def test_string(self):
freq = '440'
with pytest.raises(SPYTypeError):
scalar_parser(freq,
varname="freq",
ntype="int_like",
lims=[10, 1000])
def samplerate(self, sr):
if sr is None:
self._samplerate = None
return
try:
scalar_parser(sr, varname="samplerate", lims=[1, np.inf])
except Exception as exc:
raise exc
self._samplerate = sr
def test_integer_like(self):
freq = 440.0
scalar_parser(freq, varname="freq", ntype="int_like", lims=[10, 1000])
# not integer-like
freq = 440.5
with pytest.raises(SPYValueError):
scalar_parser(freq,
varname="freq",
ntype="int_like",
lims=[10, 1000])
def test_within_limits(self):
value = 440
scalar_parser(value,
varname="value",
ntype="int_like",
lims=[10, 1000])
freq = 2 # outside bounds
with pytest.raises(SPYValueError):
scalar_parser(freq,
varname="freq",
ntype="int_like",
lims=[10, 1000])
def validate_padding(pad_to_length, lenTrials):
"""
Simplified padding
"""
# supported padding options
not_valid = False
if not isinstance(pad_to_length, (Number, str, type(None))):
not_valid = True
elif isinstance(pad_to_length,
str) and pad_to_length not in availablePaddingOpt:
not_valid = True
if isinstance(pad_to_length,
bool): # bool is an int subclass, check for it separately...
not_valid = True
if not_valid:
lgl = "`None`, 'nextpow2' or an integer like number"
actual = f"{pad_to_length}"
raise SPYValueError(legal=lgl, varname="pad_to_length", actual=actual)
# here we check for equal lengths trials in case of no user specified absolute padding length
# we do a rough 'maxlen' padding, nextpow2 will be overruled in this case
if lenTrials.min() != lenTrials.max() and not isinstance(
pad_to_length, Number):
abs_pad = int(lenTrials.max())
msg = f"Unequal trial lengths present, automatic padding to {abs_pad} samples"
SPYWarning(msg)
# zero padding of ALL trials the same way
if isinstance(pad_to_length, Number):
scalar_parser(pad_to_length,
varname='pad_to_length',
ntype='int_like',
lims=[lenTrials.max(), np.inf])
abs_pad = pad_to_length
# or pad to optimal FFT lengths
# (not possible for unequal lengths trials)
elif pad_to_length == 'nextpow2':
# after padding
abs_pad = _nextpow2(int(lenTrials.min()))
# no padding, equal lengths trials
elif pad_to_length is None:
abs_pad = int(lenTrials.max())
# `abs_pad` is now the (soon to be padded) signal length in samples
return abs_pad
def _layout_subplot_panels(npanels,
nrow=None,
ncol=None,
ndefault=5,
maxpanels=50):
"""
Create space-optimal subplot grid given required number of panels
Parameters
----------
npanels : int
Number of required subplot panels in figure
nrow : int or None
Required number of panel rows. Note, if both `nrow` and `ncol` are not `None`,
then ``nrow * ncol >= npanels`` has to be satisfied, otherwise a
:class:`~syncopy.shared.errors.SPYValueError` is raised.
ncol : int or None
Required number of panel columns. Note, if both `nrow` and `ncol` are not `None`,
then ``nrow * ncol >= npanels`` has to be satisfied, otherwise a
:class:`~syncopy.shared.errors.SPYValueError` is raised.
ndefault: int
Default number of panel columns for grid construction (only relevant if
both `nrow` and `ncol` are `None`).
maxpanels : int
Maximally allowed number of subplot panels for which a grid is constructed
Returns
-------
nrow : int
Number of rows of constructed subplot panel grid
nrow : int
Number of columns of constructed subplot panel grid
Notes
-----
If both `nrow` and `ncol` are `None`, the constructed grid will have the
dimension `N` x `ndefault`, where `N` is chosen "optimally", i.e., the smallest
integer that satisfies ``ndefault * N >= npanels``.
Note further, that this is an auxiliary method that is intended purely for
internal use. Thus, error-checking is only performed on potentially user-provided
inputs (`nrow` and `ncol`).
See also
--------
:func:`._setup_figure` : create and prepare figures for Syncopy visualizations
Examples
--------
Create grid of default dimensions to hold eight panels
>>> _layout_subplot_panels(8, ndefault=5)
(2, 5)
Create a grid that must have 4 rows
>>> _layout_subplot_panels(8, nrow=4)
(4, 2)
Create a grid that must have 8 columns
>>> _layout_subplot_panels(8, ncol=8)
(1, 8)
"""
# Abort if requested panel count is less than one or exceeds provided maximum
try:
scalar_parser(npanels,
varname="npanels",
ntype="int_like",
lims=[1, np.inf])
except Exception as exc:
raise exc
if npanels > maxpanels:
lgl = "a maximum of {} panels in total".format(maxpanels)
raise SPYValueError(legal=lgl, actual=str(npanels), varname="npanels")
# Row specifcation was provided, cols may or may not
if nrow is not None:
try:
scalar_parser(nrow,
varname="nrow",
ntype="int_like",
lims=[1, np.inf])
except Exception as exc:
raise exc
if ncol is None:
ncol = np.ceil(npanels / nrow).astype(np.intp)
# Column specifcation was provided, rows may or may not
if ncol is not None:
try:
scalar_parser(ncol,
varname="ncol",
ntype="int_like",
lims=[1, np.inf])
except Exception as exc:
raise exc
if nrow is None:
nrow = np.ceil(npanels / ncol).astype(np.intp)
# After the preparations above, this condition is *only* satisfied if both
# `nrow` = `ncol` = `None` -> then use generic grid-layout
if nrow is None:
ncol = ndefault
nrow = np.ceil(npanels / ncol).astype(np.intp)
ncol = min(ncol, npanels)
# Complain appropriately if requested no. of panels does not fit inside grid
if nrow * ncol < npanels:
lgl = "row- and column-specification of grid to fit all panels"
act = "grid with {0} rows and {1} columns but {2} panels"
raise SPYValueError(legal=lgl,
actual=act.format(nrow, ncol, npanels),
varname="nrow/ncol")
# In case a grid was provided too big for the requested no. of panels (e.g.,
# 8 panels in an 4 x 3 grid -> would fit in 3 x 3), just warn, don't crash
if nrow * ncol - npanels >= ncol:
msg = "Grid dimension ({0} rows x {1} columns) larger than necessary " +\
"for {2} panels. "
SPYWarning(msg.format(nrow, ncol, npanels))
return nrow, ncol
def definetrial(obj,
trialdefinition=None,
pre=None,
post=None,
start=None,
trigger=None,
stop=None,
clip_edges=False):
"""(Re-)define trials of a Syncopy data object
Data can be structured into trials based on timestamps of a start, trigger
and end events::
start trigger stop
|---- pre ----|--------|---------|--- post----|
Parameters
----------
obj : Syncopy data object (:class:`BaseData`-like)
trialdefinition : :class:`EventData` object or Mx3 array
[start, stop, trigger_offset] sample indices for `M` trials
pre : float
offset time (s) before start event
post : float
offset time (s) after end event
start : int
event code (id) to be used for start of trial
stop : int
event code (id) to be used for end of trial
trigger :
event code (id) to be used center (t=0) of trial
clip_edges : bool
trim trials to actual data-boundaries.
Returns
-------
Syncopy data object (:class:`BaseData`-like))
Notes
-----
:func:`definetrial` supports the following argument combinations:
>>> # define M trials based on [start, end, offset] indices
>>> definetrial(obj, trialdefinition=[M x 3] array)
>>> # define trials based on event codes stored in <:class:`EventData` object>
>>> definetrial(obj, trialdefinition=<EventData object>,
pre=0, post=0, start=startCode, stop=stopCode,
trigger=triggerCode)
>>> # apply same trial definition as defined in <:class:`EventData` object>
>>> definetrial(<AnalogData object>,
trialdefinition=<EventData object w/sampleinfo/t0/trialinfo>)
>>> # define whole recording as single trial
>>> definetrial(obj, trialdefinition=None)
"""
# Start by vetting input object
try:
data_parser(obj, varname="obj")
except Exception as exc:
raise exc
if obj.data is None:
lgl = "non-empty Syncopy data object"
act = "empty Syncopy data object"
raise SPYValueError(legal=lgl, varname="obj", actual=act)
# Check array/object holding trial specifications
if trialdefinition is not None:
if trialdefinition.__class__.__name__ == "EventData":
try:
data_parser(trialdefinition,
varname="trialdefinition",
writable=None,
empty=False)
except Exception as exc:
raise exc
evt = True
else:
try:
array_parser(trialdefinition,
varname="trialdefinition",
dims=2)
except Exception as exc:
raise exc
if any([
"ContinuousData" in str(base)
for base in obj.__class__.__mro__
]):
scount = obj.data.shape[obj.dimord.index("time")]
else:
scount = np.inf
try:
array_parser(trialdefinition[:, :2],
varname="sampleinfo",
dims=(None, 2),
hasnan=False,
hasinf=False,
ntype="int_like",
lims=[0, scount])
except Exception as exc:
raise exc
trl = np.array(trialdefinition, dtype="float")
ref = obj
tgt = obj
evt = False
else:
# Construct object-class-specific `trl` arrays treating data-set as single trial
if any(
["ContinuousData" in str(base) for base in obj.__class__.__mro__]):
trl = np.array([[0, obj.data.shape[obj.dimord.index("time")], 0]])
else:
sidx = obj.dimord.index("sample")
trl = np.array([[
np.nanmin(obj.data[:, sidx]),
np.nanmax(obj.data[:, sidx]), 0
]])
ref = obj
tgt = obj
evt = False
# AnalogData + EventData w/sampleinfo
if obj.__class__.__name__ == "AnalogData" and evt and trialdefinition.sampleinfo is not None:
if obj.samplerate is None or trialdefinition.samplerate is None:
lgl = "non-`None` value - make sure `samplerate` is set before defining trials"
act = "None"
raise SPYValueError(legal=lgl, varname="samplerate", actual=act)
ref = trialdefinition
tgt = obj
trl = np.array(ref.trialinfo)
t0 = np.array(ref._t0).reshape((ref._t0.size, 1))
trl = np.hstack([ref.sampleinfo, t0, trl])
trl = np.round((trl / ref.samplerate) * tgt.samplerate).astype(int)
# AnalogData + EventData w/keywords or just EventData w/keywords
if any([kw is not None for kw in [pre, post, start, trigger, stop]]):
# Make sure we actually have valid data objects to work with
if obj.__class__.__name__ == "EventData" and evt is False:
ref = obj
tgt = obj
elif obj.__class__.__name__ == "AnalogData" and evt is True:
ref = trialdefinition
tgt = obj
else:
lgl = "AnalogData with associated EventData object"
act = "{} and {}".format(obj.__class__.__name__,
trialdefinition.__class__.__name__)
raise SPYValueError(legal=lgl, actual=act, varname="input")
# The only case we might actually need it: ensure `clip_edges` is valid
if not isinstance(clip_edges, bool):
raise SPYTypeError(clip_edges,
varname="clip_edges",
expected="Boolean")
