def discreteWaveletAnalysis(vx, wDict):
from utilities import dataFromDict
try:
import pywt
except:
sys.exit(' Library pywt not installed. Exiting ...')
# nlevel = 4
order = 'freq' # "normal"
wavelet = dataFromDict('wavelet', wDict, allowNone=False)
nlevel = dataFromDict('nlevel', wDict, allowNone=False)
if (wavelet in pywt.wavelist()):
try:
wp = pywt.WaveletPacket(vx, wavelet, 'sym', maxlevel=nlevel)
except:
print(" Wrong wavelet type given. Reverting to default 'db2'. ")
wavelet = 'db2'
wp = pywt.WaveletPacket(vx, wavelet, 'sym', maxlevel=nlevel)
nodes = wp.get_level(nlevel, order=order)
labels = [n.path for n in nodes]
values = np.array([n.data for n in nodes], 'd')
values = abs(values)
return values, labels
def denoise_by_wavelets(audio,wavelet = 'dmey',threshold = 9):
"""
Audio denoising by using wavelet packet decomposition
Steps 1) Wavelet Packet decomposition 2) Thresholding 3) Reconstruction of wavelet packet decomposition.
:param audio:
:param wavelet:
:param threshold: Threshold used to remove noise (Actual threshold = threshold*std of
lowest level detail coefficients of the tree of wavelet packet decomposition)
:return: Denoised audio
"""
wp = pywt.WaveletPacket(data=audio, wavelet=wavelet, mode='symmetric')
new_wp = pywt.WaveletPacket(data=None, wavelet=wavelet, mode='symmetric')
ld = wp['d'].data
threshold = threshold*np.std(ld)
print("Denoising using wavelets for {} levels ... This may take a while".format(wp.maxlevel))
for i in range(wp.maxlevel):
paths = [node.path for node in wp.get_level(i+1, 'natural')]
for path in paths:
new_wp[path] = pywt.threshold(wp[path].data,threshold)
new_wp.reconstruct(update=True)
return new_wp.data
def denoise_wpa(self, Xin, Nin):
# print("Xin len " + str(len(Xin)))
X = pywt.WaveletPacket(Xin, self.waveletName, 'symmetric',
self.wlevels)
N = pywt.WaveletPacket(Nin, self.waveletName, 'symmetric',
self.wlevels)
Y = pywt.WaveletPacket(data=None,
wavelet=self.waveletName,
mode='symmetric')
# print("X len" + str(X.maxlevel))
# the output of pywt.wavedec will be arrays (trees) of size/length wlevels+1
# so at the index of 1[wlevels] we have the "leaf node"
nBands = pow(2, self.wlevels)
XleafNodes = [node.path for node in X.get_level(self.wlevels, 'freq')]
Ak = self.linearAk(nBands, self.akSlope)
# print(Ak)
# print(XleafNodes)
bandId = 0
numOfBands = len(XleafNodes)
counter = 0
for node in XleafNodes:
bandAk = Ak[bandId]
bandId += 1
XbandData = X[node].data
NbandData = N[node].data
YbandData = self.denoise_band(XbandData, NbandData, bandAk)
Y[node] = YbandData
counter = counter + 1
# reconstructing the audio from Y
Yaudio = Y.reconstruct(update=False)
return Yaudio
def data_wt(f, lenf_r, day, n, wavelet, mode):
f_r = np.empty(shape=(lenf_r, day, pow(2, n), 3))
for s in range(lenf_r):
f_i = f[s:day + s] / np.mean(f[s:day + s])
f_i_r = np.empty(shape=(day, pow(2, n), 3))
for i in range(3):
a = f_i[:, i]
a_w = pywt.WaveletPacket(data=a,
wavelet=wavelet,
mode=mode,
maxlevel=n)
# 获取特定层数的所有节点
npn = [node.path for node in a_w.get_level(n, 'natural')]
for j in range(len(npn)):
new_aw = pywt.WaveletPacket(data=None,
wavelet=wavelet,
mode=mode)
