def remove_artifacts(self):
"""Tries to remove the artifacts.
Uses mixture of mean-subtraction and OBS-PCA-Subtraction
Needs self.slcs to be set.
Saves example data in self.examples to make evaluation of filter quality possible"""
def obs_fit_error(p,y,x):
err = n.zeros((x.shape[0]),"d")
err += p[0]
for i in range(self.obs_size):
err += y[:,i]*p[i+1]
err -= x
return abs(err**2).sum()
def obs_fit_error_lsq(p,y,x):
err = n.zeros((x.shape[0]),"d")
err += p[0]
for i in range(self.obs_size):
err += y[:,i]*p[i+1]
err -= x
return err
def las_fit_error_lsq(p,y,x):
"""Local artifact subtraction"""
err = p[0]*y
err -= x
return err
#Shortnames
eeg = self.eeg
slcs = self.slcs_ups
sw = self._new_slice_width *10
k = self.num_neighbors
self.examples=[]
num_examples=10
#Loop over channels
if show_progressbar:
pbar = ProgressBar().start()
for i_ch in range(eeg.num_channels):
#Make Highpass-Version of channel
#ch = eeg[:,i_ch].copy()
y , fn = upsample_to_memmap(eeg[:,i_ch],10)
y_hp,fn_hp = tmp_memmap(dtype=y.dtype,shape=y.shape,mode="w+")
y_hp[:] = filtfilt_high(1, y, Fs=10000.0)
y_out,fn_out = tmp_memmap(dtype=y.dtype,shape=y.shape,mode="w+")
y_out[:] = y[:]
#ch_hp = filtfilt_high(1.0,ch,Fs=eeg.Fs)
neighbors = n.zeros((sw+2*self._pa_zeros,k))
#Prefill
for i in range(k):
#print neighbors[:,i].shape, eeg[slcs[i]-sw/2:slcs[i]+sw/2,i_ch].shape
neighbors[:,i] = prepend_append_zeros( y_hp[slcs[i]-sw/2:slcs[i]+sw/2] , self._pa_zeros)
#Loop over slices and filter
next_subst = 0
for i,t in enumerate(slcs):
try:
if i>k/2 and i<len(slcs)-k/2:
neighbors[:,next_subst] = prepend_append_zeros( y_hp[slcs[i+k/2]-sw/2:slcs[i+k/2]+sw/2] , self._pa_zeros)
next_subst+=1
next_subst=next_subst%k
tmp = prepend_append_zeros( y[t-sw/2:t+sw/2] , self._pa_zeros)
#Subtraction
#Shift/scale template
template = neighbors.mean(axis=1)
p = [1.0, 0]
p_las = leastsq(las_fit_error_lsq,p,args=(template,tmp))[0]
#print p_las[0],
tmp -= p_las[0]*template
#Beispieldaten speichern: Teil 1
if i_ch == self.ch_for_slice_alignment:
if i%(len(slcs)/num_examples)==(len(slcs)/num_examples)/2:
print "examples, Teil 1"
example = {}
example["raw"] = prepend_append_zeros(y[t-sw/2:t+sw/2].copy() , self._pa_zeros )
example["mean"] = p_las[0]*template
#OBS-Fit
components = pcafilt.unmix(neighbors) #OBS will be first 5 components
components -= components.mean(axis=0).reshape(1,-1).repeat(components.shape[0],0) #demeaning column-wise
obs = components[:,:self.obs_size].copy()
#Fit OBS to artifact
p = [0]+[0]*self.obs_size
#p_lsq = fmin(obs_fit_error,p,args=(obs,eeg[t-sw/2:t+sw/2,i_ch]),maxiter=1e5,maxfun=1e5)#[0]
p_lsq = leastsq(obs_fit_error_lsq,p,args=(obs,tmp))[0]
#print i,t,"p_lsq", p_lsq
fit = n.zeros((obs.shape[0]),"d")
fit +=p_lsq[0]
for j in range(self.obs_size):
fit += obs[:,j]*p_lsq[j+1]
tmp -= fit
try:
#eeg[t/10-sw/10/2:t/10+sw/10/2,i_ch] = tmp[self._pa_zeros:-self._pa_zeros][::10]
y_out[t-sw/2:t+sw/2] = tmp[self._pa_zeros:-self._pa_zeros][:]
except ValueError, ve:
print i_ch, i,t, eeg[t/10-sw/10/2:t/10+sw/10/2,i_ch].shape, tmp[self._pa_zeros:-self._pa_zeros][::10].shape
#Beispieldaten speichern: Teil 2
if i_ch == self.ch_for_slice_alignment:
if i%(len(slcs)/num_examples)==(len(slcs)/num_examples)/2:
print "examples, Teil 2"
#example["fit"] = n.zeros(example["raw"].shape) #fit.copy()
example["fit"] = fit.copy()
example["obs"] = obs.copy()
example["filt1"] = (tmp + fit).copy()
example["filt2"] = tmp.copy()
self.examples.append(example)
if show_progressbar:
pbar.update((i_ch+i/len(slcs))*100/eeg.num_channels)
except Exception, e:
print "Error occurred at slice at t=",t,", ignored"
print e
def remove_artifacts(self):
"""Tries to remove the artifacts.