# Ensure that objects have their sampling-rates set, otherwise break
if ref.samplerate is None or tgt.samplerate is None:
lgl = "non-`None` value - make sure `samplerate` is set before defining trials"
act = "None"
raise SPYValueError(legal=lgl, varname="samplerate", actual=act)
# Get input dimensions
szin = []
for var in [pre, post, start, trigger, stop]:
if isinstance(var, (np.ndarray, list)):
szin.append(len(var))
if np.unique(szin).size > 1:
lgl = "all trial-related arrays to have the same length"
act = "arrays with sizes {}".format(
str(np.unique(szin)).replace("[", "").replace("]", ""))
raise SPYValueError(legal=lgl,
varname="trial-keywords",
actual=act)
if len(szin):
ntrials = szin[0]
ninc = 1
else:
ntrials = 1
ninc = 0
# If both `pre` and `start` or `post` and `stop` are `None`, abort
if (pre is None and start is None) or (post is None and stop is None):
lgl = "`pre` or `start` and `post` or `stop` to be not `None`"
act = "both `pre` and `start` and/or `post` and `stop` are simultaneously `None`"
raise SPYValueError(legal=lgl, actual=act)
if (trigger is None) and (pre is not None or post is not None):
lgl = "non-None `trigger` with `pre`/`post` timing information"
act = "`trigger` = `None`"
raise SPYValueError(legal=lgl, actual=act)
# If provided, ensure keywords make sense, otherwise allocate defaults
kwrds = {}
vdict = {
"pre": {
"var": pre,
"hasnan": False,
"ntype": None,
"fillvalue": 0
},
"post": {
"var": post,
"hasnan": False,
"ntype": None,
"fillvalue": 0
},
"start": {
"var": start,
"hasnan": None,
"ntype": "int_like",
"fillvalue": np.nan
},
"trigger": {
"var": trigger,
"hasnan": None,
"ntype": "int_like",
"fillvalue": np.nan
},
"stop": {
"var": stop,
"hasnan": None,
"ntype": "int_like",
"fillvalue": np.nan
}
}
for vname, opts in vdict.items():
if opts["var"] is not None:
if isinstance(opts["var"], numbers.Number):
try:
scalar_parser(opts["var"],
varname=vname,
ntype=opts["ntype"],
lims=[-np.inf, np.inf])
except Exception as exc:
raise exc
opts["var"] = np.full((ntrials, ), opts["var"])
else:
try:
array_parser(opts["var"],
varname=vname,
hasinf=False,
hasnan=opts["hasnan"],
ntype=opts["ntype"],
dims=(ntrials, ))
except Exception as exc:
raise exc
kwrds[vname] = opts["var"]
else:
kwrds[vname] = np.full((ntrials, ), opts["fillvalue"])
# Prepare `trl` and convert event-codes + sample-numbers to lists
trl = []
evtid = list(ref.data[:, ref.dimord.index("eventid")])
evtsp = list(ref.data[:, ref.dimord.index("sample")])
nevents = len(evtid)
searching = True
trialno = 0
cnt = 0
act = ""
# Do this line-by-line: halt on error (if event-id is not found in `ref`)
while searching:
# Allocate begin and end of trial
begin = None
end = None
t0 = 0
idxl = []
# First, try to assign `start`, then `t0`
if not np.isnan(kwrds["start"][trialno]):
try:
sidx = evtid.index(kwrds["start"][trialno])
except:
act = str(kwrds["start"][trialno])
vname = "start"
break
begin = evtsp[sidx] / ref.samplerate
evtid[sidx] = -np.pi
idxl.append(sidx)
if not np.isnan(kwrds["trigger"][trialno]):
try:
idx = evtid.index(kwrds["trigger"][trialno])
except:
act = str(kwrds["trigger"][trialno])
vname = "trigger"
break
t0 = evtsp[idx] / ref.samplerate
evtid[idx] = -np.pi
idxl.append(idx)
# Trial-begin is either `trigger - pre` or `start - pre`
if begin is not None:
begin -= kwrds["pre"][trialno]
else:
begin = t0 - kwrds["pre"][trialno]
# Try to assign `stop`, if we got nothing, use `t0 + post`
if not np.isnan(kwrds["stop"][trialno]):
evtid[:sidx] = [np.pi] * sidx
try:
idx = evtid.index(kwrds["stop"][trialno])
except:
act = str(kwrds["stop"][trialno])
vname = "stop"
break
end = evtsp[idx] / ref.samplerate + kwrds["post"][trialno]
evtid[idx] = -np.pi
idxl.append(idx)
else:
end = t0 + kwrds["post"][trialno]
# Off-set `t0`
t0 -= begin
# Make sure current trial setup makes (some) sense
if begin >= end:
lgl = "non-overlapping trial begin-/end-samples"
act = "trial-begin at {}, trial-end at {}".format(
str(begin), str(end))
raise SPYValueError(legal=lgl, actual=act)
# Finally, write line of `trl`
trl.append([begin, end, t0])
# Update counters and end this mess when we're done
trialno += ninc
cnt += 1
evtsp = evtsp[max(idxl, default=-1) + 1:]
evtid = evtid[max(idxl, default=-1) + 1:]
if trialno == ntrials or cnt == nevents:
searching = False
# Abort if the above loop ran into troubles
if len(trl) < ntrials:
if len(act) > 0:
raise SPYValueError(legal="existing event-id",
varname=vname,
actual=act)
# Make `trl` a NumPy array
trl = np.round(np.array(trl) * tgt.samplerate).astype(int)
# If appropriate, clip `trl` to AnalogData object's bounds (if wanted)
if clip_edges and evt:
msk = trl[:, 0] < 0
trl[msk, 0] = 0
dmax = tgt.data.shape[tgt.dimord.index("time")]
msk = trl[:, 1] > dmax
trl[msk, 1] = dmax
if np.any(trl[:, 0] >= trl[:, 1]):
lgl = "non-overlapping trials"
act = "some trials are overlapping after clipping to AnalogData object range"
raise SPYValueError(legal=lgl, actual=act)
# The triplet `sampleinfo`, `t0` and `trialinfo` works identically for
# all data genres
if trl.shape[1] < 3:
raise SPYValueError(
"array of shape (no. of trials, 3+)",
varname="trialdefinition",
actual="shape = {shp:s}".format(shp=str(trl.shape)))
# Finally: assign `sampleinfo`, `t0` and `trialinfo` (and potentially `trialid`)
tgt._trialdefinition = trl
# In the discrete case, we have some additinal work to do
if any(["DiscreteData" in str(base) for base in tgt.__class__.__mro__]):
# Compute trial-IDs by matching data samples with provided trial-bounds
samples = tgt.data[:, tgt.dimord.index("sample")]
starts = tgt.sampleinfo[:, 0]
ends = tgt.sampleinfo[:, 1]
startids = np.searchsorted(starts, samples, side="right")
endids = np.searchsorted(ends, samples, side="left")
mask = startids == endids
startids -= 1
# Samples not belonging into any trial get a trial-ID of -1
startids[mask] = int(startids.min() <= 0) * (-1)
tgt.trialid = startids
# Write log entry
if ref == tgt:
ref.log = "updated trial-definition with [" \
+ " x ".join([str(numel) for numel in trl.shape]) \
+ "] element array"
else:
ref_log = ref._log.replace("\n\n", "\n\t")
tgt.log = "trial-definition extracted from EventData object: "
tgt._log += ref_log
tgt.cfg = {
"method": sys._getframe().f_code.co_name,
"EventData object": ref.cfg
}
ref.log = "updated trial-defnition of {} object".format(
tgt.__class__.__name__)
return
def validate_taper(taper, tapsmofrq, nTaper, keeptapers, foimax, samplerate,
nSamples, output):
"""
General taper validation and Slepian/dpss input sanitization.
The default is to max out `nTaper` to achieve the desired frequency
smoothing bandwidth. For details about the Slepion settings see
"The Effective Bandwidth of a Multitaper Spectral Estimator,
A. T. Walden, E. J. McCoy and D. B. Percival"
Parameters
----------
taper : str
Windowing function, one of :data:`~syncopy.shared.const_def.availableTapers`
tapsmofrq : float or None
Taper smoothing bandwidth for `taper='dpss'`
nTaper : int_like or None
Number of tapers to use for multi-tapering (not recommended)
Other Parameters
----------------
keeptapers : bool
foimax : float
Maximum frequency for the analysis
samplerate : float
the samplerate in Hz
nSamples : int
Number of samples
output : str, one of {'abs', 'pow', 'fourier'}
Fourier transformation output type
Returns
-------
dpss_opt : dict
For multi-tapering (`taper='dpss'`) contains the
parameters `NW` and `Kmax` for `scipy.signal.windows.dpss`.
For all other tapers this is an empty dictionary.
"""
# See if taper choice is supported
if taper not in availableTapers:
lgl = "'" + "or '".join(opt + "' " for opt in availableTapers)
raise SPYValueError(legal=lgl, varname="taper", actual=taper)
# Warn user about DPSS only settings
if taper != "dpss":
if tapsmofrq is not None:
msg = "`tapsmofrq` is only used if `taper` is `dpss`!"
SPYWarning(msg)
if nTaper is not None:
msg = "`nTaper` is only used if `taper` is `dpss`!"
SPYWarning(msg)
if keeptapers:
msg = "`keeptapers` is only used if `taper` is `dpss`!"
SPYWarning(msg)
# empty dpss_opt, only Slepians have options
return {}
# direct mtm estimate (averaging) only valid for spectral power
if taper == "dpss" and not keeptapers and output != "pow":
lgl = "'pow', the only valid option for taper averaging"
raise SPYValueError(legal=lgl, varname="output", actual=output)
# Set/get `tapsmofrq` if we're working w/Slepian tapers
elif taper == "dpss":
# --- minimal smoothing bandwidth ---
# --- such that Kmax/nTaper is at least 1
minBw = 2 * samplerate / nSamples
# -----------------------------------
# user set tapsmofrq directly
if tapsmofrq is not None:
try:
scalar_parser(tapsmofrq, varname="tapsmofrq", lims=[0, np.inf])
except Exception as exc:
raise exc
if tapsmofrq < minBw:
msg = f'Setting tapsmofrq to the minimal attainable bandwidth of {minBw:.2f}Hz'
SPYInfo(msg)
tapsmofrq = minBw
# we now enforce a user submitted smoothing bw
else:
lgl = "smoothing bandwidth in Hz, typical values are in the range 1-10Hz"
raise SPYValueError(legal=lgl,
varname="tapsmofrq",
actual=tapsmofrq)
# Try to derive "sane" settings by using 3/4 octave
# smoothing of highest `foi`
# following Hill et al. "Oscillatory Synchronization in Large-Scale
# Cortical Networks Predicts Perception", Neuron, 2011
# FIX ME: This "sane setting" seems quite excessive (huuuge bwidths)
# tapsmofrq = (foimax * 2**(3 / 4 / 2) - foimax * 2**(-3 / 4 / 2)) / 2
# msg = f'Automatic setting of `tapsmofrq` to {tapsmofrq:.2f}'
# SPYInfo(msg)
# --------------------------------------------
# set parameters for scipy.signal.windows.dpss
NW = tapsmofrq * nSamples / (2 * samplerate)
# from the minBw setting NW always is at least 1
Kmax = int(2 * NW - 1) # optimal number of tapers
# --------------------------------------------
# the recommended way:
# set nTaper automatically to achieve exact effective smoothing bandwidth
if nTaper is None:
msg = f'Using {Kmax} taper(s) for multi-tapering'
SPYInfo(msg)
dpss_opt = {'NW': NW, 'Kmax': Kmax}
return dpss_opt
elif nTaper is not None:
try:
scalar_parser(nTaper,
varname="nTaper",
ntype="int_like",
lims=[1, np.inf])
except Exception as exc:
raise exc
if nTaper != Kmax:
msg = f'''
Manually setting the number of tapers is not recommended
and may (strongly) distort the effective smoothing bandwidth!\n
The optimal number of tapers is {Kmax}, you have chosen to use {nTaper}.
