def hilbert_pow(dat_ts, bands=None, pad_to_pow2=False, verbose=True):
"""
"""
# set default freq bands
if bands is None:
bands = freq_bands
# proc padding
taxis = dat_ts.get_axis(dat_ts.tdim)
npts_orig = dat_ts.shape[taxis]
if pad_to_pow2:
npts = 2**next_pow2(npts_orig)
else:
npts = npts_orig
# calc the hilbert power
if verbose:
sys.stdout.write('Hilbert Bands: ')
sys.stdout.flush()
pow = None
for band in bands:
if verbose:
sys.stdout.write('%s ' % band[0])
sys.stdout.flush()
p = TimeSeries(np.abs(
hilbert(dat_ts.filtered(band[1], filt_type='pass'),
N=npts,
axis=taxis).take(np.arange(npts_orig), axis=taxis)),
tdim=dat_ts.tdim,
samplerate=dat_ts.samplerate,
dims=dat_ts.dims.copy()).add_dim(Dim([band[0]],
'freqs'))
if pow is None:
pow = p
else:
pow = pow.extend(p, 'freqs')
if verbose:
sys.stdout.write('\n')
sys.stdout.flush()
return pow
def hilbert_pow(dat_ts, bands=None, pad_to_pow2=False, verbose=True):
"""
"""
# set default freq bands
if bands is None:
bands = freq_bands
# proc padding
taxis = dat_ts.get_axis(dat_ts.tdim)
npts_orig = dat_ts.shape[taxis]
if pad_to_pow2:
npts = 2**next_pow2(npts_orig)
else:
npts = npts_orig
# calc the hilbert power
if verbose:
sys.stdout.write('Hilbert Bands: ')
sys.stdout.flush()
pow = None
for band in bands:
if verbose:
sys.stdout.write('%s '%band[0])
sys.stdout.flush()
p = TimeSeries(np.abs(hilbert(dat_ts.filtered(band[1],
filt_type='pass'),
N=npts, axis=taxis).take(np.arange(npts_orig),
axis=taxis)),
tdim=dat_ts.tdim, samplerate=dat_ts.samplerate,
dims=dat_ts.dims.copy()).add_dim(Dim([band[0]],'freqs'))
if pow is None:
pow = p
else:
pow = pow.extend(p, 'freqs')
if verbose:
sys.stdout.write('\n')
sys.stdout.flush()
return pow
def resampled(self, resampled_rate, window=None,
loop_axis=None, num_mp_procs=0, pad_to_pow2=False):
"""
Resample the data and reset all the time ranges.
Uses the resample function from scipy. This method seems to
be more accurate than the decimate method.
Parameters
----------
resampled_rate : {float}
New sample rate to resample to.
window : {None,str,float,tuple}, optional
See scipy.signal.resample for details
loop_axis: {None,str,int}, optional
Sometimes it might be faster to loop over an axis.
num_mp_procs: int, optional
Whether to try and use multiprocessing to loop over axis.
0 means no multiprocessing
>0 specifies num procs to use
None means yes, and use all possible procs
pad_to_pow2: bool, optional
Pad along the time dimension to the next power of 2 so
that the resampling is much faster (experimental).
Returns
-------
ts : {TimeSeries}
A TimeSeries instance with the resampled data.
