def main():
# Define matrix dimensions
Nobs = 1000 # Number of observation
Nvars = 50000 # Number of variables
Ncomp = 100 # Number of components
# Simulated true sources
S_true = np.random.logistic(0,1,(Ncomp,Nvars))
# Simulated true mixing
A_true = np.random.normal(0,1,(Nobs,Ncomp))
# X = AS
X = np.dot(A_true,S_true)
# add some noise
X = X + np.random.normal(0,1,X.shape)
# apply ICA on X and ask for 2 components
model = ica1(Ncomp)
start = time.time()
A,S = model.fit(X)
total = time.time() - start
print('total time: {}'.format(total))
# compare if our estimates are accurate
# correlate A with Atrue and take
aCorr = np.abs(np.corrcoef(A.T,A_true.T)[:Ncomp,Ncomp:]).max(axis = 0).mean()
sCorr = np.abs(np.corrcoef(S,S_true)[:Ncomp,Ncomp:]).max(axis = 0).mean()
print("Accuracy of estimated sources: %.2f"%sCorr)
print("Accuracy of estimated mixing: %.2f"%aCorr)
def test_mv_rejection(self):
# three dimensional input samples
old_data, labels = synthetic_data()
gen = DataGenerator(old_data, n_components=3,
n_samples=1000, n_batches=10,
method='rejective')
new_data = [x for x in gen]
for i in range(10):
self.assertEqual(new_data[i].shape[0],1000)
old_model = gen.parameters['sample_hist']
model = ica1(3)
# foreach data batch
for i in range(10):
new_A, new_S = model.fit(new_data[i])
# Reorder components
c_SS = corrcoef(new_S, gen.sources)[3:,:3]
source_sim = abs(c_SS).max(axis=1)
# Check sources are similar
self.assertTrue(all(source_sim>0.95))
order = abs(c_SS).argmax(axis=1)
signs = array([sign(x[order[n]])
for n,x in enumerate(c_SS)])
new_S = new_S[order,:] * signs.reshape((-1,1))
new_A = new_A[:,order] * signs.reshape((1,-1))
new_model = [histogram(column, density=True, bins=20)\
for column in new_A.T]
# Check that resulting new mixing has similar histogram
for j in range(len(new_model)):
sim = abs(corrcoef(new_model[j][0], old_model[j][0]))
self.assertTrue( sim[0,1] > 0.8,
'simmilarity {} is too low'.format(sim))
def main():
# Define matrix dimensions
Nobs = 1000 # Number of observation
Nvars = 50000 # Number of variables
Ncomp = 100 # Number of components
# Simulated true sources
S_true = np.random.logistic(0, 1, (Ncomp, Nvars))
# Simulated true mixing
A_true = np.random.normal(0, 1, (Nobs, Ncomp))
# X = AS
X = np.dot(A_true, S_true)
# add some noise
X = X + np.random.normal(0, 1, X.shape)
# apply ICA on X and ask for 2 components
model = ica1(Ncomp)
start = time.time()
A, S = model.fit(X)
total = time.time() - start
#print('total time: {}'.format(total))
# compare if our estimates are accurate
# correlate A with Atrue and take
aCorr = np.abs(np.corrcoef(A.T, A_true.T)[:Ncomp,
Ncomp:]).max(axis=0).mean()
sCorr = np.abs(np.corrcoef(S, S_true)[:Ncomp, Ncomp:]).max(axis=0).mean()
def test_ICA_infomax_clean(self):
start = time.time()
A, S = ica.ica1(self.clean_data, self.NCOMP)
end = time.time()
print('\ttime: {}:.2f'.format(end - start))
# Check right dimensions of Output
self.assertEqual(A.shape, (self.NSUB, self.NCOMP))
self.assertEqual(S.shape, (self.NCOMP, self.NVOX))
idx = find_sources_order(self.sources, S)
S = S[idx, :]
A = A[:, idx]
# Check the accuracy of output
self.assertGreater(mean_corr(self.sources, S), 0.95)
self.assertGreater(mean_corr(self.loading, A), 0.95)
def test_ICA_infomax_clean(self):
start = time.time()
A, S = ica.ica1(self.clean_data, self.NCOMP)
end = time.time()
print('\ttime: {}:.2f'.format(end - start))
