def linearize_freqs(self, delta_freq=None):
"""Rebin frequencies so that the frequency axis is linear.
Parameters
----------
delta_freq : float
Difference between consecutive values on the new frequency axis.
Defaults to half of smallest delta in current frequency axis.
Compare Nyquist-Shannon sampling theorem.
"""
if delta_freq is None:
# Nyquist–Shannon sampling theorem
delta_freq = _min_delt(self.freq_axis) / 2.
nsize = (self.freq_axis.max() - self.freq_axis.min()) / delta_freq + 1
new = np.zeros((nsize, self.shape[1]), dtype=self.data.dtype)
freqs = self.freq_axis - self.freq_axis.max()
freqs = freqs / delta_freq
midpoints = np.round((freqs[:-1] + freqs[1:]) / 2)
fillto = np.concatenate(
[midpoints - 1, np.round([freqs[-1]]) - 1]
)
fillfrom = np.concatenate(
[np.round([freqs[0]]), midpoints - 1]
)
fillto = np.abs(fillto)
fillfrom = np.abs(fillfrom)
for row, from_, to_ in izip(self, fillfrom, fillto):
new[from_: to_] = row
vrs = self._get_params()
vrs.update({
'freq_axis': np.linspace(
self.freq_axis.max(), self.freq_axis.min(), nsize
)
})
return self.__class__(new, **vrs)
def interpolate(self, frequency):
"""
Linearly interpolate intensity at unknown frequency using linear
interpolation of its two neighbours.
Parameters
----------
frequency : float or int
Unknown frequency for which to linearly interpolate the intensities.
freq_axis[0] >= frequency >= self_freq_axis[-1]
"""
lfreq, lvalue = None, None
for freq, value in izip(self.freq_axis, self.data[:, :]):
if freq < frequency:
break
lfreq, lvalue = freq, value
else:
raise ValueError("Frequency not in interpolation range")
if lfreq is None:
raise ValueError("Frequency not in interpolation range")
diff = frequency - freq # pylint: disable=W0631
ldiff = lfreq - frequency
return (ldiff * value + diff * lvalue) / (diff + ldiff) # pylint: disable=W0631
def interpolate(self, frequency):
"""
Linearly interpolate intensity at unknown frequency using linear
interpolation of its two neighbours.
Parameters
----------
frequency : float or int
Unknown frequency for which to linearly interpolate the intensities.
freq_axis[0] >= frequency >= self_freq_axis[-1]
"""
lfreq, lvalue = None, None
for freq, value in izip(self.freq_axis, self.data[:, :]):
if freq < frequency:
break
lfreq, lvalue = freq, value
else:
raise ValueError("Frequency not in interpolation range")
if lfreq is None:
raise ValueError("Frequency not in interpolation range")
diff = frequency - freq # pylint: disable=W0631
ldiff = lfreq - frequency
return (ldiff * value + diff * lvalue) / (diff + ldiff) # pylint: disable=W0631
def linearize_freqs(self, delta_freq=None):
"""Rebin frequencies so that the frequency axis is linear.
Parameters
----------
delta_freq : float
Difference between consecutive values on the new frequency axis.
Defaults to half of smallest delta in current frequency axis.
Compare Nyquist-Shannon sampling theorem.
"""
if delta_freq is None:
# Nyquist–Shannon sampling theorem
delta_freq = _min_delt(self.freq_axis) / 2.
nsize = (self.freq_axis.max() - self.freq_axis.min()) / delta_freq + 1
new = np.zeros((nsize, self.shape[1]), dtype=self.data.dtype)
freqs = self.freq_axis - self.freq_axis.max()
freqs = freqs / delta_freq
midpoints = np.round((freqs[:-1] + freqs[1:]) / 2)
fillto = np.concatenate([midpoints - 1, np.round([freqs[-1]]) - 1])
fillfrom = np.concatenate([np.round([freqs[0]]), midpoints - 1])
fillto = np.abs(fillto)
fillfrom = np.abs(fillfrom)
for row, from_, to_ in izip(self, fillfrom, fillto):
new[from_:to_] = row
vrs = self._get_params()
vrs.update({
'freq_axis':
np.linspace(self.freq_axis.max(), self.freq_axis.min(), nsize)
})
return self.__class__(new, **vrs)
def join_many(cls,
specs,
mk_arr=None,
nonlinear=False,
maxgap=0,
fill=JOIN_REPEAT):
"""Produce new Spectrogram that contains spectrograms
joined together in time.