See Also
--------
scipy.signal.resample
"""
# resample the data, getting new time range
time_range = self[self.tdim]
new_length = int(np.round(len(time_range)*
resampled_rate/self.samplerate))
if pad_to_pow2:
padded_length = 2**next_pow2(len(time_range))
padded_new_length = int(np.round(padded_length*resampled_rate/self.samplerate))
time_range = np.hstack([time_range,
(np.arange(1,padded_length-len(time_range)+1)*np.diff(time_range[-2:]))+time_range[-1]])
if loop_axis is None:
# just do standard method on all data at once
if pad_to_pow2:
newdat,new_time_range = resample(pad_to_next_pow2(np.asarray(self),axis=self.taxis),
padded_new_length, t=time_range,
axis=self.taxis, window=window)
else:
newdat,new_time_range = resample(np.asarray(self),
new_length, t=time_range,
axis=self.taxis, window=window)
else:
# loop over specified axis
# get the loop axis name and length
loop_dim = self.get_dim_name(loop_axis)
loop_dim_len = len(self[loop_dim])
# specify empty boolean index
ind = np.zeros(loop_dim_len,dtype=np.bool)
newdat = []
if has_mp and num_mp_procs != 0:
po = mp.Pool(num_mp_procs)
for i in range(loop_dim_len):
ind[i] = True
dat = self.select(**{loop_dim:ind})
taxis = dat.taxis
if has_mp and num_mp_procs != 0:
# start async proc
if pad_to_pow2:
dat = pad_to_next_pow2(np.asarray(dat), axis=dat.taxis)
newdat.append(po.apply_async(resample,
(np.asarray(dat), padded_new_length, time_range,
taxis, window)))
else:
newdat.append(po.apply_async(resample,
(np.asarray(dat), new_length, time_range,
taxis, window)))
else:
# just call on that dataset
sys.stdout.write('%d '%i)
sys.stdout.flush()
if pad_to_pow2:
dat = pad_to_next_pow2(np.asarray(dat), axis=dat.taxis)
ndat,new_time_range = resample(np.asarray(dat), padded_new_length, t=time_range,
axis=taxis, window=window)
else:
ndat,new_time_range = resample(np.asarray(dat), new_length, t=time_range,
axis=taxis, window=window)
newdat.append(ndat)
ind[i] = False
if has_mp and num_mp_procs != 0:
# aggregate mp results
po.close()
#po.join()
out = []
for i in range(len(newdat)):
sys.stdout.write('%d '%i)
sys.stdout.flush()
out.append(newdat[i].get())
#out = [newdat[i].get() for i in range(len(newdat))]
newdat = [out[i][0] for i in range(len(out))]
new_time_range = out[i][1]
# concatenate the new data
newdat = np.concatenate(newdat,axis=self.get_axis(loop_axis))
sys.stdout.write('\n')
sys.stdout.flush()
# remove pad if we padded it
if pad_to_pow2:
newdat = newdat.take(range(new_length),axis=self.taxis)
new_time_range = new_time_range[:new_length]
# set the time dimension
newdims = self.dims.copy()
attrs = self.dims[self.taxis]._attrs.copy()
for k in self.dims[self.taxis]._required_attrs.keys():
attrs.pop(k,None)
newdims[self.taxis] = Dim(new_time_range,
self.dims[self.taxis].name,
**attrs)
attrs = self._attrs.copy()
for k in self._required_attrs.keys():
attrs.pop(k,None)
return TimeSeries(newdat, self.tdim, resampled_rate,
dims=newdims, **attrs)
def resampled(self, resampled_rate, window=None,
loop_axis=None, num_mp_procs=0, pad_to_pow2=False):
"""
Resample the data and reset all the time ranges.
Uses the resample function from scipy. This method seems to
be more accurate than the decimate method.
Parameters
----------
resampled_rate : {float}
New sample rate to resample to.
window : {None,str,float,tuple}, optional
See scipy.signal.resample for details
loop_axis: {None,str,int}, optional
Sometimes it might be faster to loop over an axis.
num_mp_procs: int, optional
Whether to try and use multiprocessing to loop over axis.
0 means no multiprocessing
>0 specifies num procs to use
None means yes, and use all possible procs
pad_to_pow2: bool, optional
Pad along the time dimension to the next power of 2 so
that the resampling is much faster (experimental).
Returns
-------
ts : {TimeSeries}
A TimeSeries instance with the resampled data.