Uses mixture of mean-subtraction and OBS-PCA-Subtraction
Needs self.slcs to be set.
Saves example data in self.examples to make evaluation of filter quality possible"""
def obs_fit_error(p, y, x):
err = n.zeros((x.shape[0]), "d")
err += p[0]
for i in range(self.obs_size):
err += y[:, i] * p[i + 1]
err -= x
return abs(err**2).sum()
def obs_fit_error_lsq(p, y, x):
err = n.zeros((x.shape[0]), "d")
err += p[0]
for i in range(self.obs_size):
err += y[:, i] * p[i + 1]
err -= x
return err
def las_fit_error_lsq(p, y, x):
"""Local artifact subtraction"""
err = p[0] * y
err -= x
return err
#Shortnames
eeg = self.eeg
slcs = self.slcs_ups
sw = self._new_slice_width * 10
k = self.num_neighbors
self.examples = []
num_examples = 10
#Loop over channels
if show_progressbar:
pbar = ProgressBar().start()
for i_ch in range(eeg.num_channels):
#Make Highpass-Version of channel
#ch = eeg[:,i_ch].copy()
y, fn = upsample_to_memmap(eeg[:, i_ch], 10)
y_hp, fn_hp = tmp_memmap(dtype=y.dtype, shape=y.shape, mode="w+")
y_hp[:] = filtfilt_high(1, y, Fs=10000.0)
y_out, fn_out = tmp_memmap(dtype=y.dtype, shape=y.shape, mode="w+")
y_out[:] = y[:]
#ch_hp = filtfilt_high(1.0,ch,Fs=eeg.Fs)
neighbors = n.zeros((sw + 2 * self._pa_zeros, k))
#Prefill
for i in range(k):
#print neighbors[:,i].shape, eeg[slcs[i]-sw/2:slcs[i]+sw/2,i_ch].shape
neighbors[:, i] = prepend_append_zeros(
y_hp[slcs[i] - sw / 2:slcs[i] + sw / 2], self._pa_zeros)
#Loop over slices and filter
next_subst = 0
for i, t in enumerate(slcs):
try:
if i > k / 2 and i < len(slcs) - k / 2:
neighbors[:, next_subst] = prepend_append_zeros(
y_hp[slcs[i + k / 2] - sw / 2:slcs[i + k / 2] +
sw / 2], self._pa_zeros)
next_subst += 1
next_subst = next_subst % k
tmp = prepend_append_zeros(y[t - sw / 2:t + sw / 2],
self._pa_zeros)
#Subtraction
#Shift/scale template
template = neighbors.mean(axis=1)
p = [1.0, 0]
p_las = leastsq(las_fit_error_lsq, p,
args=(template, tmp))[0]
#print p_las[0],
tmp -= p_las[0] * template
#Beispieldaten speichern: Teil 1
if i_ch == self.ch_for_slice_alignment:
if i % (len(slcs) / num_examples) == (
len(slcs) / num_examples) / 2:
print "examples, Teil 1"
example = {}
example["raw"] = prepend_append_zeros(
y[t - sw / 2:t + sw / 2].copy(),
self._pa_zeros)
example["mean"] = p_las[0] * template
#OBS-Fit
components = pcafilt.unmix(
neighbors) #OBS will be first 5 components
components -= components.mean(axis=0).reshape(
1, -1).repeat(components.shape[0],
0) #demeaning column-wise
obs = components[:, :self.obs_size].copy()
#Fit OBS to artifact
p = [0] + [0] * self.obs_size
#p_lsq = fmin(obs_fit_error,p,args=(obs,eeg[t-sw/2:t+sw/2,i_ch]),maxiter=1e5,maxfun=1e5)#[0]
p_lsq = leastsq(obs_fit_error_lsq, p, args=(obs, tmp))[0]
#print i,t,"p_lsq", p_lsq
fit = n.zeros((obs.shape[0]), "d")
fit += p_lsq[0]
for j in range(self.obs_size):
fit += obs[:, j] * p_lsq[j + 1]
tmp -= fit
try:
#eeg[t/10-sw/10/2:t/10+sw/10/2,i_ch] = tmp[self._pa_zeros:-self._pa_zeros][::10]
y_out[t - sw / 2:t +
sw / 2] = tmp[self._pa_zeros:-self._pa_zeros][:]
except ValueError, ve:
print i_ch, i, t, eeg[
t / 10 - sw / 10 / 2:t / 10 + sw / 10 / 2,
i_ch].shape, tmp[self._pa_zeros:-self.