new_aw[npn[j]] = a_w[npn[j]]
a_r_j = new_aw.reconstruct(update=True)
f_i_r[:, j, i] = a_r_j
f_r[s] = f_i_r
print(f_r[-1])
c = np.zeros(shape=(day, ))
for i in range(pow(2, n)):
c = c + f_r[-1, :, i, 0]
print(c - f_i[:, 0])
return f_r
def transform(self, X, y=None):
def Energy(coeffs, k):
return np.sqrt(np.sum(np.array(coeffs[-k])**2)) / len(coeffs[-k])
def getEnergy(wp):
coefs = np.asarray([n.data for n in wp.get_leaf_nodes(True)])
return np.asarray(
[Energy(coefs, i) for i in range(2**wp.maxlevel)])
if X.shape[2] == 1:
return np.array([
getEnergy(
pywt.WaveletPacket(data=x,
wavelet='db4',
mode='symmetric',
maxlevel=4)) for x in X[:, :, 0]
])
de = np.array([
getEnergy(
pywt.WaveletPacket(data=x,
wavelet='db4',
mode='symmetric',
maxlevel=4)) for x in X[:, :, 0]
])
fe = np.array([
getEnergy(
pywt.WaveletPacket(data=x,
wavelet='db4',
mode='symmetric',
maxlevel=4)) for x in X[:, :, 1]
])
return np.concatenate((de, fe), axis=1)
def reconstruct():
""" Take only the lowest-frequency component of the WPT. """
npzfile = np.load("./data/rawInputs.npz")
raw = npzfile['arr_0']
wave, waveTS, dataE, dataRT = raw[0], raw[1], raw[2], raw[3]
wp = pywt.WaveletPacket(data=wave,
wavelet='haar',
mode='symmetric',
maxlevel=3)
new_wp = pywt.WaveletPacket(data=None, wavelet='haar', mode='symmetric')
new_wp['aaa'] = wp['aaa'].data
newWave = new_wp.reconstruct(update=False)
# plots
fig = plt.figure(figsize=(8, 5), facecolor='w')
a1 = plt.subplot(111)
a1.set_title("Energy %.1f keV RiseTime %.1f ns" % (dataE, dataRT))
a1.set_ylabel("ADC [A.U.]", y=0.95, ha='right')
a1.set_xlabel("Time (ns)", x=0.95, ha='right')
a1.plot(waveTS, wave, color='blue', alpha=0.8, label='Original Waveform')
a1.plot(np.arange(0,
len(newWave) * 10, 10),
newWave,
color='red',
label='Lvl3 Haar Denoised WF')
a1.legend(loc=4)
plt.show(block=False)
plt.savefig("./plots/denoised-wf.pdf")
def test_reconstructing_data():
x = [1, 2, 3, 4, 5, 6, 7, 8]
wp = pywt.WaveletPacket(data=x, wavelet='db1', mode='sym')
# Create another Wavelet Packet and feed it with some data.
new_wp = pywt.WaveletPacket(data=None, wavelet='db1', mode='sym')
new_wp['aa'] = wp['aa'].data
new_wp['ad'] = [-2., -2.]
# For convenience, :attr:`Node.data` gets automatically extracted
# from the :class:`Node` object:
new_wp['d'] = wp['d']
# Reconstruct data from aa, ad, and d packets.
assert_allclose(new_wp.reconstruct(update=False), x, rtol=1e-12)
# The node's :attr:`~Node.data` will not be updated
assert_(new_wp.data is None)
# When `update` is True:
assert_allclose(new_wp.reconstruct(update=True), x, rtol=1e-12)
assert_allclose(new_wp.data, np.arange(1, 9), rtol=1e-12)
assert_([n.path
for n in new_wp.get_leaf_nodes(False)] == ['aa', 'ad', 'd'])
assert_([n.path for n in new_wp.get_leaf_nodes(True)] ==
['aaa', 'aad', 'ada', 'add', 'daa', 'dad', 'dda', 'ddd'])
def reconstructExample():
x = [1, 2, 3, 4, 5, 6, 7, 8]
wp = pywt.WaveletPacket(data=x, wavelet='db1', mode='symmetric')
# Now create a new Wavelet Packet and set its nodes with some data.
new_wp = pywt.WaveletPacket(data=None, wavelet='db1', mode='symmetric')
new_wp['aa'] = wp['aa'].data
new_wp['ad'] = [-2., -2.]