# Check right dimensions of Output
self.assertEqual(A.shape, (self.NSUB, self.NCOMP))
self.assertEqual(S.shape, (self.NCOMP, self.NVOX))
idx = find_sources_order(self.sources, S)
S = S[idx, :]
A = A[:, idx]
# Check the accuracy of output
self.assertGreater(mean_corr(self.sources, S), 0.95)
self.assertGreater(mean_corr(self.loading, A), 0.95)
def test_mv_normal(self):
# three dimensional input samples
old_data, labels = synthetic_data()
gen = DataGenerator(old_data, n_components=3,
n_samples=1000, n_batches=10,
method='normal')
new_data = [x for x in gen]
# Checking that there is enough samples
for i in range(10):
self.assertEqual(new_data[i].shape[0],1000)
old_model = (gen.parameters['sample_mean'],
gen.parameters['sample_cov'])
model = ica1(3)
# foreach data batch
for i in range(10):
new_A, new_S = model.fit(new_data[i])
# Reorder components
c_SS = corrcoef(new_S, gen.sources)[3:,:3]
source_sim = abs(c_SS).max(axis=1)
# Check sources are similar
self.assertTrue(all(source_sim>0.95))
order = abs(c_SS).argmax(axis=1)
signs = array([sign(x[order[n]])
for n,x in enumerate(c_SS)])
new_S = new_S[order,:] * signs.reshape((-1,1))
new_A = new_A[:,order] * signs.reshape((1,-1))
new_model = (new_A.mean(axis=0),
cov(new_A, rowvar=0))
# Check that resulting new mixing has similar mean and cov
error = norm(abs(new_model[0] - old_model[0]))/3
self.assertTrue(error < 0.01)
error = norm(abs(new_model[1] - old_model[1]).ravel())/9
self.assertTrue(error < 0.01)
def test_mv_normal(self):
# three dimensional input samples
old_data, labels = synthetic_data()
gen = DataGenerator(old_data,
n_components=3,
n_samples=1000,
n_batches=10,
method='normal')
new_data = [x for x in gen]
# Checking that there is enough samples
for i in range(10):
self.assertEqual(new_data[i].shape[0], 1000)
old_model = (gen.parameters['sample_mean'],
gen.parameters['sample_cov'])
model = ica1(3)
# foreach data batch
for i in range(10):
new_A, new_S = model.fit(new_data[i])
# Reorder components
c_SS = corrcoef(new_S, gen.sources)[3:, :3]
source_sim = abs(c_SS).max(axis=1)
# Check sources are similar
self.assertTrue(all(source_sim > 0.95))
order = abs(c_SS).argmax(axis=1)
signs = array([sign(x[order[n]]) for n, x in enumerate(c_SS)])
new_S = new_S[order, :] * signs.reshape((-1, 1))
new_A = new_A[:, order] * signs.reshape((1, -1))
new_model = (new_A.mean(axis=0), cov(new_A, rowvar=0))
# Check that resulting new mixing has similar mean and cov
error = norm(abs(new_model[0] - old_model[0])) / 3
self.assertTrue(error < 0.01)
error = norm(abs(new_model[1] - old_model[1]).ravel()) / 9
self.assertTrue(error < 0.01)
def test_mv_rejection(self):
# three dimensional input samples
old_data, labels = synthetic_data()
gen = DataGenerator(old_data,
n_components=3,
n_samples=1000,
n_batches=10,
method='rejective')
new_data = [x for x in gen]
for i in range(10):
self.assertEqual(new_data[i].shape[0], 1000)
old_model = gen.parameters['sample_hist']
model = ica1(3)
# foreach data batch
for i in range(10):
new_A, new_S = model.fit(new_data[i])
# Reorder components
c_SS = corrcoef(new_S, gen.sources)[3:, :3]
source_sim = abs(c_SS).max(axis=1)
# Check sources are similar
self.assertTrue(all(source_sim > 0.95))
order = abs(c_SS).argmax(axis=1)
signs = array([sign(x[order[n]]) for n, x in enumerate(c_SS)])
new_S = new_S[order, :] * signs.reshape((-1, 1))
new_A = new_A[:, order] * signs.reshape((1, -1))