'''
SPYWarning(msg)
dpss_opt = {'NW': NW, 'Kmax': nTaper}
return dpss_opt
def freqanalysis(data, method='mtmfft', output='fourier',
keeptrials=True, foi=None, foilim=None,
pad_to_length=None, polyremoval=None,
taper="hann", tapsmofrq=None, nTaper=None, keeptapers=False,
toi="all", t_ftimwin=None, wavelet="Morlet", width=6, order=None,
order_max=None, order_min=1, c_1=3, adaptive=False,
out=None, **kwargs):
"""
Perform (time-)frequency analysis of Syncopy :class:`~syncopy.AnalogData` objects
**Usage Summary**
Options available in all analysis methods:
* **output** : one of :data:`~syncopy.specest.const_def.availableOutputs`;
return power spectra, complex Fourier spectra or absolute values.
* **foi**/**foilim** : frequencies of interest; either array of frequencies or
frequency window (not both)
* **keeptrials** : return individual trials or grand average
* **polyremoval** : de-trending method to use (0 = mean, 1 = linear or `None`)
List of available analysis methods and respective distinct options:
"mtmfft" : (Multi-)tapered Fourier transform
Perform frequency analysis on time-series trial data using either a single
taper window (Hanning) or many tapers based on the discrete prolate
spheroidal sequence (DPSS) that maximize energy concentration in the main
lobe.
* **taper** : one of :data:`~syncopy.shared.const_def.availableTapers`
* **tapsmofrq** : spectral smoothing box for slepian tapers (in Hz)
* **nTaper** : number of orthogonal tapers for slepian tapers
* **keeptapers** : return individual tapers or average
* **pad_to_length**: either pad to an absolute length or set to `'nextpow2'`
"mtmconvol" : (Multi-)tapered sliding window Fourier transform
Perform time-frequency analysis on time-series trial data based on a sliding
window short-time Fourier transform using either a single Hanning taper or
multiple DPSS tapers.
* **taper** : one of :data:`~syncopy.specest.const_def.availableTapers`
* **tapsmofrq** : spectral smoothing box for slepian tapers (in Hz)
* **nTaper** : number of orthogonal tapers for slepian tapers
* **keeptapers** : return individual tapers or average
* **toi** : time-points of interest; can be either an array representing
analysis window centroids (in sec), a scalar between 0 and 1 encoding
the percentage of overlap between adjacent windows or "all" to center
a window on every sample in the data.
* **t_ftimwin** : sliding window length (in sec)
"wavelet" : (Continuous non-orthogonal) wavelet transform
Perform time-frequency analysis on time-series trial data using a non-orthogonal
continuous wavelet transform.
* **wavelet** : one of :data:`~syncopy.specest.const_def.availableWavelets`
* **toi** : time-points of interest; can be either an array representing
time points (in sec) or "all"(pre-trimming and subsampling of results)
* **width** : Nondimensional frequency constant of Morlet wavelet function (>= 6)
* **order** : Order of Paul wavelet function (>= 4) or derivative order
of real-valued DOG wavelets (2 = mexican hat)
"superlet" : Superlet transform
Perform time-frequency analysis on time-series trial data using
the super-resolution superlet transform (SLT) from [Moca2021]_.
* **order_max** : Maximal order of the superlet
* **order_min** : Minimal order of the superlet
* **c_1** : Number of cycles of the base Morlet wavelet
* **adaptive** : If set to `True` perform fractional adaptive SLT,
otherwise perform multiplicative SLT
**Full documentation below**
Parameters
----------
data : `~syncopy.AnalogData`
A non-empty Syncopy :class:`~syncopy.datatype.AnalogData` object
method : str
Spectral estimation method, one of :data:`~syncopy.specest.const_def.availableMethods`
(see below).
output : str
Output of spectral estimation. One of :data:`~syncopy.specest.const_def.availableOutputs` (see below);
use `'pow'` for power spectrum (:obj:`numpy.float32`), `'fourier'` for complex
Fourier coefficients (:obj:`numpy.complex64`) or `'abs'` for absolute
values (:obj:`numpy.float32`).
keeptrials : bool
If `True` spectral estimates of individual trials are returned, otherwise
results are averaged across trials.
foi : array-like or None
Frequencies of interest (Hz) for output. If desired frequencies cannot be
matched exactly, the closest possible frequencies are used. If `foi` is `None`
or ``foi = "all"``, all attainable frequencies (i.e., zero to Nyquist / 2)
are selected.
foilim : array-like (floats [fmin, fmax]) or None or "all"
Frequency-window ``[fmin, fmax]`` (in Hz) of interest. Window
specifications must be sorted (e.g., ``[90, 70]`` is invalid) and not NaN
but may be unbounded (e.g., ``[-np.inf, 60.5]`` is valid). Edges `fmin`
and `fmax` are included in the selection. If `foilim` is `None` or
``foilim = "all"``, all frequencies are selected.
pad_to_length : int, None or 'nextpow2'
Padding of the input data, if set to a number pads all trials
to this absolute length. For instance ``pad_to_length = 2000`` pads all
trials to an absolute length of 2000 samples, if and only if the longest
trial contains at maximum 2000 samples.
Alternatively if all trials have the same initial lengths
setting `pad_to_length='nextpow2'` pads all trials to
the next power of two.
If `None` and trials have unequal lengths all trials are padded to match
the longest trial.
polyremoval : int or None
Order of polynomial used for de-trending data in the time domain prior
to spectral analysis. A value of 0 corresponds to subtracting the mean
("de-meaning"), ``polyremoval = 1`` removes linear trends (subtracting the
least squares fit of a linear polynomial).
If `polyremoval` is `None`, no de-trending is performed. Note that
for spectral estimation de-meaning is very advisable and hence also the
default.
taper : str
Only valid if `method` is `'mtmfft'` or `'mtmconvol'`. Windowing function,
one of :data:`~syncopy.specest.const_def.availableTapers` (see below).
tapsmofrq : float
Only valid if `method` is `'mtmfft'` or `'mtmconvol'` and `taper` is `'dpss'`.
The amount of spectral smoothing through multi-tapering (Hz).
Note that smoothing frequency specifications are one-sided,
i.e., 4 Hz smoothing means plus-minus 4 Hz, i.e., a 8 Hz smoothing box.
nTaper : int or None
Only valid if `method` is `'mtmfft'` or `'mtmconvol'` and `taper='dpss'`.
Number of orthogonal tapers to use. It is not recommended to set the number
of tapers manually! Leave at `None` for the optimal number to be set automatically.
keeptapers : bool
Only valid if `method` is `'mtmfft'` or `'mtmconvol'`.
If `True`, return spectral estimates for each taper.
Otherwise power spectrum is averaged across tapers,
if and only if `output` is `pow`.
toi : float or array-like or "all"
**Mandatory input** for time-frequency analysis methods (`method` is either
`"mtmconvol"` or `"wavelet"` or `"superlet"`).
If `toi` is scalar, it must be a value between 0 and 1 indicating the
percentage of overlap between time-windows specified by `t_ftimwin` (only
valid if `method` is `'mtmconvol'`).
If `toi` is an array it explicitly selects the centroids of analysis
windows (in seconds), if `toi` is `"all"`, analysis windows are centered
on all samples in the data for `method="mtmconvol"`. For wavelet based
methods (`"wavelet"` or `"superlet"`) toi needs to be either an
equidistant array of time points or "all".
t_ftimwin : positive float
Only valid if `method` is `'mtmconvol'`. Sliding window length (in seconds).
wavelet : str
Only valid if `method` is `'wavelet'`. Wavelet function to use, one of
:data:`~syncopy.specest.const_def.availableWavelets` (see below).
width : positive float
Only valid if `method` is `'wavelet'` and `wavelet` is `'Morlet'`. Nondimensional
frequency constant of Morlet wavelet function. This number should be >= 6,
which corresponds to 6 cycles within the analysis window to ensure sufficient
spectral sampling.
order : positive int
Only valid if `method` is `'wavelet'` and `wavelet` is `'Paul'` or `'DOG'`. Order
of the wavelet function. If `wavelet` is `'Paul'`, `order` should be chosen
>= 4 to ensure that the analysis window contains at least a single oscillation.
At an order of 40, the Paul wavelet exhibits about the same number of cycles
as the Morlet wavelet with a `width` of 6.
All other supported wavelets functions are *real-valued* derivatives of
Gaussians (DOGs). Hence, if `wavelet` is `'DOG'`, `order` represents the derivative order.
The special case of a second order DOG yields a function known as "Mexican Hat",
"Marr" or "Ricker" wavelet, which can be selected alternatively by setting
`wavelet` to `'Mexican_hat'`, `'Marr'` or `'Ricker'`. **Note**: A real-valued
wavelet function encodes *only* information about peaks and discontinuities
in the signal and does *not* provide any information about amplitude or phase.
order_max : int
Only valid if `method` is `'superlet'`.
Maximal order of the superlet set. Controls the maximum
number of cycles within a SL together
with the `c_1` parameter: c_max = c_1 * order_max
order_min : int
Only valid if `method` is `'superlet'`.
Minimal order of the superlet set. Controls
the minimal number of cycles within a SL together
with the `c_1` parameter: c_min = c_1 * order_min
Note that for admissability reasons c_min should be at least 3!
c_1 : int
Only valid if `method` is `'superlet'`.
Number of cycles of the base Morlet wavelet. If set to lower
than 3 increase `order_min` as to never have less than 3 cycles
in a wavelet!
adaptive : bool
Only valid if `method` is `'superlet'`.
Wether to perform multiplicative SLT or fractional adaptive SLT.
If set to True, the order of the wavelet set will increase
linearly with the frequencies of interest from `order_min`
to `order_max`. If set to False the same SL will be used for
all frequencies.
out : None or :class:`SpectralData` object
None if a new :class:`SpectralData` object is to be created, or an empty :class:`SpectralData` object
Returns
-------
spec : :class:`~syncopy.SpectralData`
(Time-)frequency spectrum of input data
Notes
-----
.. [Moca2021] Moca, Vasile V., et al. "Time-frequency super-resolution with superlets."
Nature communications 12.1 (2021): 1-18.
**Options**
.. autodata:: syncopy.specest.const_def.availableMethods
.. autodata:: syncopy.specest.const_def.availableOutputs
.. autodata:: syncopy.specest.const_def.availableTapers
.. autodata:: syncopy.specest.const_def.availableWavelets
Examples
--------
Coming soon...