# For convenience, Node.data gets automatically extracted from the Node object:
new_wp['d'] = wp['d']
# And reconstruct the data from the aa, ad and d packets.
print(new_wp.reconstruct(update=False))
# [ 1. 2. 3. 4. 5. 6. 7. 8.]
# If the update param in the reconstruct method is set to False, the node's data will not be updated.
print(new_wp.data)
# None
# Otherwise, the data attribute will be set to the reconstructed value.
print(new_wp.reconstruct(update=True))
# [ 1. 2. 3. 4. 5. 6. 7. 8.]
print(new_wp.data)
# [ 1. 2. 3. 4. 5. 6. 7. 8.]
print([n.path for n in new_wp.get_leaf_nodes(False)])
# ['aa', 'ad', 'd']
print([n.path for n in new_wp.get_leaf_nodes(True)])
def gety():
ys = sampley()
ploty(ys, "original")
wp = pywt.WaveletPacket(data=ys, wavelet='db4')
new_wp = pywt.WaveletPacket(data=None, wavelet='db4')
new_wp['a'] = wp['a'].data
recy = new_wp.reconstruct()
afilt = np.fft.fft(recy)
abins = np.fft.fftfreq(len(wp['a'].data), d=1 / 400)
todrop = np.where(abins < len(wp['a'].data) / 4) or np.where(
abins > 3 * len(wp['a'].data) / 4)
afilt[todrop] = 0
recy = np.fft.ifft(afilt)
ploty(recy, "a")
new_wp = pywt.WaveletPacket(data=None, wavelet='db4')
#new_wp['a'] = np.zeros(len(wp['a'].data))
new_wp['d'] = wp['d'].data
recy = new_wp.reconstruct()
afilt = np.fft.fft(recy)
abins = np.fft.fftfreq(len(wp['d'].data), d=1 / 400)
todrop = np.where(abins < len(wp['d'].data) / 4) or np.where(
abins > 3 * len(wp['d'].data) / 4)
afilt[todrop] = 0
recy = np.fft.ifft(afilt)
ploty(recy, "d")
plt.show()
def run(self, data) -> list:
close = data.close.values[:-7]
wavelet = 'db3'
level = 4
wp = pywt.WaveletPacket(close, wavelet, maxlevel=level)
wl_list = []
for i in range(level):
new_wp = pywt.WaveletPacket(None, wavelet, maxlevel=level)
node = 'a' * (i + 1)
new_wp[node] = wp[node]
wl = new_wp.reconstruct()
wl_list.append(wl)
for i, wl in enumerate(wl_list):
if i < level - 1:
continue
self._add_technique('WL{0}'.format(i + 1),
wl,
x_axis=np.arange(wl.size))
# node_list = ['daa', 'ada', 'aad']
# wld_list = []
# for i in range(level):
# new_wp = pywt.WaveletPacket(None, wavelet, maxlevel=level)
# node = node_list[i]
# new_wp[node] = wp[node]
# wld = new_wp.reconstruct()
# wld_list.append(wld)
# for i, wl in enumerate(wld_list):
# self._add_technique('D{0}'.format(i + 1), wl, 2, x_axis=np.arange(wl.size))
return super().run(data)
def __check_pywt(features=None):
"""Check for available functionality within pywt
Parameters
----------
features : list of str
List of known features to check such as 'wp reconstruct',
'wp reconstruct fixed'
"""
import pywt
import numpy as np
data = np.array(
[0.57316901, 0.65292526, 0.75266733, 0.67020084, 0.46505364,
0.76478331, 0.33034164, 0.49165547, 0.32979941, 0.09696717,
0.72552711, 0.4138999, 0.54460628, 0.786351, 0.50096306,
0.72436454, 0.2193098, -0.0135051, 0.34283984, 0.65596245,