new_model = [histogram(column, density=True, bins=20)\
for column in new_A.T]
# Check that resulting new mixing has similar histogram
for j in range(len(new_model)):
sim = abs(corrcoef(new_model[j][0], old_model[j][0]))
self.assertTrue(sim[0, 1] > 0.8,
'simmilarity {} is too low'.format(sim))
# Import ica function from ica import ica1 import numpy as np import matplotlib.pyplot as plt # Define matrix dimensions Nobs = 100 # Number of observations Nvars = 10000 # Number of variables Ncomp = 2 # Number of components # Simulated true sources S_true = np.random.logistic(0, 1, (Ncomp, Nvars)) # Simulated true mixing A_true = np.random.normal(0, 1, (Nobs, Ncomp)) # X = AS X = np.dot(A_true, S_true) # add some noise X = X + np.random.normal(0, 5, X.shape) # apply ICA on X and ask for 2 components A, S = ica1(X, 2) # compare if our estimates are accurate # correlate A with Atrue and take aCorr = np.abs(np.corrcoef(A.T, A_true.T)[:Ncomp, Ncomp:]).max(axis=0).mean() sCorr = np.abs(np.corrcoef(S, S_true)[:Ncomp, Ncomp:]).max(axis=0).mean() print "Accuracy of estimated sources: %.2f" % sCorr print "Accuracy of estimated mixing: %.2f" % aCorr
def sliding_window_ica(data, win, step, comp): # runs Infomax ICA on chunked array
C = create_chunked_array(data, win, step)
output = ica1(np.asarray(C),comp)
return output[1] # returns sources array
def eawica(sample,
constants,
wavelet='db4',
low_k=5,
up_k=95,
low_r=5,
up_r=95,
alpha=6):
n_epochs = constants.SECONDS
n_channels = sample.shape[0]
n_samples = constants.WINDOW
fb = filter_bank_class(constants)
# COMPUTE WAVELET DECOMPOSED wcs_delta
wcs, wcs_beta, wcs_gamma = [], [], []
for i in range(n_channels):
GAMMA, BETA, ALPHA, THETA, DELTA = fb.eawica_wavelet_band_pass(
sample[i, :], wavelet)
pos = i * 3
wcs.append([GAMMA, pos])
wcs.append([BETA, pos + 1])
wcs.append([ALPHA, pos + 2])
wcs_beta.append(BETA)
wcs_gamma.append(GAMMA)
# CHECKING FIRST CONDITION OVER ALL wcs_delta
kurt_list = []
renyi_list = []
for i in range(len(wcs)):
# -- kurtosis --
k = kurtosis(wcs[i][0])
kurt_list.append(k)
# -- renyi entropy --
pdf = np.histogram(wcs[i][0], bins=10)[0] / wcs[i][0].shape[0]
r = entropy.renyientropy(pdf, alpha=alpha, logbase=2, measure='R')
renyi_list.append(r)
# -- scaling --
kurt_list_scaled = zscore(kurt_list)
renyi_list_scaled = zscore(renyi_list)
# -- threshold --
low_kurt_threshold, up_kurt_threshold = np.percentile(
kurt_list_scaled, low_k), np.percentile(kurt_list_scaled, up_k)
low_renyi_threshold, up_renyi_threshold = np.percentile(
renyi_list_scaled, low_r), np.percentile(kurt_list_scaled, up_r)
cond_11 = np.logical_or(kurt_list_scaled > up_kurt_threshold,
kurt_list_scaled < low_kurt_threshold)
cond_12 = np.logical_or(renyi_list_scaled > up_renyi_threshold,
renyi_list_scaled < low_renyi_threshold)
cond_1 = cond_11 + cond_12
# SELECT wcs_delta MARKED AS CONTAINING ARTIFACTUAL INFORMATION
signals2check = np.zeros((np.sum(cond_1), n_samples + 1))
indices = np.where(cond_1 == True)[0]
for indx in range(len(indices)):
if cond_1[indices[indx]]:
signals2check[indx, :-1] = wcs[indices[indx]][0]
signals2check[indx, -1] = wcs[indices[indx]][1]
# ICA INFOMAX DECOMPOSITION OF MARKED signals TO OBTAIN ICs
n_components = signals2check.shape[0]
A, S, W = ica1(signals2check[:, :-1], n_components)
# CHECK SECOND CONDITION OVER EACH EPOCH ON WICs
control_k = np.zeros((S.shape[0], (n_epochs)))
control_r = np.zeros((S.shape[0], (n_epochs)))