_pa_zeros][::10].shape
#Beispieldaten speichern: Teil 2
if i_ch == self.ch_for_slice_alignment:
if i % (len(slcs) / num_examples) == (
len(slcs) / num_examples) / 2:
print "examples, Teil 2"
#example["fit"] = n.zeros(example["raw"].shape) #fit.copy()
example["fit"] = fit.copy()
example["obs"] = obs.copy()
example["filt1"] = (tmp + fit).copy()
example["filt2"] = tmp.copy()
self.examples.append(example)
if show_progressbar:
pbar.update(
(i_ch + i / len(slcs)) * 100 / eeg.num_channels)
except Exception, e:
print "Error occurred at slice at t=", t, ", ignored"
print e
def find_all_slice_artifacts(self):
def update_pbar(num):
"""Callback for find_all_maxs"""
if show_progressbar:
pbar.update(num/2)
eeg = self.eeg
#y = abs(smooth_windowed_eeg(eeg,[self.cfsa],self._slice_width))[:,0]
#y = smooth_windowed_eeg_power(eeg,[self.cfsa],self._slice_width)[:,0]
y = filtfilt_band(1,eeg.Fs/self._slice_width,eeg[:,self.cfsa])
#pylab.plot(y[::10])
#pylab.plot(eeg[:,14])
print y.shape, self._slice_width
#import pylab
#pylab.ion()
#pylab.plot(y[0:20000:1])
#pylab.show()
#raw_input()
#pylab.plot(y[13000:20000:1])
slcs_raw = find_all_maxs(y[:1000],ratio=0.6) # First segment
slcs_raw.sort()
#print "t=", t
#slcs_raw.append(t)
offset=0
t=int(0.5*self._slice_width)
while (t>0.4*self._slice_width or (y.shape[0]-offset)>10000) and (y.shape[0]-offset)>self._slice_width*2:
#print (y.shape[0]-offset)
#print t, offset, "-",
offset = slcs_raw[-1]+self._slice_width/2
#print t, offset, "-",
#pylab.plot(y[offset:offset+self._slice_width])
#pylab.show()
#raw_input()
t=y[offset:offset+self._slice_width].argmax()
slcs_raw.append(offset+t)
#print slcs_raw[-1], slcs_raw[-1]-slcs_raw[-2], " - ",
#time.sleep(0.1)
#print t, offset
print ""
#pylab.plot(y[::10])
if show_progressbar:
pbar = ProgressBar(maxval=eeg.shape[0]/self._slice_width).start()
#slcs_raw = find_all_maxs(y[:,0],0.3,self._slice_width,20,callback=update_pbar)
print "Raw slice-positions found", len(slcs_raw), np.mean(np.diff(slcs_raw)), np.min(slcs_raw), np.max(slcs_raw)
slcs_raw_diff = np.diff(slcs_raw)
print "slcs_raw_diff: ", scoreatpercentile(slcs_raw_diff,5), scoreatpercentile(slcs_raw_diff,50), scoreatpercentile(slcs_raw_diff,95)
#raise Exception("Abbruch")
y , fn = upsample_to_memmap(eeg[:,self.cfsa],10)
slcs_raw_ups = [x*10 for x in slcs_raw]
t = slcs_raw_ups[len(slcs_raw)/2]
template = y[t-self._slice_width*10/2:t+self._slice_width*10/2]
for i in range(5):
t = slcs_raw_ups[len(slcs_raw)/2+i]
template += y[t-self._slice_width*10/2:t+self._slice_width*10/2]
template /= 6
offsets = []
for i,t in enumerate(slcs_raw_ups):
#offset = find_max_overlap(template, eeg[t-self._slice_width/2:t+self._slice_width/2,self.cfsa], 100)
offset = find_max_overlap(template, y[t-self._slice_width*10/2:t+self._slice_width*10/2], 100)
offsets.append(offset)
self.slcs_ups = [slcs_raw_ups[i]+offsets[i]+self.slice_shift for i in range(len(slcs_raw_ups))]
if show_progressbar:
pbar.finish()
print "Refined slice-positions found. Finished.", len(offsets), np.mean(offsets), np.median(offsets), np.min(offsets), np.max(offsets)
print "Percentile 0.5,5,95,99.5 of offsets: ", scoreatpercentile(offsets,0.5), scoreatpercentile(offsets,5), scoreatpercentile(offsets,95), scoreatpercentile(offsets,99.5)
#Adjusting _slice_width...