0.49598417, 0.39935064, 0.26370727, 0.05572373, 0.40194438,
0.47004551, 0.60327258, 0.25628266, 0.32964893, 0.24009889])
mode = 'per'
wp = pywt.WaveletPacket(data, 'sym2', mode)
wp2 = pywt.WaveletPacket(data=None, wavelet='sym2', mode=mode)
try:
for node in wp.get_level(2): wp2[node.path] = node.data
except:
raise ImportError(
"Failed to reconstruct WP by specifying data in the layer")
if 'wp reconstruct fixed' in features:
rec = wp2.reconstruct()
if np.linalg.norm(rec[:len(data)] - data) > 1e-3:
raise ImportError(
"Failed to reconstruct WP correctly")
return True
def test_wavelet_packet_dtypes():
N = 32
for dtype in [np.float32, np.float64, np.complex64, np.complex128]:
x = np.random.randn(N).astype(dtype)
if np.iscomplexobj(x):
x = x + 1j * np.random.randn(N).astype(x.real.dtype)
wp = pywt.WaveletPacket(data=x, wavelet='db1', mode='symmetric')
# no unnecessary copy made
assert_(wp.data is x)
# assiging to a node should not change supported dtypes
wp['d'] = wp['d'].data
assert_equal(wp['d'].data.dtype, x.dtype)
# full decomposition
wp.get_level(wp.maxlevel)
# reconstruction from coefficients should preserve dtype
r = wp.reconstruct(False)
assert_equal(r.dtype, x.dtype)
assert_allclose(r, x, atol=1e-5, rtol=1e-5)
# first element of the tuple is the input dtype
# second element of the tuple is the transform dtype
dtype_pairs = [
(np.uint8, np.float64),
(np.intp, np.float64),
]
if hasattr(np, "complex256"):
dtype_pairs += [
(np.complex256, np.complex128),
]
if hasattr(np, "half"):
dtype_pairs += [
(np.half, np.float32),
]
for (dtype, transform_dtype) in dtype_pairs:
x = np.arange(N, dtype=dtype)
wp = pywt.WaveletPacket(x, wavelet='db1', mode='symmetric')
# no unnecessary copy made of top-level data
assert_(wp.data is x)
# full decomposition
wp.get_level(wp.maxlevel)
# reconstructed data will have modified dtype
r = wp.reconstruct(False)
assert_equal(r.dtype, transform_dtype)
assert_allclose(r, x.astype(transform_dtype), atol=1e-5, rtol=1e-5)
def return_iwv_PACKET_fix_levels(x, fs, max_level, mother_wv, idx_levels):
n = len(x)
nico = pywt.WaveletPacket(data=x, wavelet=mother_wv, maxlevel=max_level)
mylevels = [node.path for node in nico.get_level(max_level, 'freq')]
print('inverse WV!')
gato = pywt.WaveletPacket(data=None, wavelet=mother_wv, maxlevel=max_level)
for idx in idx_levels:
gato[mylevels[idx]] = nico[mylevels[idx]].data
outsignal = gato.reconstruct(update=False)
return outsignal
def frequency_draw(data2: np.ndarray):
wp = pywt.WaveletPacket(data=data2,
wavelet='db3',
mode='symmetric',
maxlevel=2) #2层小波包分解
new_wp = pywt.WaveletPacket(data=None,
wavelet='db3',
mode='symmetric',
maxlevel=2) #新建小波包树用来重构4个节点系数
new_wp['aa'] = wp['aa']
LL = new_wp.reconstruct(update=False)
del (new_wp['aa'])
new_wp['ad'] = wp['ad']
LH = new_wp.reconstruct(update=False)
del (new_wp['a'])
new_wp['da'] = wp['da']
HL = new_wp.reconstruct(update=False)