See Also
--------
scipy.signal.resample
"""
# resample the data, getting new time range
time_range = self[self.tdim]
new_length = int(np.round(len(time_range)*resampled_rate/self.samplerate))
if pad_to_pow2:
padded_length = 2**next_pow2(len(time_range))
padded_new_length = int(np.round(padded_length*resampled_rate/self.samplerate))
time_range = np.hstack([time_range,
(np.arange(1,padded_length-len(time_range)+1)*np.diff(time_range[-2:]))+time_range[-1]])
if loop_axis is None:
# just do standard method on all data at once
if pad_to_pow2:
newdat,new_time_range = resample(pad_to_next_pow2(np.asarray(self),axis=self.taxis),
padded_new_length, t=time_range,
axis=self.taxis, window=window)
else:
newdat,new_time_range = resample(np.asarray(self),
new_length, t=time_range,
axis=self.taxis, window=window)
else:
# loop over specified axis
# get the loop axis name and length
loop_dim = self.get_dim_name(loop_axis)
loop_dim_len = len(self[loop_dim])
# specify empty boolean index
ind = np.zeros(loop_dim_len,dtype=np.bool)
newdat = []
if has_mp and num_mp_procs != 0:
po = mp.Pool(num_mp_procs)
for i in range(loop_dim_len):
ind[i] = True
dat = self.select(**{loop_dim:ind})
taxis = dat.taxis
if has_mp and num_mp_procs != 0:
# start async proc
if pad_to_pow2:
dat = pad_to_next_pow2(np.asarray(dat), axis=dat.taxis)
newdat.append(po.apply_async(resample,
(np.asarray(dat), padded_new_length, time_range,
taxis, window)))
else:
newdat.append(po.apply_async(resample,
(np.asarray(dat), new_length, time_range,
taxis, window)))
else:
# just call on that dataset
sys.stdout.write('%d '%i)
sys.stdout.flush()
if pad_to_pow2:
dat = pad_to_next_pow2(np.asarray(dat), axis=dat.taxis)
ndat,new_time_range = resample(np.asarray(dat), padded_new_length, t=time_range,
axis=taxis, window=window)
else:
ndat,new_time_range = resample(np.asarray(dat), new_length, t=time_range,
axis=taxis, window=window)
newdat.append(ndat)
ind[i] = False
if has_mp and num_mp_procs != 0:
# aggregate mp results
po.close()
#po.join()
out = []
for i in range(len(newdat)):
sys.stdout.write('%d '%i)
sys.stdout.flush()
out.append(newdat[i].get())
#out = [newdat[i].get() for i in range(len(newdat))]
newdat = [out[i][0] for i in range(len(out))]
new_time_range = out[i][1]
# concatenate the new data
newdat = np.concatenate(newdat,axis=self.get_axis(loop_axis))
sys.stdout.write('\n')
sys.stdout.flush()
# remove pad if we padded it
if pad_to_pow2:
newdat = newdat.take(range(new_length),axis=self.taxis)
new_time_range = new_time_range[:new_length]
# set the time dimension
newdims = self.dims.copy()
attrs = self.dims[self.taxis]._attrs.copy()
for k in self.dims[self.taxis]._required_attrs.keys():
attrs.pop(k,None)
newdims[self.taxis] = Dim(new_time_range,
self.dims[self.taxis].name,
**attrs)
attrs = self._attrs.copy()
for k in self._required_attrs.keys():
attrs.pop(k,None)
return TimeSeries(newdat, self.tdim, resampled_rate,
dims=newdims, **attrs)
def fconv_multi(in1, in2, mode='full'):
"""
Convolve multiple 1-dimensional arrays using FFT.
Calls scipy.signal.fft on every row in in1 and in2, multiplies
every possible pairwise combination of the transformed rows, and
returns an inverse fft (by calling scipy.signal.ifft) of the
result. Therefore the output array has as many rows as the product
of the number of rows in in1 and in2 (the number of colums depend
on the mode).
Parameters
----------
in1 : {array_like}
First input array. Must be arranged such that each row is a
1-D array with data to convolve.
in2 : {array_like}
Second input array. Must be arranged such that each row is a
1-D array with data to convolve.
mode : {'full','valid','same'},optional
Specifies the size of the output. See the docstring for
scipy.signal.convolve() for details.
Returns
-------
Array with in1.shape[0]*in2.shape[0] rows with the convolution of
the 1-D signals in the rows of in1 and in2.