for indx1 in range(S.shape[0]):
for indx2 in range((n_epochs)):
ini = int((indx2 * S.shape[1] / n_epochs))
end = int(ini + S.shape[1] / n_epochs)
if end + 1 == S.shape[1]:
end += 1
epoch = S[indx1, ini:end]
control_k[indx1, indx2] = kurtosis(epoch)
pdf = np.histogram(epoch, bins=10)[0] / epoch.shape[0]
r = entropy.renyientropy(pdf, alpha=alpha, logbase=2, measure='R')
control_r[indx1, indx2] = r
table = np.zeros((S.shape[0], n_epochs))
for indx1 in range(n_epochs):
control_k[:, indx1] = preprocessing.scale(control_k[:, indx1])
control_r[:, indx1] = preprocessing.scale(control_r[:, indx1])
table = np.logical_or(control_k > up_kurt_threshold,
control_k < low_kurt_threshold) + np.logical_or(
control_r > up_renyi_threshold,
control_r < low_renyi_threshold)
# ZEROING THOSE EPOCHS IN WICs MARKED AS ARTIFACTUAL EPOCHS
for indx1 in range(S.shape[0]):
for indx2 in range(n_epochs):
if table[indx1, indx2]:
ini = indx2 * int(n_samples / n_epochs)
end = ini + int(n_samples / n_epochs)
# epochs zeroing
S[indx1, ini:end] = 0
# wcs_delta RECONSTRUCTION FROM WICs
reconstructed = A.dot(S)
for i in range(reconstructed.shape[0]):
wcs[int(signals2check[i, -1])][0] = reconstructed[i, :]
data_cleaned = np.zeros(sample.shape)
for i in range(n_channels):
pos = i * 3
data_cleaned[i, :] = wcs[pos][0] + wcs[pos + 1][0] + wcs[
pos + 2][0] + wcs_beta[i] + wcs_gamma[i]
return data_cleaned
import numpy as np from ica import ica1 import time NSUB = 1000 NCOMP = 100 NVOX = 50000 true_a = np.random.normal(0,1,NSUB*NCOMP).reshape((NSUB,NCOMP)) true_s = np.random.logistic(0,1,NCOMP*NVOX).reshape((NCOMP, NVOX)) true_x = np.dot(true_a, true_s) + np.random.normal(0,1, NSUB*NVOX).reshape((NSUB,NVOX)) import time start = time.time() A,S = ica1(true_x,NCOMP) end = time.time() print end - start
import numpy as np
from ica import ica1
import time
NSUB = 1000
NCOMP = 100
NVOX = 50000
true_a = np.random.normal(0, 1, NSUB * NCOMP).reshape((NSUB, NCOMP))
true_s = np.random.logistic(0, 1, NCOMP * NVOX).reshape((NCOMP, NVOX))
true_x = np.dot(true_a, true_s) + np.random.normal(0, 1, NSUB * NVOX).reshape(
(NSUB, NVOX))
import time
start = time.time()
A, S = ica1(true_x, NCOMP)
end = time.time()
print end - start
def eawica(sample, constants, wavelet='db4', low_k=5, up_k=95, low_r=5, up_r=95, alpha=6): #Recibe la muestra y las constantes
#DESDE Data_Manager. También se fijan
#los valores de los umbrales
n_channels = sample.shape[0] #Primer elemento del vector SHAPE = Filas = Canales = Sensores
n_epochs = constants.SECONDS
n_samples = constants.WINDOW
fb = filter_bank_class(constants) #Envía las constantes a Filter_Bank
# Descomposición de la onda en rangos de frecuencias usando filtro EAWICA en tres listas diferentes
wcs, wcs_beta, wcs_gamma = [], [], []
# Para cada canal se descompone la señal en los rangos de frecuencia y se agregan GAMMA, BETA y ALPHA
for i in range(n_channels):
GAMMA, BETA, ALPHA, THETA, DELTA = fb.eawica_wavelet_band_pass(sample[i, :], wavelet) #Aplica el filtro EAWICA paso banda definido en Filter_Bank
# La lista WCS tendrá la forma [[GAMMA1,0], [BETA1,1], [ALPHA1,2], [GAMMA2,3], [BETA2,4], [ALPHA2,5], ...]
# [[GAMMA1,0], [BETA1,1], [ALPHA1,2], [GAMMA2,3], [BETA2,4], [ALPHA2,5], ...]
# [[GAMMA1,0], [BETA1,1], [ALPHA1,2], [GAMMA2,3], [BETA2,4], [ALPHA2,5], ...]
pos = i*3
wcs.append([GAMMA, pos])
wcs.append([BETA, pos+1])
wcs.append([ALPHA, pos+2])