Parameters
----------
specs : list
List of spectrograms to join together in time.
nonlinear : bool
If True, leave out gaps between spectrograms. Else, fill them with
the value specified in fill.
maxgap : float, int or None
Largest gap to allow in second. If None, allow gap of arbitrary
size.
fill : float or int
Value to fill missing values (assuming nonlinear=False) with.
Can be LinearTimeSpectrogram.JOIN_REPEAT to repeat the values for
the time just before the gap.
mk_array: function
Function that is called to create the resulting array. Can be set
to LinearTimeSpectrogram.memap(filename) to create a memory mapped
result array.
"""
# XXX: Only load header and load contents of files
# on demand.
mask = None
if mk_arr is None:
mk_arr = cls.make_array
specs = sorted(specs, key=lambda x: x.start)
freqs = specs[0].freq_axis
if not all(np.array_equal(freqs, sp.freq_axis) for sp in specs):
raise ValueError("Frequency channels do not match.")
# Smallest time-delta becomes the common time-delta.
min_delt = min(sp.t_delt for sp in specs)
dtype_ = max(sp.dtype for sp in specs)
specs = [sp.resample_time(min_delt) for sp in specs]
size = sum(sp.shape[1] for sp in specs)
data = specs[0]
start_day = data.start
xs = []
last = data
for elem in specs[1:]:
e_init = (SECONDS_PER_DAY *
(get_day(elem.start) - get_day(start_day)).days +
elem.t_init)
x = int((e_init - last.t_init) / min_delt)
xs.append(x)
diff = last.shape[1] - x
if maxgap is not None and -diff > maxgap / min_delt:
raise ValueError("Too large gap.")
# If we leave out undefined values, we do not want to
# add values here if x > t_res.
if nonlinear:
size -= max(0, diff)
else:
size -= diff
last = elem
# The non existing element after the last one starts after
# the last one. Needed to keep implementation below sane.
xs.append(specs[-1].shape[1])
# We do that here so the user can pass a memory mapped
# array if they'd like to.
arr = mk_arr((data.shape[0], size), dtype_)
time_axis = np.zeros((size, ))
sx = 0
# Amount of pixels left out due to non-linearity. Needs to be
# considered for correct time axes.
sd = 0
for x, elem in izip(xs, specs):
diff = x - elem.shape[1]
e_time_axis = elem.time_axis
elem = elem.data
if x > elem.shape[1]:
if nonlinear:
x = elem.shape[1]
else:
# If we want to stay linear, fill up the missing
# pixels with placeholder zeros.
filler = np.zeros((data.shape[0], diff))
if fill is cls.JOIN_REPEAT:
filler[:, :] = elem[:, -1, np.newaxis]
else:
filler[:] = fill
minimum = e_time_axis[-1]
e_time_axis = np.concatenate([
e_time_axis,
np.linspace(minimum + min_delt,
minimum + diff * min_delt, diff)
])
elem = np.concatenate([elem, filler], 1)
arr[:, sx:sx + x] = elem[:, :x]
if diff > 0:
if mask is None:
mask = np.zeros((data.shape[0], size), dtype=np.uint8)
mask[:, sx + x - diff:sx + x] = 1
time_axis[sx:sx + x] = e_time_axis[:x] + data.t_delt * (sx + sd)
if nonlinear:
sd += max(0, diff)
sx += x
params = {
'time_axis': time_axis,
'freq_axis': data.freq_axis,
'start': data.start,
'end': specs[-1].end,
't_delt': data.t_delt,
't_init': data.t_init,
't_label': data.t_label,
'f_label': data.f_label,
'content': data.content,
'instruments': _union(spec.instruments for spec in specs),
}
if mask is not None:
arr = ma.array(arr, mask=mask)
if nonlinear:
del params['t_delt']
return Spectrogram(arr, **params)
return common_base(specs)(arr, **params)
def join_many(cls, specs, mk_arr=None, nonlinear=False,
maxgap=0, fill=JOIN_REPEAT):
"""Produce new Spectrogram that contains spectrograms
joined together in time.