See also
--------
syncopy.specest.mtmfft.mtmfft : (multi-)tapered Fourier transform of multi-channel time series data
syncopy.specest.mtmconvol.mtmconvol : time-frequency analysis of multi-channel time series data with a sliding window FFT
syncopy.specest.wavelet.wavelet : time-frequency analysis of multi-channel time series data using a wavelet transform
numpy.fft.fft : NumPy's reference FFT implementation
scipy.signal.stft : SciPy's Short Time Fourier Transform
"""
# Make sure our one mandatory input object can be processed
try:
data_parser(data, varname="data", dataclass="AnalogData",
writable=None, empty=False)
except Exception as exc:
raise exc
timeAxis = data.dimord.index("time")
# Get everything of interest in local namespace
defaults = get_defaults(freqanalysis)
lcls = locals()
# check for ineffective additional kwargs
check_passed_kwargs(lcls, defaults, frontend_name="freqanalysis")
# Ensure a valid computational method was selected
if method not in availableMethods:
lgl = "'" + "or '".join(opt + "' " for opt in availableMethods)
raise SPYValueError(legal=lgl, varname="method", actual=method)
# Ensure a valid output format was selected
if output not in spectralConversions.keys():
lgl = "'" + "or '".join(opt + "' " for opt in spectralConversions.keys())
raise SPYValueError(legal=lgl, varname="output", actual=output)
# Parse all Boolean keyword arguments
for vname in ["keeptrials", "keeptapers"]:
if not isinstance(lcls[vname], bool):
raise SPYTypeError(lcls[vname], varname=vname, expected="Bool")
# If only a subset of `data` is to be processed, make some necessary adjustments
# of the sampleinfo and trial lengths
if data._selection is not None:
sinfo = data._selection.trialdefinition[:, :2]
trialList = data._selection.trials
else:
trialList = list(range(len(data.trials)))
sinfo = data.sampleinfo
lenTrials = np.diff(sinfo).squeeze()
if not lenTrials.shape:
lenTrials = lenTrials[None]
numTrials = len(trialList)
# check polyremoval
if polyremoval is not None:
scalar_parser(polyremoval, varname="polyremoval", ntype="int_like", lims=[0, 1])
# --- Padding ---
# Sliding window FFT does not support "fancy" padding
if method == "mtmconvol" and isinstance(pad_to_length, str):
msg = "method 'mtmconvol' only supports in-place padding for windows " +\
"exceeding trial boundaries. Your choice of `pad_to_length = '{}'` will be ignored. "
SPYWarning(msg.format(pad_to_length))
if method == 'mtmfft':
# the actual number of samples in case of later padding
minSampleNum = validate_padding(pad_to_length, lenTrials)
else:
minSampleNum = lenTrials.min()
# Compute length (in samples) of shortest trial
minTrialLength = minSampleNum / data.samplerate
# Shortcut to data sampling interval
dt = 1 / data.samplerate
foi, foilim = validate_foi(foi, foilim, data.samplerate)
# see also https://docs.obspy.org/_modules/obspy/signal/detrend.html#polynomial
if polyremoval is not None:
try:
scalar_parser(polyremoval, varname="polyremoval", lims=[0, 1], ntype="int_like")
except Exception as exc:
raise exc
# Prepare keyword dict for logging (use `lcls` to get actually provided
# keyword values, not defaults set above)
log_dct = {"method": method,
"output": output,
"keeptapers": keeptapers,
"keeptrials": keeptrials,
"polyremoval": polyremoval,
"pad_to_length": pad_to_length}
# --------------------------------
# 1st: Check time-frequency inputs
# to prepare/sanitize `toi`
# --------------------------------
if method in ["mtmconvol", "wavelet", "superlet"]:
# Get start/end timing info respecting potential in-place selection
if toi is None:
raise SPYTypeError(toi, varname="toi", expected="scalar or array-like or 'all'")
if data._selection is not None:
tStart = data._selection.trialdefinition[:, 2] / data.samplerate
else:
tStart = data._t0 / data.samplerate
tEnd = tStart + lenTrials / data.samplerate
# for these methods only 'all' or an equidistant array
# of time points (sub-sampling, trimming) are valid
if method in ["wavelet", "superlet"]:
valid = True
if isinstance(toi, Number):
valid = False
elif isinstance(toi, str):
if toi != "all":
valid = False
else:
# take everything
preSelect = [slice(None)] * numTrials
postSelect = [slice(None)] * numTrials
elif not iter(toi):
valid = False
# this is the sequence type - can only be an interval!
else:
try:
array_parser(toi, varname="toi", hasinf=False, hasnan=False,
lims=[tStart.min(), tEnd.max()], dims=(None,))
except Exception as exc:
raise exc
toi = np.array(toi)
# check for equidistancy
if not np.allclose(np.diff(toi, 2), np.zeros(len(toi) - 2)):
valid = False
# trim (preSelect) and subsample output (postSelect)
else:
preSelect = []
postSelect = []
# get sample intervals and relative indices from toi
for tk in range(numTrials):
start = int(data.samplerate * (toi[0] - tStart[tk]))
stop = int(data.samplerate * (toi[-1] - tStart[tk]) + 1)
preSelect.append(slice(max(0, start), max(stop, stop - start)))
smpIdx = np.minimum(lenTrials[tk] - 1,
data.samplerate * (toi - tStart[tk]) - start)
postSelect.append(smpIdx.astype(np.intp))
# get out if sth wasn't right
if not valid:
lgl = "array of equidistant time-points or 'all' for wavelet based methods"
raise SPYValueError(legal=lgl, varname="toi", actual=toi)
# Update `log_dct` w/method-specific options (use `lcls` to get actually
# provided keyword values, not defaults set in here)
log_dct["toi"] = lcls["toi"]
# --------------------------------------------
# Check options specific to mtm*-methods
# (particularly tapers and foi/freqs alignment)
# --------------------------------------------
if "mtm" in method:
if method == "mtmconvol":
# get the sliding window size
try:
scalar_parser(t_ftimwin, varname="t_ftimwin",
lims=[dt, minTrialLength])
except Exception as exc:
SPYInfo("Please specify 't_ftimwin' parameter.. exiting!")
raise exc
# this is the effective sliding window FFT sample size
minSampleNum = int(t_ftimwin * data.samplerate)
# Construct array of maximally attainable frequencies
freqs = np.fft.rfftfreq(minSampleNum, dt)
# Match desired frequencies as close as possible to
# actually attainable freqs
# these are the frequencies attached to the SpectralData by the CR!
if foi is not None:
foi, _ = best_match(freqs, foi, squash_duplicates=True)
elif foilim is not None:
foi, _ = best_match(freqs, foilim, span=True, squash_duplicates=True)
else:
msg = (f"Automatic FFT frequency selection from {freqs[0]:.1f}Hz to "
f"{freqs[-1]:.1f}Hz")
SPYInfo(msg)
foi = freqs
log_dct["foi"] = foi
# Abort if desired frequency selection is empty
if foi.size == 0:
lgl = "non-empty frequency specification"
act = "empty frequency selection"
raise SPYValueError(legal=lgl, varname="foi/foilim", actual=act)
# sanitize taper selection and retrieve dpss settings
taper_opt = validate_taper(taper,
tapsmofrq,
nTaper,
keeptapers,
foimax=foi.max(),
samplerate=data.samplerate,
nSamples=minSampleNum,
output=output)
# Update `log_dct` w/method-specific options
log_dct["taper"] = taper
# only dpss returns non-empty taper_opt dict
if taper_opt:
log_dct["nTaper"] = taper_opt["Kmax"]
log_dct["tapsmofrq"] = tapsmofrq
# -------------------------------------------------------
# Now, prepare explicit compute-classes for chosen method
# -------------------------------------------------------
if method == "mtmfft":
check_effective_parameters(MultiTaperFFT, defaults, lcls)
# method specific parameters
method_kwargs = {
'samplerate': data.samplerate,
'taper': taper,
'taper_opt': taper_opt,
'nSamples': minSampleNum
}
# Set up compute-class
specestMethod = MultiTaperFFT(
foi=foi,
timeAxis=timeAxis,
keeptapers=keeptapers,
polyremoval=polyremoval,
output_fmt=output,
method_kwargs=method_kwargs)
elif method == "mtmconvol":
check_effective_parameters(MultiTaperFFTConvol, defaults, lcls)
# Process `toi` for sliding window multi taper fft,
# we have to account for three scenarios: (1) center sliding
# windows on all samples in (selected) trials (2) `toi` was provided as
# percentage indicating the degree of overlap b/w time-windows and (3) a set
# of discrete time points was provided. These three cases are encoded in
# `overlap, i.e., ``overlap > 1` => all, `0 < overlap < 1` => percentage,
# `overlap < 0` => discrete `toi`
# overlap = None
if isinstance(toi, str):
if toi != "all":
lgl = "`toi = 'all'` to center analysis windows on all time-points"
raise SPYValueError(legal=lgl, varname="toi", actual=toi)
equidistant = True
overlap = np.inf
elif isinstance(toi, Number):
try:
scalar_parser(toi, varname="toi", lims=[0, 1])
except Exception as exc:
raise exc
overlap = toi
equidistant = True
# this captures all other cases, e.i. toi is of sequence type
else:
overlap = -1
try:
array_parser(toi, varname="toi", hasinf=False, hasnan=False,
lims=[tStart.min(), tEnd.max()], dims=(None,))
except Exception as exc:
raise exc
toi = np.array(toi)
tSteps = np.diff(toi)
if (tSteps < 0).any():
lgl = "ordered list/array of time-points"
act = "unsorted list/array"
raise SPYValueError(legal=lgl, varname="toi", actual=act)
# Account for round-off errors: if toi spacing is almost at sample interval
# manually correct it
if np.isclose(tSteps.min(), dt):
tSteps[np.isclose(tSteps, dt)] = dt
if tSteps.min() < dt:
msg = f"`toi` selection too fine, max. time resolution is {dt}s"
SPYWarning(msg)
# This is imho a bug in NumPy - even `arange` and `linspace` may produce
# arrays that are numerically not exactly equidistant - `unique` will
# show several entries here - use `allclose` to identify "even" spacings
equidistant = np.allclose(tSteps, [tSteps[0]] * tSteps.size)
# If `toi` was 'all' or a percentage, use entire time interval of (selected)
# trials and check if those trials have *approximately* equal length
if toi is None:
if not np.allclose(lenTrials, [minSampleNum] * lenTrials.size):
msg = "processing trials of different lengths (min = {}; max = {} samples)" +\
" with `toi = 'all'`"
SPYWarning(msg.format(int(minSampleNum), int(lenTrials.max())))
# number of samples per window
nperseg = int(t_ftimwin * data.samplerate)
halfWin = int(nperseg / 2)
postSelect = slice(None) # select all is the default
if 0 <= overlap <= 1: # `toi` is percentage
noverlap = min(nperseg - 1, int(overlap * nperseg))