# Las listas WCS_Beta y WCS_Gamma tendrán la forma [BETA1, BETA2...] y [GAMMA1, GAMMA2...]
wcs_beta.append(BETA)
wcs_gamma.append(GAMMA)
# CHECKING FIRST CONDITION OVER ALL wcs_delta
kurt_list = []
renyi_list = []
for i in range(len(wcs)):
# Calcular la kurtosis: E(desviación/variaza)^4
k = kurtosis(wcs[i][0]) #Aplica la kurtosis al primer elemento de cada pareja de la lista WCS
kurt_list.append(k)
# Calcular la entropía Renyi
pdf = np.histogram(wcs[i][0], bins=10)[0]/wcs[i][0].shape[0] #Calcula el histograma (diagrama de 10 barras),
#coge el primer elemento (número de repeticiones)
#y la divide entre el número de filas de esa componente de WCS
r = entropy.renyientropy(pdf,alpha=alpha,logbase=2,measure='R')
renyi_list.append(r)
# Escalado: transformar los valores en una distribución normal
kurt_list_scaled = zscore(kurt_list)
renyi_list_scaled = zscore(renyi_list)
# Umbrales superior e inferior: calcula los percentiles = valor por debajo del cual se encuentra el porcentaje de los datos especificado
low_kurt_threshold, up_kurt_threshold = np.percentile(kurt_list_scaled, low_k), np.percentile(kurt_list_scaled, up_k)
low_renyi_threshold, up_renyi_threshold = np.percentile(renyi_list_scaled, low_r), np.percentile(kurt_list_scaled, up_r)
# Aplica un OR entre las matrices elemento a elemento para hallar los datos extremos = fuera de los percentiles
cond_11 = np.logical_or(kurt_list_scaled > up_kurt_threshold, kurt_list_scaled < low_kurt_threshold)
cond_12 = np.logical_or(renyi_list_scaled > up_renyi_threshold, renyi_list_scaled < low_renyi_threshold)
cond_1 = cond_11 + cond_12 #Vector con el valor TRUE/FALSE para cada elemento y para las dos condiciones
# SELECT wcs_delta MARKED AS CONTAINING ARTIFACTUAL INFORMATION
signals2check = np.zeros((np.sum(cond_1), n_samples+1)) #Crear una matriz de ceros: FILAS = número de TRUE en las condiciones, COLUMNAS = número de muestras
indices = np.where(cond_1 == True)[0] #Extraer los índices de los elementos cuya condición sea TRUE
for indx in range(len(indices)):
if cond_1[indices[indx]]:
signals2check[indx, :-1] = wcs[indices[indx]][0]
signals2check[indx, -1] = wcs[indices[indx]][1]
# ICA INFOMAX DECOMPOSITION OF MARKED signals TO OBTAIN ICs
n_components = signals2check.shape[0]
A,S,W = ica1(signals2check[:,:-1], n_components)
# CHECK SECOND CONDITION OVER EACH EPOCH ON WICs
control_k = np.zeros((S.shape[0], (n_epochs)))
control_r = np.zeros((S.shape[0], (n_epochs)))
for indx1 in range(S.shape[0]):
for indx2 in range((n_epochs)):
ini = int((indx2*S.shape[1]/n_epochs))
end = int(ini + S.shape[1]/n_epochs)
if end+1==S.shape[1]:
end+=1
epoch = S[indx1,ini:end]
control_k[indx1,indx2] = kurtosis(epoch)
pdf = np.histogram(epoch, bins=10)[0]/epoch.shape[0]
r = entropy.renyientropy(pdf,alpha=alpha,logbase=2,measure='R')
control_r[indx1,indx2] = r
table = np.zeros((S.shape[0], n_epochs))
for indx1 in range(n_epochs):
control_k[:,indx1] = preprocessing.scale(control_k[:,indx1])
control_r[:,indx1] = preprocessing.scale(control_r[:,indx1])
table = np.logical_or(control_k > up_kurt_threshold, control_k < low_kurt_threshold) + np.logical_or(control_r > up_renyi_threshold, control_r < low_renyi_threshold)
# ZEROING THOSE EPOCHS IN WICs MARKED AS ARTIFACTUAL EPOCHS
for indx1 in range(S.shape[0]):
for indx2 in range(n_epochs):
if table[indx1,indx2]:
ini = indx2*int(n_samples/n_epochs)
end = ini + int(n_samples/n_epochs)
# epochs zeroing
S[indx1,ini:end] = 0
# wcs_delta RECONSTRUCTION FROM WICs
reconstructed = A.dot(S)
for i in range(reconstructed.shape[0]):
wcs[int(signals2check[i,-1])][0] = reconstructed[i,:]
data_cleaned = np.zeros(sample.shape)
for i in range(n_channels):
pos = i*3
data_cleaned[i,:] = wcs[pos][0]+wcs[pos+1][0]+wcs[pos+2][0]+wcs_beta[i]+wcs_gamma[i]
return data_cleaned