del (new_wp['da'])
new_wp['dd'] = wp['dd']
HH = new_wp.reconstruct(update=False)
#4个频率分量的能量
LLPE = np.sum(LL * LL)
LHPE = np.sum(LH * LH)
HLPE = np.sum(HL * HL)
HHPE = np.sum(HH * HH)
E = [LLPE, LHPE, HLPE, HHPE]
P = E / np.sum(E)
WE = -np.sum(np.log2(P) * P) #信息熵
WIQ = np.sum(P * E) #信息量
S = np.array([LL, LH, HL, HH]) #小波系数矩阵4*33
U, sigma, VT = np.linalg.svd(S, full_matrices=1, compute_uv=1)
feature = np.empty(10)
feature[0] = LLPE
feature[1] = LHPE
feature[2] = HLPE
feature[3] = HHPE
feature[4] = WE
feature[5] = WIQ
feature[6] = sigma[0]
feature[7] = sigma[1]
feature[8] = sigma[2]
feature[9] = sigma[3]
return feature
def run(self, data) -> list:
close_all = data.close.values
wavelet = 'db2'
level = 4
wla = [np.nan]
for i in range(close_all.size):
close = close_all[:i + 1]
wp = pywt.WaveletPacket(close, wavelet, maxlevel=level)
new_wp = pywt.WaveletPacket(None, wavelet, maxlevel=level)
node = 'a' * level
new_wp[node] = wp[node]
wl = new_wp.reconstruct()
wla.append(wl[i])
self._add_technique('WLA', wla, x_axis=np.arange(len(wla)))
return super().run(data)
def mjdWaveform():
# Load a data and a temp waveform
npzfile = np.load("./data/rawInputs.npz")
raw = npzfile['arr_0']
wave, waveTS, dataE, dataRT = raw[0], raw[1], raw[2], raw[3]
wp = pywt.WaveletPacket(wave, 'db2', 'symmetric', maxlevel=4)
nodes = wp.get_level(4, order='freq', decompose=True)
yWT = np.array([n.data for n in nodes], 'd')
yWT = abs(yWT)
waveletYTrans = yWT # final coeff's (used in process-waveforms)
rec = wp.reconstruct(update=True)
# plots
fig = plt.figure(figsize=(8, 8), facecolor='w')
a1 = plt.subplot(311)
a1.plot(waveTS, wave, color='blue')
a2 = plt.subplot(312)
a2.imshow(waveletYTrans,
interpolation='nearest',
aspect="auto",
origin="lower",
extent=[0, 1, 0, len(waveletYTrans)],
cmap='jet')
a3 = plt.subplot(313)
a3.plot(waveTS, rec, color='red')
plt.show()
def _compare_trees1(
wavelet_str: str,
max_lev: int = 3,
pywt_boundary: str = "zero",
ptwt_boundary: str = "zero",
length: int = 256,
batch_size: int = 1,
):
data = np.random.rand(batch_size, length)
wavelet = pywt.Wavelet(wavelet_str)
twp = WaveletPacket(torch.from_numpy(data), wavelet, mode=ptwt_boundary)
nodes = twp.get_level(max_lev)
twp_lst = []
for node in nodes:
twp_lst.append(twp[node])
torch_res = torch.cat(twp_lst, -1).numpy()
np_batches = []
for batch_index in range(batch_size):
wp = pywt.WaveletPacket(data=data[batch_index],
wavelet=wavelet,
mode=pywt_boundary)
nodes = [node.path for node in wp.get_level(max_lev, "freq")]
np_lst = []
for node in nodes:
np_lst.append(wp[node].data)
np_res = np.concatenate(np_lst, -1)
np_batches.append(np_res)
np_batches = np.stack(np_batches, 0)
assert np.allclose(torch_res, np_batches)
def __reverse_single_level(self, wp):
# local bindings
level_paths = self.__level_paths
# define wavelet packet to use
WP = pywt.WaveletPacket(data=None,
wavelet=self._wavelet,