# windows get shifted exactly 1 sample
# to get a spectral estimate at each sample
else:
noverlap = nperseg - 1
# `toi` is array
if overlap < 0:
# Compute necessary padding at begin/end of trials to fit sliding windows
offStart = ((toi[0] - tStart) * data.samplerate).astype(np.intp)
padBegin = halfWin - offStart
padBegin = ((padBegin > 0) * padBegin).astype(np.intp)
offEnd = ((tEnd - toi[-1]) * data.samplerate).astype(np.intp)
padEnd = halfWin - offEnd
padEnd = ((padEnd > 0) * padEnd).astype(np.intp)
# Compute sample-indices (one slice/list per trial) from time-selections
soi = []
if equidistant:
# soi just trims the input data to the [toi[0], toi[-1]] interval
# postSelect then subsamples the spectral esimate to the user given toi
postSelect = []
for tk in range(numTrials):
start = max(0, int(round(data.samplerate * (toi[0] - tStart[tk]) - halfWin)))
stop = int(round(data.samplerate * (toi[-1] - tStart[tk]) + halfWin + 1))
soi.append(slice(start, max(stop, stop - start)))
# chosen toi subsampling interval in sample units, min. is 1;
# compute `delta_idx` s.t. stop - start / delta_idx == toi.size
delta_idx = int(round((soi[0].stop - soi[0].start) / toi.size))
delta_idx = delta_idx if delta_idx > 1 else 1
postSelect = slice(None, None, delta_idx)
else:
for tk in range(numTrials):
starts = (data.samplerate * (toi - tStart[tk]) - halfWin).astype(np.intp)
starts += padBegin[tk]
stops = (data.samplerate * (toi - tStart[tk]) + halfWin + 1).astype(np.intp)
stops += padBegin[tk]
stops = np.maximum(stops, stops - starts, dtype=np.intp)
soi.append([slice(start, stop) for start, stop in zip(starts, stops)])
# postSelect here remains slice(None), as resulting spectrum
# has exactly one entry for each soi
# `toi` is percentage or "all"
else:
soi = [slice(None)] * numTrials
# Collect keyword args for `mtmconvol` in dictionary
method_kwargs = {"samplerate": data.samplerate,
"nperseg": nperseg,
"noverlap": noverlap,
"taper" : taper,
"taper_opt" : taper_opt}
# Set up compute-class
specestMethod = MultiTaperFFTConvol(
soi,
postSelect,
equidistant=equidistant,
toi=toi,
foi=foi,
timeAxis=timeAxis,
keeptapers=keeptapers,
polyremoval=polyremoval,
output_fmt=output,
method_kwargs=method_kwargs)
elif method == "wavelet":
check_effective_parameters(WaveletTransform, defaults, lcls)
# Check wavelet selection
if wavelet not in availableWavelets:
lgl = "'" + "or '".join(opt + "' " for opt in availableWavelets)
raise SPYValueError(legal=lgl, varname="wavelet", actual=wavelet)
if wavelet not in ["Morlet", "Paul"]:
msg = "the chosen wavelet '{}' is real-valued and does not provide " +\
"any information about amplitude or phase of the data. This wavelet function " +\
"may be used to isolate peaks or discontinuities in the signal. "
SPYWarning(msg.format(wavelet))
# Check for consistency of `width`, `order` and `wavelet`
if wavelet == "Morlet":
try:
scalar_parser(width, varname="width", lims=[1, np.inf])
except Exception as exc:
raise exc
wfun = getattr(spywave, wavelet)(w0=width)
else:
if width != lcls["width"]:
msg = "option `width` has no effect for wavelet '{}'"
SPYWarning(msg.format(wavelet))
if wavelet == "Paul":
try:
scalar_parser(order, varname="order", lims=[4, np.inf], ntype="int_like")
except Exception as exc:
raise exc
wfun = getattr(spywave, wavelet)(m=order)
elif wavelet == "DOG":
try:
scalar_parser(order, varname="order", lims=[1, np.inf], ntype="int_like")
except Exception as exc:
raise exc
wfun = getattr(spywave, wavelet)(m=order)
else:
if order is not None:
msg = "option `order` has no effect for wavelet '{}'"
SPYWarning(msg.format(wavelet))
wfun = getattr(spywave, wavelet)()
# automatic frequency selection
if foi is None and foilim is None:
scales = get_optimal_wavelet_scales(
wfun.scale_from_period, # all availableWavelets sport one!
int(minTrialLength * data.samplerate),
dt)
foi = 1 / wfun.fourier_period(scales)
msg = (f"Setting frequencies of interest to {foi[0]:.1f}-"
f"{foi[-1]:.1f}Hz")
SPYInfo(msg)
else:
if foilim is not None:
foi = np.arange(foilim[0], foilim[1] + 1, dtype=float)
# 0 frequency is not valid
foi[foi < 0.01] = 0.01
scales = wfun.scale_from_period(1 / foi)
# Update `log_dct` w/method-specific options (use `lcls` to get actually
# provided keyword values, not defaults set in here)
log_dct["foi"] = foi
log_dct["wavelet"] = lcls["wavelet"]
log_dct["width"] = lcls["width"]
log_dct["order"] = lcls["order"]
# method specific parameters
method_kwargs = {
'samplerate' : data.samplerate,
'scales' : scales,
'wavelet' : wfun
}
# Set up compute-class
specestMethod = WaveletTransform(
preSelect,
postSelect,
toi=toi,
timeAxis=timeAxis,
polyremoval=polyremoval,
output_fmt=output,
method_kwargs=method_kwargs)
elif method == "superlet":
check_effective_parameters(SuperletTransform, defaults, lcls)
# check and parse superlet specific arguments
if order_max is None:
lgl = "Positive integer needed for order_max"
raise SPYValueError(legal=lgl, varname="order_max",
actual=None)
else:
scalar_parser(
order_max,
varname="order_max",
lims=[1, np.inf],
ntype="int_like"
)
scalar_parser(
order_min, varname="order_min",
lims=[1, order_max],
ntype="int_like"
)
scalar_parser(c_1, varname="c_1", lims=[1, np.inf], ntype="int_like")
# if no frequencies are user selected, take a sensitive default
if foi is None and foilim is None:
scales = get_optimal_wavelet_scales(
superlet.scale_from_period,
int(minTrialLength * data.samplerate),
dt)
foi = 1 / superlet.fourier_period(scales)
msg = (f"Setting frequencies of interest to {foi[0]:.1f}-"
f"{foi[-1]:.1f}Hz")
SPYInfo(msg)
else:
if foilim is not None:
# frequency range in 1Hz steps
foi = np.arange(foilim[0], foilim[1] + 1, dtype=float)
# 0 frequency is not valid
foi[foi < 0.01] = 0.01
scales = superlet.scale_from_period(1. / foi)
# FASLT needs ordered frequencies low - high
# meaning the scales have to go high - low
if adaptive:
if len(scales) < 2:
lgl = "A range of frequencies"
act = "Single frequency"
raise SPYValueError(legal=lgl, varname="foi", actual=act)
if np.any(np.diff(scales) > 0):
msg = "Sorting frequencies low to high for adaptive SLT.."
SPYWarning(msg)
scales = np.sort(scales)[::-1]
log_dct["foi"] = foi
log_dct["c_1"] = lcls["c_1"]
log_dct["order_max"] = lcls["order_max"]
log_dct["order_min"] = lcls["order_min"]
# method specific parameters
method_kwargs = {
'samplerate' : data.samplerate,
'scales' : scales,
'order_max' : order_max,
'order_min' : order_min,
'c_1' : c_1,
'adaptive' : adaptive
}
# Set up compute-class
specestMethod = SuperletTransform(
preSelect,
postSelect,
toi=toi,
timeAxis=timeAxis,
polyremoval=polyremoval,
output_fmt=output,
method_kwargs=method_kwargs)
# -------------------------------------------------
# Sanitize output and call the ComputationalRoutine
# -------------------------------------------------
# If provided, make sure output object is appropriate
if out is not None:
try:
data_parser(out, varname="out", writable=True, empty=True,
dataclass="SpectralData",
dimord=SpectralData().dimord)
except Exception as exc:
raise exc
new_out = False
else:
out = SpectralData(dimord=SpectralData._defaultDimord)
new_out = True
# Perform actual computation
specestMethod.initialize(data,
out._stackingDim,
chan_per_worker=kwargs.get("chan_per_worker"),
keeptrials=keeptrials)
specestMethod.compute(data, out, parallel=kwargs.get("parallel"), log_dict=log_dct)
# Either return newly created output object or simply quit
return out if new_out else None
def cleanup(older_than=24, **kwargs):
"""
Delete old files in temporary Syncopy folder
The location of the temporary folder is stored in `syncopy.__storage__`.
Parameters
----------
older_than : int
Files older than `older_than` hours will be removed
Examples
--------
>>> spy.cleanup()
"""
# Make sure age-cutoff is valid
try:
scalar_parser(older_than, varname="older_than", ntype="int_like",
lims=[0, np.inf])
except Exception as exc:
raise exc
older_than = int(older_than)
# For clarification: show location of storage folder that is scanned here
funcName = "Syncopy <{}>".format(inspect.currentframe().f_code.co_name)
dirInfo = \
"\n{name:s} Analyzing temporary storage folder {dir:s}...\n"
print(dirInfo.format(name=funcName, dir=__storage__))
# Parse "hidden" interactive keyword: if `False`, don't ask, just delete
interactive = kwargs.get("interactive", True)
if not isinstance(interactive, bool):
raise SPYTypeError(interactive, varname="interactive", expected="bool")
# Get current date + time and scan package's temp directory for session files
now = datetime.now()
sessions = glob(os.path.join(__storage__, "session*"))
allIds = []
for sess in sessions:
allIds.append(os.path.splitext(os.path.basename(sess))[0].split("_")[1])
# Also check for dangling data (not associated to any session)
data = glob(os.path.join(__storage__, "spy_*"))
dangling = []
for dat in data:
sessid = os.path.splitext(os.path.basename(dat))[0].split("_")[1]
if sessid not in allIds:
dangling.append(dat)
# Cycle through session-logs and identify stuff older than `older_than` hrs
sesList = [] # full path to session files
ageList = [] # session age in days
usrList = [] # session users
sizList = [] # raw session sizes in bytes
ownList = [] # session owners (user@machine)
flsList = [] # files/directories associated to session
for sk, sess in enumerate(sessions):
sessid = allIds[sk]
if sessid != __sessionid__:
with open(sess, "r") as fid:
sesslog = fid.read()
timestr = sesslog[sesslog.find("<") + 1:sesslog.find(">")]
timeobj = datetime.strptime(timestr, '%Y-%m-%d %H:%M:%S')
age = round((now - timeobj).total_seconds()/3600) # age in hrs
if age >= older_than:
sesList.append(sess)
files = glob(os.path.join(__storage__, "*_{}_*".format(sessid)))
flsList.append(files)
ageList.append(round(age/24)) # age in days
usrList.append(sesslog[:sesslog.find("@")])
ownList.append(sesslog[:sesslog.find(":")])
sizList.append(sum(os.path.getsize(file) if os.path.isfile(file) else
sum(os.path.getsize(os.path.join(dirpth, fname)) \
for dirpth, _, fnames in os.walk(file)
for fname in fnames) for file in files))
# Farewell if nothing's to do here
if not sesList and not dangling:
ext = \
"Did not find any dangling data or Syncopy session remains " +\
"older than {age:d} hours."