print "Old slice_width:", self._slice_width
self._new_slice_width = int(n.ceil(n.mean(n.diff(self.slcs_ups))))/10
self._new_slice_width += 3 # Make slice wider to have no zombie-timepoints
self._new_slice_width = self._new_slice_width+self._new_slice_width%2
#self._new_slice_width = (self._new_slice_width/2)*2 # make sw%2==0 (divisible by 2)
print "New slice_width:", self._new_slice_width
#raise Exception("Abbruch")
return [x/10 for x in self.slcs_ups]
def find_all_slice_artifacts(self):
def update_pbar(num):
"""Callback for find_all_maxs"""
if show_progressbar:
pbar.update(num / 2)
eeg = self.eeg
#y = abs(smooth_windowed_eeg(eeg,[self.cfsa],self._slice_width))[:,0]
#y = smooth_windowed_eeg_power(eeg,[self.cfsa],self._slice_width)[:,0]
y = filtfilt_band(1, eeg.Fs / self._slice_width, eeg[:, self.cfsa])
#pylab.plot(y[::10])
#pylab.plot(eeg[:,14])
print y.shape, self._slice_width
#import pylab
#pylab.ion()
#pylab.plot(y[0:20000:1])
#pylab.show()
#raw_input()
#pylab.plot(y[13000:20000:1])
slcs_raw = find_all_maxs(y[:1000], ratio=0.6) # First segment
slcs_raw.sort()
#print "t=", t
#slcs_raw.append(t)
offset = 0
t = int(0.5 * self._slice_width)
while (t > 0.4 * self._slice_width or (y.shape[0] - offset) > 10000
) and (y.shape[0] - offset) > self._slice_width * 2:
#print (y.shape[0]-offset)
#print t, offset, "-",
offset = slcs_raw[-1] + self._slice_width / 2
#print t, offset, "-",
#pylab.plot(y[offset:offset+self._slice_width])
#pylab.show()
#raw_input()
t = y[offset:offset + self._slice_width].argmax()
slcs_raw.append(offset + t)
#print slcs_raw[-1], slcs_raw[-1]-slcs_raw[-2], " - ",
#time.sleep(0.1)
#print t, offset
print ""
#pylab.plot(y[::10])
if show_progressbar:
pbar = ProgressBar(maxval=eeg.shape[0] / self._slice_width).start()
#slcs_raw = find_all_maxs(y[:,0],0.3,self._slice_width,20,callback=update_pbar)
print "Raw slice-positions found", len(slcs_raw), np.mean(
np.diff(slcs_raw)), np.min(slcs_raw), np.max(slcs_raw)
slcs_raw_diff = np.diff(slcs_raw)
print "slcs_raw_diff: ", scoreatpercentile(
slcs_raw_diff,
5), scoreatpercentile(slcs_raw_diff,
50), scoreatpercentile(slcs_raw_diff, 95)
#raise Exception("Abbruch")
y, fn = upsample_to_memmap(eeg[:, self.cfsa], 10)
slcs_raw_ups = [x * 10 for x in slcs_raw]
t = slcs_raw_ups[len(slcs_raw) / 2]
template = y[t - self._slice_width * 10 / 2:t +
self._slice_width * 10 / 2]
for i in range(5):
t = slcs_raw_ups[len(slcs_raw) / 2 + i]
template += y[t - self._slice_width * 10 / 2:t +
self._slice_width * 10 / 2]
template /= 6
offsets = []
for i, t in enumerate(slcs_raw_ups):
#offset = find_max_overlap(template, eeg[t-self._slice_width/2:t+self._slice_width/2,self.cfsa], 100)
offset = find_max_overlap(
template, y[t - self._slice_width * 10 / 2:t +
self._slice_width * 10 / 2], 100)
offsets.append(offset)
self.slcs_ups = [
slcs_raw_ups[i] + offsets[i] + self.slice_shift
for i in range(len(slcs_raw_ups))
]
if show_progressbar:
pbar.finish()
print "Refined slice-positions found. Finished.", len(
offsets), np.mean(offsets), np.median(offsets), np.min(
offsets), np.max(offsets)
print "Percentile 0.5,5,95,99.5 of offsets: ", scoreatpercentile(
offsets, 0.5), scoreatpercentile(offsets, 5), scoreatpercentile(
offsets, 95), scoreatpercentile(offsets, 99.5)
#Adjusting _slice_width...
print "Old slice_width:", self._slice_width
self._new_slice_width = int(n.ceil(n.mean(n.diff(self.slcs_ups)))) / 10
self._new_slice_width += 3 # Make slice wider to have no zombie-timepoints
self._new_slice_width = self._new_slice_width + self._new_slice_width % 2
#self._new_slice_width = (self._new_slice_width/2)*2 # make sw%2==0 (divisible by 2)
print "New slice_width:", self._new_slice_width
#raise Exception("Abbruch")
return [x / 10 for x in self.slcs_ups]