mode=self._mode,
maxlevel=self.__level)
# prepare storage
signal_shape = wp.shape[:1] + self._inshape[1:]
signal = np.zeros(signal_shape)
Ntime_points = self._intimepoints
for indexes in _get_indexes(signal_shape, self._dim):
if __debug__:
debug('MAP_', " %s" % (indexes, ), lf=False, cr=True)
for path, level_data in zip(level_paths, wp[indexes]):
WP[path] = level_data
signal[indexes] = WP.reconstruct(True)[:Ntime_points]
return signal
def test_removing_nodes():
x = [1, 2, 3, 4, 5, 6, 7, 8]
wp = pywt.WaveletPacket(data=x, wavelet='db1', mode='sym')
wp.get_level(2)
dataleafs = [n.data for n in wp.get_leaf_nodes(False)]
expected = np.array([[5., 13.], [-2, -2], [-1, -1], [0, 0]])
for i in range(4):
assert_allclose(dataleafs[i], expected[i, :], atol=1e-12)
node = wp['ad']
del (wp['ad'])
dataleafs = [n.data for n in wp.get_leaf_nodes(False)]
expected = np.array([[5., 13.], [-1, -1], [0, 0]])
for i in range(3):
assert_allclose(dataleafs[i], expected[i, :], atol=1e-12)
wp.reconstruct()
# The reconstruction is:
assert_allclose(wp.reconstruct(),
np.array([2., 3., 2., 3., 6., 7., 6., 7.]),
rtol=1e-12)
# Restore the data
wp['ad'].data = node.data
dataleafs = [n.data for n in wp.get_leaf_nodes(False)]
expected = np.array([[5., 13.], [-2, -2], [-1, -1], [0, 0]])
for i in range(4):
assert_allclose(dataleafs[i], expected[i, :], atol=1e-12)
assert_allclose(wp.reconstruct(), np.arange(1, 9), rtol=1e-12)
def waveletPacket(self, packlevel):
"""After obtaining the frame, the Wavelet Packet Energy (WPE) feature
is obtain from the frame using the Wavelet Packe method.
Parameters
----------
packlevel: int, the quantity of the frequency bands of the frequency. Larger
packlevel, higher frequency resolution and more generated features.
2^ packlevel must smaller than the frame data.
"""
Energy = []
Flatness = []
self.maxWPE = []
for clipindex in xrange(len(self.cutclip)):
tempE = []
tempF = []
wp= pywt.WaveletPacket(data=self.cutclip[clipindex,:],wavelet='db1',mode='symmetric', maxlevel = packlevel)
for i in xrange(packlevel+1):
for index, node in enumerate(wp.get_level(i)):
d = wp[node.path].data
E = np.log(np.sqrt(np.sum(d ** 2)))
F = np.exp(np.mean(np.log(np.abs(d)))) / np.mean(np.abs(d))
tempE.append(E)
tempF.append(F)
maxnumE = float(max(tempE))
temp = list(np.array(tempE) / maxnumE) # this function will deliminate the effect of the amplitude
self.maxWPE.append(maxnumE)
Energy.append(tempE)
Flatness.append(tempF)
self.maxWPE = np.array(self.maxWPE)
self.WPE = np.matrix(Energy)
self.WPE = self.checkmatrix(self.WPE)
self.WPF = np.matrix(Flatness)
self.WPF = self.checkmatrix(self.WPF)
def wpe2(matrix):
t_wpe = np.zeros((matrix.shape[0], 16))
# Step1: The Wavelet Packet Decomposition using the Daubechies wavelet in mode symmetric.
for i in range(matrix.shape[0]):
wp = pywt.WaveletPacket(data=matrix[i], wavelet='db1', mode='symmetric')