print(ext.format(name=funcName, age=older_than))
return
# Prepare session-related info prompt
if sesList:
usrList = list(set(usrList))
gbList = [sz/1024**3 for sz in sizList]
sesInfo = \
"Found data of {numsess:d} syncopy sessions {ageinfo:s} " +\
"created by user{users:s}'\ntaking up {gbinfo:s} of disk space. \n"
sesInfo = sesInfo.format(numsess=len(sesList),
ageinfo="between {agemin:d} and {agemax:d} days old".format(agemin=min(ageList),
agemax=max(ageList)) \
if min(ageList) < max(ageList) else "from {} days ago".format(ageList[0]),
users="(s) '" + ",".join(usr + ", " for usr in usrList)[:-2] \
if len(usrList) > 1 else " '" + usrList[0],
gbinfo="a total of {gbsz:4.1f} GB".format(gbsz=sum(gbList)) \
if sum(gbList) > 1 else "less than 1 GB")
sesOptions = \
"[I]NTERACTIVE walkthrough to decide which session to remove \n" +\
"[S]ESSION removal to delete all sessions at once " +\
"(you will not be prompted for confirmation) \n"
sesValid = ["I", "S"]
promptInfo = sesInfo
promptOptions = sesOptions
promptValid = sesValid
# Prepare info prompt for dangling files
if dangling:
dangInfo = \
"Found {numdang:d} dangling files not associated to any session " +\
"using {szdang:4.1f} GB of disk space. \n"
dangInfo = dangInfo.format(numdang=len(dangling),
szdang=sum(os.path.getsize(file)/1024**3 if os.path.isfile(file) else \
sum(os.path.getsize(os.path.join(dirpth, fname))/1024**3 \
for dirpth, _, fnames in os.walk(file) \
for fname in fnames) for file in dangling))
dangOptions = \
"[D]ANGLING FILE removal to delete anything not associated to sessions " +\
"(you will not be prompted for confirmation) \n"
dangValid = ["D"]
promptInfo = dangInfo
promptOptions = dangOptions
promptValid = dangValid
# Put together actual prompt message message
promptChoice = "\nPlease choose one of the following options:\n"
abortOption = "[C]ANCEL\n"
abortValid = ["C"]
if sesList and dangling:
rmAllOption = \
"[R]EMOVE all (sessions and dangling files) at once " +\
"(you will not be prompted for confirmation)\n"
rmAllValid = ["R"]
promptInfo = sesInfo + dangInfo
promptOptions = sesOptions + dangOptions + rmAllOption
promptValid = sesValid + dangValid + rmAllValid
# By default, ask what to do; if `interactive` is `False`, remove everything
if interactive:
choice = user_input(promptInfo + promptChoice + promptOptions + abortOption,
valid=promptValid + abortValid)
else:
choice = "R"
# Query removal of data session by session
if choice == "I":
promptYesNo = \
"Found{numf:s} files created by session {sess:s} {age:d} " +\
"days ago{sizeinfo:s} Do you want to permanently delete these files?"
for sk in range(len(sesList)):
if user_yesno(promptYesNo.format(numf=" " + str(len(flsList[sk])),
sess=ownList[sk],
age=ageList[sk],
sizeinfo=" using " + \
str(round(sizList[sk]/1024**2)) + \
" MB of disk space.")):
_rm_session(flsList[sk])
# Delete all session-remains at once
elif choice == "S":
for fls in tqdm(flsList, desc="Deleting session data..."):
_rm_session(fls)
# Deleate all dangling files at once
elif choice == "D":
for dat in tqdm(dangling, desc="Deleting dangling data..."):
_rm_session([dat])
# Delete everything
elif choice == "R":
for contents in tqdm(flsList + [[dat] for dat in dangling],
desc="Deleting temporary data..."):
_rm_session(contents)
# Don't do anything for now, continue w/dangling data
else:
print("Aborting...")
return
def connectivityanalysis(data,
method="coh",
keeptrials=False,
output="abs",
foi=None,
foilim=None,
pad_to_length=None,
polyremoval=None,
taper="hann",
tapsmofrq=None,
nTaper=None,
out=None,
**kwargs):
"""
Perform connectivity analysis of Syncopy :class:`~syncopy.AnalogData` objects
**Usage Summary**
Options available in all analysis methods:
* **foi**/**foilim** : frequencies of interest; either array of frequencies or
frequency window (not both)
* **polyremoval** : de-trending method to use (0 = mean, 1 = linear or `None`)
List of available analysis methods and respective distinct options:
"coh" : (Multi-) tapered coherency estimate
Compute the normalized cross spectral densities
between all channel combinations
* **output** : one of ('abs', 'pow', 'fourier')
* **taper** : one of :data:`~syncopy.shared.const_def.availableTapers`
* **tapsmofrq** : spectral smoothing box for slepian tapers (in Hz)
* **nTaper** : (optional) number of orthogonal tapers for slepian tapers
* **pad_to_length**: either pad to an absolute length or set to `'nextpow2'`
"corr" : Cross-correlations
Computes the one sided (positive lags) cross-correlations
between all channel combinations. The maximal lag is half
the trial lengths.
* **keeptrials** : set to `True` for single trial cross-correlations
"granger" : Spectral Granger-Geweke causality
Computes linear causality estimates between
all channel combinations. The intermediate cross-spectral
densities can be computed via multi-tapering.
* **taper** : one of :data:`~syncopy.shared.const_def.availableTapers`
* **tapsmofrq** : spectral smoothing box for slepian tapers (in Hz)
* **nTaper** : (optional, not recommended) number of slepian tapers
* **pad_to_length**: either pad to an absolute length or set to `'nextpow2'`
Parameters
----------
data : `~syncopy.AnalogData`
A non-empty Syncopy :class:`~syncopy.datatype.AnalogData` object
method : str
Connectivity estimation method, one of 'coh', 'corr', 'granger'
output : str
Relevant for cross-spectral density estimation (`method='coh'`)
Use `'pow'` for absolute squared coherence, `'abs'` for absolute value of coherence
and`'fourier'` for the complex valued coherency.
keeptrials : bool
Relevant for cross-correlations (`method='corr'`).
If `True` single-trial cross-correlations are returned.
foi : array-like or None
Frequencies of interest (Hz) for output. If desired frequencies cannot be
matched exactly, the closest possible frequencies are used. If `foi` is `None`
or ``foi = "all"``, all attainable frequencies (i.e., zero to Nyquist / 2)
are selected.
foilim : array-like (floats [fmin, fmax]) or None or "all"
Frequency-window ``[fmin, fmax]`` (in Hz) of interest. The
`foi` array will be constructed in 1Hz steps from `fmin` to
`fmax` (inclusive).
pad_to_length : int, None or 'nextpow2'
Padding of the (tapered) signal, if set to a number pads all trials
to this absolute length. E.g. `pad_to_length=2000` pads all
trials to 2000 samples, if and only if the longest trial is
at maximum 2000 samples.
Alternatively if all trials have the same initial lengths
setting `pad_to_length='nextpow2'` pads all trials to
the next power of two.
If `None` and trials have unequal lengths all trials are padded to match
the longest trial.
taper : str
Only valid if `method` is `'coh'` or `'granger'`. Windowing function,
one of :data:`~syncopy.specest.const_def.availableTapers`
tapsmofrq : float
Only valid if `method` is `'coh'` or `'granger'` and `taper` is `'dpss'`.
The amount of spectral smoothing through multi-tapering (Hz).
Note that smoothing frequency specifications are one-sided,
i.e., 4 Hz smoothing means plus-minus 4 Hz, i.e., a 8 Hz smoothing box.
nTaper : int or None
Only valid if `method` is `'coh'` or `'granger'` and ``taper = 'dpss'``.
Number of orthogonal tapers to use. It is not recommended to set the number
of tapers manually! Leave at `None` for the optimal number to be set automatically.
Examples
--------
Coming soon...
"""
# Make sure our one mandatory input object can be processed
try:
data_parser(data,
varname="data",
dataclass="AnalogData",
writable=None,
empty=False)
except Exception as exc:
raise exc
timeAxis = data.dimord.index("time")
# Get everything of interest in local namespace
defaults = get_defaults(connectivityanalysis)
lcls = locals()
# check for ineffective additional kwargs
check_passed_kwargs(lcls, defaults, frontend_name="connectivity")
# Ensure a valid computational method was selected
if method not in availableMethods:
lgl = "'" + "or '".join(opt + "' " for opt in availableMethods)
raise SPYValueError(legal=lgl, varname="method", actual=method)
# if a subset selection is present
# get sampleinfo and check for equidistancy
if data._selection is not None:
sinfo = data._selection.trialdefinition[:, :2]
trialList = data._selection.trials
# user picked discrete set of time points
if isinstance(data._selection.time[0], list):
lgl = "equidistant time points (toi) or time slice (toilim)"
actual = "non-equidistant set of time points"
raise SPYValueError(legal=lgl, varname="select", actual=actual)
else:
trialList = list(range(len(data.trials)))
sinfo = data.sampleinfo
lenTrials = np.diff(sinfo).squeeze()
# check polyremoval
if polyremoval is not None:
scalar_parser(polyremoval,
varname="polyremoval",
ntype="int_like",
lims=[0, 1])
# --- Padding ---
if method == "corr" and pad_to_length:
lgl = "`None`, no padding needed/allowed for cross-correlations"
actual = f"{pad_to_length}"
raise SPYValueError(legal=lgl, varname="pad_to_length", actual=actual)
# the actual number of samples in case of later padding
nSamples = validate_padding(pad_to_length, lenTrials)
# --- Basic foi sanitization ---
foi, foilim = validate_foi(foi, foilim, data.samplerate)
# only now set foi array for foilim in 1Hz steps
if foilim is not None:
foi = np.arange(foilim[0], foilim[1] + 1, dtype=float)
# Prepare keyword dict for logging (use `lcls` to get actually provided
# keyword values, not defaults set above)
log_dict = {
"method": method,
"output": output,
"keeptrials": keeptrials,
"polyremoval": polyremoval,
"pad_to_length": pad_to_length
}
# --- Setting up specific Methods ---
if method in ['coh', 'granger']:
# --- set up computation of the single trial CSDs ---
if keeptrials is not False:
lgl = "False, trial averaging needed!"