Parameters
----------
specs : list
List of spectrograms to join together in time.
nonlinear : bool
If True, leave out gaps between spectrograms. Else, fill them with
the value specified in fill.
maxgap : float, int or None
Largest gap to allow in second. If None, allow gap of arbitrary
size.
fill : float or int
Value to fill missing values (assuming nonlinear=False) with.
Can be LinearTimeSpectrogram.JOIN_REPEAT to repeat the values for
the time just before the gap.
mk_array: function
Function that is called to create the resulting array. Can be set
to LinearTimeSpectrogram.memap(filename) to create a memory mapped
result array.
"""
# XXX: Only load header and load contents of files
# on demand.
mask = None
if mk_arr is None:
mk_arr = cls.make_array
specs = sorted(specs, key=lambda x: x.start)
freqs = specs[0].freq_axis
if not all(np.array_equal(freqs, sp.freq_axis) for sp in specs):
raise ValueError("Frequency channels do not match.")
# Smallest time-delta becomes the common time-delta.
min_delt = min(sp.t_delt for sp in specs)
dtype_ = max(sp.dtype for sp in specs)
specs = [sp.resample_time(min_delt) for sp in specs]
size = sum(sp.shape[1] for sp in specs)
data = specs[0]
start_day = data.start
xs = []
last = data
for elem in specs[1:]:
e_init = (
SECONDS_PER_DAY * (
get_day(elem.start) - get_day(start_day)
).days + elem.t_init
)
x = int((e_init - last.t_init) / min_delt)
xs.append(x)
diff = last.shape[1] - x
if maxgap is not None and -diff > maxgap / min_delt:
raise ValueError("Too large gap.")
# If we leave out undefined values, we do not want to
# add values here if x > t_res.
if nonlinear:
size -= max(0, diff)
else:
size -= diff
last = elem
# The non existing element after the last one starts after
# the last one. Needed to keep implementation below sane.
xs.append(specs[-1].shape[1])
# We do that here so the user can pass a memory mapped
# array if they'd like to.
arr = mk_arr((data.shape[0], size), dtype_)
time_axis = np.zeros((size,))
sx = 0
# Amount of pixels left out due to non-linearity. Needs to be
# considered for correct time axes.
sd = 0
for x, elem in izip(xs, specs):
diff = x - elem.shape[1]
e_time_axis = elem.time_axis
elem = elem.data
if x > elem.shape[1]:
if nonlinear:
x = elem.shape[1]
else:
# If we want to stay linear, fill up the missing
# pixels with placeholder zeros.
filler = np.zeros((data.shape[0], diff))
if fill is cls.JOIN_REPEAT:
filler[:, :] = elem[:, -1, np.newaxis]
else:
filler[:] = fill
minimum = e_time_axis[-1]
e_time_axis = np.concatenate([
e_time_axis,
np.linspace(
minimum + min_delt,
minimum + diff * min_delt,
diff
)
])
elem = np.concatenate([elem, filler], 1)
arr[:, sx:sx + x] = elem[:, :x]
if diff > 0:
if mask is None:
mask = np.zeros((data.shape[0], size), dtype=np.uint8)
mask[:, sx + x - diff:sx + x] = 1
time_axis[sx:sx + x] = e_time_axis[:x] + data.t_delt * (sx + sd)
if nonlinear:
sd += max(0, diff)
sx += x
params = {
'time_axis': time_axis,
'freq_axis': data.freq_axis,
'start': data.start,
'end': specs[-1].end,
't_delt': data.t_delt,
't_init': data.t_init,
't_label': data.t_label,
'f_label': data.f_label,
'content': data.content,
'instruments': _union(spec.instruments for spec in specs),
}
if mask is not None:
arr = ma.array(arr, mask=mask)
if nonlinear:
del params['t_delt']
return Spectrogram(arr, **params)
return common_base(specs)(arr, **params)
