] * next((j for j in SCVSR[i + 1:] if j != 0), 0) < 0:
zero_cross_SCVSR += 1
zc_idx_SCVSR = np.append(zc_idx_SCVSR, i)
for i in range(zc_idx_SCSR.size - 1):
peaks_value_SCSR += np.absolute(
SCSR[zc_idx_SCSR[i]:zc_idx_SCSR[i + 1]]).max()
peaks_cnt_SCSR += 1
for i in range(zc_idx_SCVSR.size - 1):
peaks_value_SCVSR += np.absolute(
SCVSR[zc_idx_SCVSR[i]:zc_idx_SCVSR[i + 1]]).max()
peaks_cnt_SCVSR += 1
zcr_SCSR = zero_cross_SCSR / (nor_con_signals.size / 128.)
zcr_SCVSR = zero_cross_SCVSR / (nor_con_signals.size / 128.)
mean_peak_SCSR = peaks_value_SCSR / peaks_cnt_SCSR if peaks_cnt_SCSR != 0 else 0
mean_peak_SCVSR = peaks_value_SCVSR / peaks_cnt_SCVSR if peaks_value_SCVSR != 0 else 0
features = [mean, der_mean, neg_der_mean, neg_der_pro, local_min, avg_rising_time] + \
power_0_24 + [zcr_SCSR, zcr_SCVSR, mean_peak_SCSR, mean_peak_SCVSR]
return features