act = keeptrials
raise SPYValueError(lgl, varname="keeptrials", actual=act)
# Construct array of maximally attainable frequencies
freqs = np.fft.rfftfreq(nSamples, 1 / data.samplerate)
# Match desired frequencies as close as possible to
# actually attainable freqs
# these are the frequencies attached to the SpectralData by the CR!
if foi is not None:
foi, _ = best_match(freqs, foi, squash_duplicates=True)
elif foilim is not None:
foi, _ = best_match(freqs,
foilim,
span=True,
squash_duplicates=True)
elif foi is None and foilim is None:
# Construct array of maximally attainable frequencies
msg = (f"Setting frequencies of interest to {freqs[0]:.1f}-"
f"{freqs[-1]:.1f}Hz")
SPYInfo(msg)
foi = freqs
# sanitize taper selection and retrieve dpss settings
taper_opt = validate_taper(
taper,
tapsmofrq,
nTaper,
keeptapers=False, # ST_CSD's always average tapers
foimax=foi.max(),
samplerate=data.samplerate,
nSamples=nSamples,
output="pow") # ST_CSD's always have this unit/norm
log_dict["foi"] = foi
log_dict["taper"] = taper
# only dpss returns non-empty taper_opt dict
if taper_opt:
log_dict["nTaper"] = taper_opt["Kmax"]
log_dict["tapsmofrq"] = tapsmofrq
check_effective_parameters(ST_CrossSpectra, defaults, lcls)
# parallel computation over trials
st_compRoutine = ST_CrossSpectra(samplerate=data.samplerate,
nSamples=nSamples,
taper=taper,
taper_opt=taper_opt,
polyremoval=polyremoval,
timeAxis=timeAxis,
foi=foi)
# hard coded as class attribute
st_dimord = ST_CrossSpectra.dimord
if method == 'coh':
# final normalization after trial averaging
av_compRoutine = NormalizeCrossSpectra(output=output)
if method == 'granger':
# after trial averaging
# hardcoded numerical parameters
av_compRoutine = GrangerCausality(rtol=1e-8, nIter=100, cond_max=1e4)
if method == 'corr':
if lcls['foi'] is not None:
msg = 'Parameter `foi` has no effect for `corr`'
SPYWarning(msg)
check_effective_parameters(ST_CrossCovariance, defaults, lcls)
# single trial cross-correlations
if keeptrials:
av_compRoutine = None # no trial average
norm = True # normalize individual trials within the ST CR
else:
av_compRoutine = NormalizeCrossCov()
norm = False
# parallel computation over trials
st_compRoutine = ST_CrossCovariance(samplerate=data.samplerate,
polyremoval=polyremoval,
timeAxis=timeAxis,
norm=norm)
# hard coded as class attribute
st_dimord = ST_CrossCovariance.dimord
# -------------------------------------------------
# Call the chosen single trial ComputationalRoutine
# -------------------------------------------------
# the single trial results need a new DataSet
st_out = CrossSpectralData(dimord=st_dimord)
# Perform the trial-parallelized computation of the matrix quantity
st_compRoutine.initialize(
data,
st_out._stackingDim,
chan_per_worker=None, # no parallelisation over channels possible
keeptrials=keeptrials) # we most likely need trial averaging!
st_compRoutine.compute(data,
st_out,
parallel=kwargs.get("parallel"),
log_dict=log_dict)
# if ever needed..
# for single trial cross-corr results <-> keeptrials is True
if keeptrials and av_compRoutine is None:
if out is not None:
msg = "Single trial processing does not support `out` argument but directly returns the results"
SPYWarning(msg)
return st_out
# ----------------------------------------------------------------------------------
# Sanitize output and call the chosen ComputationalRoutine on the averaged ST output
# ----------------------------------------------------------------------------------
# If provided, make sure output object is appropriate
if out is not None:
try:
data_parser(out,
varname="out",
writable=True,
empty=True,
dataclass="CrossSpectralData",
dimord=st_dimord)
except Exception as exc:
raise exc
new_out = False
else:
out = CrossSpectralData(dimord=st_dimord)
new_out = True
# now take the trial average from the single trial CR as input
av_compRoutine.initialize(st_out, out._stackingDim, chan_per_worker=None)
av_compRoutine.pre_check() # make sure we got a trial_average
av_compRoutine.compute(st_out, out, parallel=False, log_dict=log_dict)
# Either return newly created output object or simply quit
return out if new_out else None
def esi_cluster_setup(partition="8GBS", n_jobs=2, mem_per_job=None,
timeout=180, interactive=True, start_client=True,
**kwargs):
"""
Start a distributed Dask cluster of parallel processing workers using SLURM
(or local multi-processing)
Parameters
----------
partition : str
Name of SLURM partition/queue to use
n_jobs : int
Number of jobs to spawn
mem_per_job : None or str
Memory booking for each job. Can be specified either in megabytes
(e.g., ``mem_per_job = 1500MB``) or gigabytes (e.g., ``mem_per_job = "2GB"``).
If `mem_per_job` is `None`, it is attempted to infer a sane default value
from the chosen queue, e.g., for ``partition = "8GBS"`` `mem_per_job` is
automatically set to the allowed maximum of `'8GB'`. However, even in
queues with guaranted memory bookings, it is possible to allocate less
memory than the allowed maximum per job to spawn numerous low-memory
jobs. See Examples for details.
timeout : int
Number of seconds to wait for requested jobs to start up.
interactive : bool
If `True`, user input is required in case not all jobs could
be started in the provided waiting period (determined by `timeout`).
If `interactive` is `False` and the jobs could not be started
within `timeout` seconds, a `TimeoutError` is raised.
start_client : bool
If `True`, a distributed computing client is launched and attached to
the workers. If `start_client` is `False`, only a distributed
computing cluster is started to which compute-clients can connect.
**kwargs : dict
Additional keyword arguments can be used to control job-submission details.
Returns
-------
proc : object
A distributed computing client (if ``start_client = True``) or
a distributed computing cluster (otherwise).
Examples
--------
The following command launches 10 SLURM jobs with 2 gigabytes memory each
in the `8GBS` partition
>>> spy.esi_cluster_setup(n_jobs=10, partition="8GBS", mem_per_job="2GB")
If you want to access properties of the created distributed computing client,
assign an explicit return quantity, i.e.,
>>> client = spy.esi_cluster_setup(n_jobs=10, partition="8GBS", mem_per_job="2GB")
The underlying distributed computing cluster can be accessed using
>>> client.cluster
Notes
-----
Syncopy's parallel computing engine relies on the concurrent processing library
`Dask <https://docs.dask.org/en/latest/>`_. Thus, the distributed computing
clients used by Syncopy are in fact instances of :class:`dask.distributed.Client`.
This function specifically acts as a wrapper for :class:`dask_jobqueue.SLURMCluster`.
Users familiar with Dask in general and its distributed scheduler and cluster
objects in particular, may leverage Dask's entire API to fine-tune parallel
processing jobs to their liking (if wanted).
See also
--------
cluster_cleanup : remove dangling parallel processing job-clusters
"""
# For later reference: dynamically fetch name of current function
funcName = "Syncopy <{}>".format(inspect.currentframe().f_code.co_name)
# Be optimistic: prepare success message
successMsg = "{name:s} Cluster dashboard accessible at {dash:s}"
# Retrieve all partitions currently available in SLURM
out, err = subprocess.Popen("sinfo -h -o %P",
stdout=subprocess.PIPE, stderr=subprocess.PIPE,
text=True, shell=True).communicate()
if len(err) > 0:
# SLURM is not installed, either allocate `LocalCluster` or just leave
if "sinfo: not found" in err:
if interactive:
msg = "{name:s} SLURM does not seem to be installed on this machine " +\
"({host:s}). Do you want to start a local multi-processing " +\
"computing client instead? "
startLocal = user_yesno(msg.format(name=funcName, host=socket.gethostname()),
default="no")
else:
startLocal = True
if startLocal:
client = Client()
successMsg = "{name:s} Local parallel computing client ready. \n" + successMsg
print(successMsg.format(name=funcName, dash=client.cluster.dashboard_link))
if start_client:
return client
return client.cluster
return
# SLURM is installed, but something's wrong
msg = "SLURM queuing system from node {node:s}. " +\
"Original error message below:\n{error:s}"
raise SPYIOError(msg.format(node=socket.gethostname(), error=err))
options = out.split()
# Make sure we're in a valid partition (exclude IT partitions from output message)
if partition not in options:
valid = list(set(options).difference(["DEV", "PPC"]))
raise SPYValueError(legal="'" + "or '".join(opt + "' " for opt in valid),
varname="partition", actual=partition)
# Parse job count
try:
scalar_parser(n_jobs, varname="n_jobs", ntype="int_like", lims=[1, np.inf])
except Exception as exc:
raise exc
# Get requested memory per job
if mem_per_job is not None:
if not isinstance(mem_per_job, str):
raise SPYTypeError(mem_per_job, varname="mem_per_job", expected="string")
if not any(szstr in mem_per_job for szstr in ["MB", "GB"]):
lgl = "string representation of requested memory (e.g., '8GB', '12000MB')"
raise SPYValueError(legal=lgl, varname="mem_per_job", actual=mem_per_job)
# Query memory limit of chosen partition and ensure that `mem_per_job` is
# set for partitions w/o limit
idx = partition.find("GB")
if idx > 0:
mem_lim = int(partition[:idx]) * 1000
else:
if partition == "PREPO":
mem_lim = 16000
else:
if mem_per_job is None:
lgl = "explicit memory amount as required by partition '{}'"
raise SPYValueError(legal=lgl.format(partition),
varname="mem_per_job", actual=mem_per_job)
mem_lim = np.inf
# Consolidate requested memory with chosen partition (or assign default memory)
if mem_per_job is None:
mem_per_job = str(mem_lim) + "MB"
else:
if "MB" in mem_per_job:
mem_req = int(mem_per_job[:mem_per_job.find("MB")])
else:
mem_req = int(round(float(mem_per_job[:mem_per_job.find("GB")]) * 1000))
if mem_req > mem_lim:
msg = "`mem_per_job` exceeds limit of {lim:d}GB for partition {par:s}. " +\
"Capping memory at partition limit. "
SPYWarning(msg.format(lim=mem_lim, par=partition))
mem_per_job = str(int(mem_lim)) + "GB"
# Parse requested timeout period
try:
scalar_parser(timeout, varname="timeout", ntype="int_like", lims=[1, np.inf])
except Exception as exc:
raise exc
# Determine if cluster allocation is happening interactively
if not isinstance(interactive, bool):
raise SPYTypeError(interactive, varname="interactive", expected="bool")
# Determine if a dask client was requested
if not isinstance(start_client, bool):
raise SPYTypeError(start_client, varname="start_client", expected="bool")
# Set/get "hidden" kwargs
workers_per_job = kwargs.get("workers_per_job", 1)
try:
scalar_parser(workers_per_job, varname="workers_per_job",
ntype="int_like", lims=[1, 8])
except Exception as exc:
raise exc
n_cores = kwargs.get("n_cores", 1)
try:
scalar_parser(n_cores, varname="n_cores",
ntype="int_like", lims=[1, np.inf])
except Exception as exc:
raise exc
slurm_wdir = kwargs.get("slurmWorkingDirectory", None)
if slurm_wdir is None:
usr = getpass.getuser()
slurm_wdir = "/mnt/hpx/slurm/{usr:s}/{usr:s}_{date:s}"
slurm_wdir = slurm_wdir.format(usr=usr,
date=datetime.now().strftime('%Y%m%d-%H%M%S'))
os.makedirs(slurm_wdir, exist_ok=True)
else:
try:
io_parser(slurm_wdir, varname="slurmWorkingDirectory", isfile=False)
except Exception as exc:
raise exc
# Hotfix for upgraded cluster-nodes: point to correct Python executable if working from /home
pyExec = sys.executable
if sys.executable.startswith("/home"):
pyExec = "/mnt/gs" + sys.executable
# Create `SLURMCluster` object using provided parameters
out_files = os.path.join(slurm_wdir, "slurm-%j.out")
cluster = SLURMCluster(cores=n_cores,
memory=mem_per_job,
processes=workers_per_job,
local_directory=slurm_wdir,
queue=partition,
name="spyswarm",
python=pyExec,
header_skip=["-t", "--mem"],
job_extra=["--output={}".format(out_files)])
# interface="asdf", # interface is set via `psutil.net_if_addrs()`
# job_extra=["--hint=nomultithread",
# "--threads-per-core=1"]
# Compute total no. of workers and up-scale cluster accordingly
total_workers = n_jobs * workers_per_job
cluster.scale(total_workers)
# Fire up waiting routine to avoid premature cluster setups
if _cluster_waiter(cluster, funcName, total_workers, timeout, interactive):
return
# Kill a zombie cluster in non-interactive mode
if not interactive and _count_running_workers(cluster) == 0:
cluster.close()
err = "SLURM jobs could not be started within given time-out " +\
"interval of {0:d} seconds"
raise TimeoutError(err.format(timeout))
# Highlight how to connect to dask performance monitor
print(successMsg.format(name=funcName, dash=cluster.dashboard_link))
# If client was requested, return that instead of the created cluster
if start_client:
return Client(cluster)
return cluster
def freqanalysis(data,
method='mtmfft',
output='fourier',
keeptrials=True,
foi=None,
foilim=None,
pad=None,
padtype='zero',
padlength=None,
prepadlength=None,
postpadlength=None,
polyremoval=None,
taper="hann",
tapsmofrq=None,
keeptapers=False,
toi=None,
t_ftimwin=None,
wav="Morlet",
width=6,
order=None,
out=None,
**kwargs):
"""
Perform (time-)frequency analysis of Syncopy :class:`~syncopy.AnalogData` objects
**Usage Summary**
Options available in all analysis methods:
* **output** : one of :data:`~.availableOutputs`; return power spectra, complex
Fourier spectra or absolute values.