# Step2: Computing the Wavelet Packet Energy of each of 16 Wavelet Packet Decomposition sub-bands.
t_wpe[i, 0] = wpe(wp['aaaa'].data)
t_wpe[i, 1] = wpe(wp['aaad'].data)
t_wpe[i, 2] = wpe(wp['aada'].data)
t_wpe[i, 3] = wpe(wp['aadd'].data)
t_wpe[i, 4] = wpe(wp['adaa'].data)
t_wpe[i, 5] = wpe(wp['adad'].data)
t_wpe[i, 6] = wpe(wp['adda'].data)
t_wpe[i, 7] = wpe(wp['addd'].data)
t_wpe[i, 8] = wpe(wp['daaa'].data)
t_wpe[i, 9] = wpe(wp['daad'].data)
t_wpe[i,10] = wpe(wp['dada'].data)
t_wpe[i,11] = wpe(wp['dadd'].data)
t_wpe[i,12] = wpe(wp['ddaa'].data)
t_wpe[i,13] = wpe(wp['ddad'].data)
t_wpe[i,14] = wpe(wp['ddda'].data)
t_wpe[i,15] = wpe(wp['dddd'].data)
return t_wpe
def extration_wavelet_packet(wavFile):
sig, fs = librosa.load(wavFile)
# print("fs is ",fs)
# print("signal length is ", len(sig))
N_signal = len(sig)
n_level = 6
#N_sub = N_signal/(2**n_level)
#print sig
#print ("N sub is ",N_sub)
# x = sig.reshape((len(sig),1))
# print x
wp = pywt.WaveletPacket(data=sig, wavelet='db3', mode='symmetric')
#
# print wp.data
#print wp.maxlevel
#print len(wp.data)
Node = []
N_limit = 16
limit = 0
for node in wp.get_level(6, 'natural'):
Node.append(node.path)
print("sub sample is ",len(Node))
TC = get_teager_energy(wp,Node,n_level,N_limit)
print ("TC is ",TC)
return TC
def test_packet_harbo_lvl3():
"""From Jensen, La Cour-Harbo, Rippels in Mathematics, Chapter 8 (page 89)."""
data = np.array([56.0, 40.0, 8.0, 24.0, 48.0, 48.0, 40.0, 16.0])
class _MyHaarFilterBank(object):
@property
def filter_bank(self):
"""Unscaled Haar wavelet filters."""
return (
[1 / 2, 1 / 2.0],
[-1 / 2.0, 1 / 2.0],
[1 / 2.0, 1 / 2.0],
[1 / 2.0, -1 / 2.0],
)
wavelet = pywt.Wavelet("unscaled Haar Wavelet",
filter_bank=_MyHaarFilterBank())
twp = WaveletPacket(torch.from_numpy(data), wavelet, mode="reflect")
twp_nodes = twp.get_level(3)
twp_lst = []
for node in twp_nodes:
twp_lst.append(torch.squeeze(twp[node]))
torch_res = torch.stack(twp_lst).numpy()
wp = pywt.WaveletPacket(data=data, wavelet=wavelet, mode="reflect")
pywt_nodes = [node.path for node in wp.get_level(3, "freq")]
np_lst = []
for node in pywt_nodes:
np_lst.append(wp[node].data)
np_res = np.concatenate(np_lst)
assert np.allclose(torch_res, np_res)
def waveletTransform(signalRaw, level=4, wavelet='db2', order='freq'):
""" Use PyWavelets to do a wavelet transform. """
wp = pywt.WaveletPacket(signalRaw, wavelet, 'symmetric', maxlevel=level)
nodes = wp.get_level(level, order=order)
yWT = np.array([n.data for n in nodes], 'd')
yWT = abs(yWT)
return wp.data, yWT
def get_wavelet_envelopes(frames, level=5, window_type='hann'):
'''
Decomposes input audio with wavelet packets and calculates energy envelopes of their components
Parameters:
frames : overlapping signal frames for short-time analysis
level : number of levels of wavelet decomposition, 2**level gives the final number of wavelet components
window_type : type of windowing function to apply
Returns numpy 2D array af
'''
frame_size = len(frames[0])
num_bands = 2**level
output_envelopes = {i: [] for i in range(num_bands)}
windowing = es.Windowing(type='hann', size=frame_size)
for frame in frames:
wp = pywt.WaveletPacket(data=windowing(frame),
wavelet='db1',
mode='zero',
maxlevel=level)
for i in range(num_bands):
band_key = bin(i).replace('0b', '').zfill(level).replace(
'0', 'a').replace('1', 'd')