"""
# ensure proper number of dimensions
in1 = np.atleast_2d(in1)
in2 = np.atleast_2d(in2)
# get the number of signals and samples in each input
num1,s1 = in1.shape
num2,s2 = in2.shape
# see if we will be returning a complex result
complex_result = (np.issubdtype(in1.dtype, np.complex) or
np.issubdtype(in2.dtype, np.complex))
# determine the size based on the next power of 2
actual_size = s1+s2-1
size = np.power(2,next_pow2(actual_size))
# perform the fft of each row of in1 and in2:
#in1_fft = np.empty((num1,size),dtype=np.complex128)
in1_fft = np.empty((num1,size),dtype=np.complex)
for i in xrange(num1):
in1_fft[i] = fft(in1[i],size)
#in2_fft = np.empty((num2,size),dtype=np.complex128)
in2_fft = np.empty((num2,size),dtype=np.complex)
for i in xrange(num2):
in2_fft[i] = fft(in2[i],size)
# duplicate the signals and multiply before taking the inverse
in1_fft = in1_fft.repeat(num2,axis=0)
in1_fft *= np.vstack([in2_fft]*num1)
ret = ifft(in1_fft)
# ret = ifft(in1_fft.repeat(num2,axis=0) * \
# np.vstack([in2_fft]*num1))
# delete to save memory
del in1_fft, in2_fft
# strip of extra space if necessary
ret = ret[:,:actual_size]
# determine if complex, keeping only real if not
if not complex_result:
ret = ret.real
# now only keep the requested portion
if mode == "full":
return ret
elif mode == "same":
if s1 > s2:
osize = s1
else:
osize = s2
return centered(ret,(num1*num2,osize))
elif mode == "valid":
return centered(ret,(num1*num2,np.abs(s2-s1)+1))
def fconv_multi(in1, in2, mode='full'):
"""
Convolve multiple 1-dimensional arrays using FFT.
Calls scipy.signal.fft on every row in in1 and in2, multiplies
every possible pairwise combination of the transformed rows, and
returns an inverse fft (by calling scipy.signal.ifft) of the
result. Therefore the output array has as many rows as the product
of the number of rows in in1 and in2 (the number of colums depend
on the mode).
Parameters
----------
in1 : {array_like}
First input array. Must be arranged such that each row is a
1-D array with data to convolve.
in2 : {array_like}
Second input array. Must be arranged such that each row is a
1-D array with data to convolve.
mode : {'full','valid','same'},optional
Specifies the size of the output. See the docstring for
scipy.signal.convolve() for details.
Returns
-------
Array with in1.shape[0]*in2.shape[0] rows with the convolution of
the 1-D signals in the rows of in1 and in2.
"""
# ensure proper number of dimensions
in1 = np.atleast_2d(in1)
in2 = np.atleast_2d(in2)
# get the number of signals and samples in each input
num1,s1 = in1.shape
num2,s2 = in2.shape
# see if we will be returning a complex result
complex_result = (np.issubdtype(in1.dtype, np.complex) or
np.issubdtype(in2.dtype, np.complex))
# determine the size based on the next power of 2
actual_size = s1+s2-1
size = np.power(2,next_pow2(actual_size))
# perform the fft of each row of in1 and in2:
#in1_fft = np.empty((num1,size),dtype=np.complex128)
in1_fft = np.empty((num1,size),dtype=np.complex)
for i in xrange(num1):
in1_fft[i] = fft(in1[i],size)
#in2_fft = np.empty((num2,size),dtype=np.complex128)
in2_fft = np.empty((num2,size),dtype=np.complex)
for i in xrange(num2):
in2_fft[i] = fft(in2[i],size)
# duplicate the signals and multiply before taking the inverse
in1_fft = in1_fft.repeat(num2,axis=0)
in1_fft *= np.vstack([in2_fft]*num1)
ret = ifft(in1_fft)
# ret = ifft(in1_fft.repeat(num2,axis=0) * \
# np.vstack([in2_fft]*num1))
# delete to save memory
del in1_fft, in2_fft
# strip of extra space if necessary
ret = ret[:,:actual_size]
# determine if complex, keeping only real if not
if not complex_result:
ret = ret.real
# now only keep the requested portion
if mode == "full":
return ret
elif mode == "same":
if s1 > s2:
osize = s1
else:
osize = s2
return centered(ret,(num1*num2,osize))
elif mode == "valid":
return centered(ret,(num1*num2,np.abs(s2-s1)+1))