output_envelopes[i].append(np.std(wp[band_key].data))
output_array = []
for item in output_envelopes.values():
output_array.append(list(item))
return np.array(output_array)
def WPT(X):
y = []
for x in X:
wp = pywt.WaveletPacket(data=x, wavelet='db1', mode='symmetric')
temp = [node.data[0] for node in wp.get_level(wp.maxlevel, 'natural')]
y.append(temp)
return np.array(y)
def wavelet_tree(self, wavelet='db1', analyze_level=6):
sound = self.org_sound
# 进行小波包分解
wp = pywt.WaveletPacket(data=sound, wavelet=wavelet, mode='sym')
# 获取小波包分解树的各个节点名称
node_names = [
node.path for node in wp.get_level(analyze_level, 'natural')
]
# 获取 概要部分节点名称 和 细节部分节点名称
node_names_approx = node_names[:(2**analyze_level // 2)]
node_names_detail = node_names[(2**analyze_level // 2):]
# 分别获取 概要部分的信号 和 细节部分的信号
approx_array = [pywt.upcoef('a', wp[node].data, wavelet, level=analyze_level) \
for node in node_names_approx]
detail_array = [pywt.upcoef('d', wp[node].data, wavelet, level=analyze_level) \
for node in node_names_detail]
self.approx_array = approx_array
self.detail_array = detail_array
self.wave_mode = 'wavelet_tree'
self.level = analyze_level
self.col_name = [node.path for node in \
wp.get_level(analyze_level, 'natural')]
return approx_array, detail_array
def getFeature(path):
with open(path, 'r', encoding='utf-8') as f:
for line in f:
a = line.split(',')
a = np.array(a).astype(float)
x, y = signal.butter(1, 0.4, 'highpass')
a = signal.filtfilt(x, y, a)
u = [0 for i in range(8)]
for i in range(100):
wp = pywt.WaveletPacket(a[10000 * i:10000 * (i + 1)],
wavelet='db3',
mode='symmetric',
maxlevel=3)
u[0] += np.linalg.norm(wp['aaa'].data, ord=None)
u[1] += np.linalg.norm(wp['aad'].data, ord=None)
u[2] += np.linalg.norm(wp['ada'].data, ord=None)
u[3] += np.linalg.norm(wp['add'].data, ord=None)
u[4] += np.linalg.norm(wp['daa'].data, ord=None)
u[5] += np.linalg.norm(wp['dad'].data, ord=None)
u[6] += np.linalg.norm(wp['dda'].data, ord=None)
u[7] += np.linalg.norm(wp['ddd'].data, ord=None)
u[0] /= 100
u[1] /= 100
u[2] /= 100
u[3] /= 100
u[4] /= 100
u[5] /= 100
u[6] /= 100
u[7] /= 100
return u
def create_wavelet_packet(self):
for item in self.hamming_window_list:
item = np.array(item)
wp = pywt.WaveletPacket(data=item,
wavelet='db1',
mode='symmetric',
maxlevel=4)
self.wavelet_list.append(wp['aaaa'].data)
self.wavelet_list.append(wp['aaad'].data)
self.wavelet_list.append(wp['aada'].data)
self.wavelet_list.append(wp['aadd'].data)
self.wavelet_list.append(wp['adaa'].data)
self.wavelet_list.append(wp['adad'].data)
self.wavelet_list.append(wp['adda'].data)
self.wavelet_list.append(wp['addd'].data)
self.wavelet_list.append(wp['daaa'].data)
self.wavelet_list.append(wp['daad'].data)
self.wavelet_list.append(wp['dada'].data)
self.wavelet_list.append(wp['dadd'].data)
self.wavelet_list.append(wp['ddaa'].data)
self.wavelet_list.append(wp['ddad'].data)
self.wavelet_list.append(wp['ddda'].data)
self.wavelet_list.append(wp['dddd'].data)
self.standard_deviation()
print(
f'Wavelet packet created + calculated std. deviation from {[self.data_type]}'
)
def test_wavelet_packet_structure():
x = [1, 2, 3, 4, 5, 6, 7, 8]
wp = pywt.WaveletPacket(data=x, wavelet='db1', mode='sym')
assert_(wp.data == [1, 2, 3, 4, 5, 6, 7, 8])
assert_(wp.path == '')
assert_(wp.level == 0)
assert_(wp['ad'].maxlevel == 3)