* **foi**/**foilim** : frequencies of interest; either array of frequencies or
frequency window (not both)
* **keeptrials** : return individual trials or grand average
* **polyremoval** : de-trending method to use (0 = mean, 1 = linear, 2 = quadratic,
3 = cubic, etc.)
List of available analysis methods and respective distinct options:
:func:`~syncopy.specest.mtmfft.mtmfft` : (Multi-)tapered Fourier transform
Perform frequency analysis on time-series trial data using either a single
taper window (Hanning) or many tapers based on the discrete prolate
spheroidal sequence (DPSS) that maximize energy concentration in the main
lobe.
* **taper** : one of :data:`~.availableTapers`
* **tapsmofrq** : spectral smoothing box for tapers (in Hz)
* **keeptapers** : return individual tapers or average
* **pad** : padding method to use (`None`, `True`, `False`, `'absolute'`,
`'relative'`, `'maxlen'` or `'nextpow2'`). If `None`, then `'nextpow2'`
is selected by default.
* **padtype** : values to pad data with (`'zero'`, `'nan'`, `'mean'`, `'localmean'`,
`'edge'` or `'mirror'`)
* **padlength** : number of samples to pre-pend and/or append to each trial
* **prepadlength** : number of samples to pre-pend to each trial
* **postpadlength** : number of samples to append to each trial
:func:`~syncopy.specest.mtmconvol.mtmconvol` : (Multi-)tapered sliding window Fourier transform
Perform time-frequency analysis on time-series trial data based on a sliding
window short-time Fourier transform using either a single Hanning taper or
multiple DPSS tapers.
* **taper** : one of :data:`~.availableTapers`
* **tapsmofrq** : spectral smoothing box for tapers (in Hz)
* **keeptapers** : return individual tapers or average
* **pad** : flag indicating, whether or not to pad trials. If `None`,
trials are padded only if sliding window centroids are too close
to trial boundaries for the entire window to cover available data-points.
* **toi** : time-points of interest; can be either an array representing
analysis window centroids (in sec), a scalar between 0 and 1 encoding
the percentage of overlap between adjacent windows or "all" to center
a window on every sample in the data.
* **t_ftimwin** : sliding window length (in sec)
:func:`~syncopy.specest.wavelet.wavelet` : (Continuous non-orthogonal) wavelet transform
Perform time-frequency analysis on time-series trial data using a non-orthogonal
continuous wavelet transform.
* **wav** : one of :data:`~.availableWavelets`
* **toi** : time-points of interest; can be either an array representing
time points (in sec) to center wavelets on or "all" to center a wavelet
on every sample in the data.
* **width** : Nondimensional frequency constant of Morlet wavelet function (>= 6)
* **order** : Order of Paul wavelet function (>= 4) or derivative order
of real-valued DOG wavelets (2 = mexican hat)
**Full documentation below**
Parameters
----------
data : `~syncopy.AnalogData`
A non-empty Syncopy :class:`~syncopy.datatype.AnalogData` object
method : str
Spectral estimation method, one of :data:`~.availableMethods`
(see below).
output : str
Output of spectral estimation. One of :data:`~.availableOutputs` (see below);
use `'pow'` for power spectrum (:obj:`numpy.float32`), `'fourier'` for complex
Fourier coefficients (:obj:`numpy.complex128`) or `'abs'` for absolute
values (:obj:`numpy.float32`).
keeptrials : bool
If `True` spectral estimates of individual trials are returned, otherwise
results are averaged across trials.
foi : array-like or None
Frequencies of interest (Hz) for output. If desired frequencies cannot be
matched exactly, the closest possible frequencies are used. If `foi` is `None`
or ``foi = "all"``, all attainable frequencies (i.e., zero to Nyquist / 2)
are selected.
foilim : array-like (floats [fmin, fmax]) or None or "all"
Frequency-window ``[fmin, fmax]`` (in Hz) of interest. Window
specifications must be sorted (e.g., ``[90, 70]`` is invalid) and not NaN
but may be unbounded (e.g., ``[-np.inf, 60.5]`` is valid). Edges `fmin`
and `fmax` are included in the selection. If `foilim` is `None` or
``foilim = "all"``, all frequencies are selected.
pad : str or None or bool
One of `None`, `True`, `False`, `'absolute'`, `'relative'`, `'maxlen'` or
`'nextpow2'`.
If `pad` is `None` or ``pad = True``, then method-specific defaults are
chosen. Specifically, if `method` is `'mtmfft'` then `pad` is set to
`'nextpow2'` so that all trials in `data` are padded to the next power of
two higher than the sample-count of the longest (selected) trial in `data`. Conversely,
time-frequency analysis methods (`'mtmconvol'` and `'wavelet'`), only perform
padding if necessary, i.e., if time-window centroids are chosen too close
to trial boundaries for the entire window to cover available data-points.
If `pad` is `False`, then no padding is performed. Then in case of
``method = 'mtmfft'`` all trials have to have approximately the same
length (up to the next even sample-count), if ``method = 'mtmconvol'`` or
``method = 'wavelet'``, window-centroids have to keep sufficient
distance from trial boundaries. For more details on the padding methods
`'absolute'`, `'relative'`, `'maxlen'` and `'nextpow2'` see :func:`syncopy.padding`.
padtype : str
Values to be used for padding. Can be `'zero'`, `'nan'`, `'mean'`,
`'localmean'`, `'edge'` or `'mirror'`. See :func:`syncopy.padding` for
more information.
padlength : None, bool or positive int
Only valid if `method` is `'mtmfft'` and `pad` is `'absolute'` or `'relative'`.
Number of samples to pad data with. See :func:`syncopy.padding` for more
information.
prepadlength : None or bool or int
Only valid if `method` is `'mtmfft'` and `pad` is `'relative'`. Number of
samples to pre-pend to each trial. See :func:`syncopy.padding` for more
information.
postpadlength : None or bool or int
Only valid if `method` is `'mtmfft'` and `pad` is `'relative'`. Number of
samples to append to each trial. See :func:`syncopy.padding` for more
information.
polyremoval : int or None
**FIXME: Not implemented yet**
Order of polynomial used for de-trending data in the time domain prior
to spectral analysis. A value of 0 corresponds to subtracting the mean
("de-meaning"), ``polyremoval = 1`` removes linear trends (subtracting the
least squares fit of a linear polynomial), ``polyremoval = N`` for `N > 1`
subtracts a polynomial of order `N` (``N = 2`` quadratic, ``N = 3`` cubic
etc.). If `polyremoval` is `None`, no de-trending is performed.
taper : str
Only valid if `method` is `'mtmfft'` or `'mtmconvol'`. Windowing function,
one of :data:`~.availableTapers` (see below).
tapsmofrq : float
Only valid if `method` is `'mtmfft'` or `'mtmconvol'`. The amount of spectral
smoothing through multi-tapering (Hz). Note that smoothing frequency
specifications are one-sided, i.e., 4 Hz smoothing means plus-minus 4 Hz,
i.e., a 8 Hz smoothing box.
keeptapers : bool
Only valid if `method` is `'mtmfft'` or `'mtmconvol'`. If `True`, return
spectral estimates for each taper, otherwise results are averaged across
tapers.
toi : float or array-like or "all"
**Mandatory input** for time-frequency analysis methods (`method` is either
`"mtmconvol"` or `"wavelet"`).
If `toi` is scalar, it must be a value between 0 and 1 indicating the
percentage of overlap between time-windows specified by `t_ftimwin` (only
valid if `method` is `'mtmconvol'`, invalid for `'wavelet'`).
If `toi` is an array it explicitly selects the centroids of analysis
windows (in seconds). If `toi` is `"all"`, analysis windows are centered
on all samples in the data.
t_ftimwin : positive float
Only valid if `method` is `'mtmconvol'`. Sliding window length (in seconds).
wav : str
Only valid if `method` is `'wavelet'`. Wavelet function to use, one of
:data:`~.availableWavelets` (see below).
width : positive float
Only valid if `method` is `'wavelet'` and `wav` is `'Morlet'`. Nondimensional
frequency constant of Morlet wavelet function. This number should be >= 6,
which corresponds to 6 cycles within the analysis window to ensure sufficient
spectral sampling.
order : positive int
Only valid if `method` is `'wavelet'` and `wav` is `'Paul'` or `'DOG'`. Order
of the wavelet function. If `wav` is `'Paul'`, `order` should be chosen
>= 4 to ensure that the analysis window contains at least a single oscillation.
At an order of 40, the Paul wavelet exhibits about the same number of cycles
as the Morlet wavelet with a `width` of 6.
All other supported wavelets functions are *real-valued* derivatives of
Gaussians (DOGs). Hence, if `wav` is `'DOG'`, `order` represents the derivative order.
The special case of a second order DOG yields a function known as "Mexican Hat",
"Marr" or "Ricker" wavelet, which can be selected alternatively by setting
`wav` to `'Mexican_hat'`, `'Marr'` or `'Ricker'`. **Note**: A real-valued
wavelet function encodes *only* information about peaks and discontinuities
in the signal and does *not* provide any information about amplitude or phase.
out : None or :class:`SpectralData` object
None if a new :class:`SpectralData` object is to be created, or an empty :class:`SpectralData` object
Returns
-------
spec : :class:`~syncopy.SpectralData`
(Time-)frequency spectrum of input data
Notes
-----
Coming soon...
Examples
--------
Coming soon...
.. autodata:: syncopy.specest.freqanalysis.availableMethods
.. autodata:: syncopy.specest.freqanalysis.availableOutputs
.. autodata:: syncopy.specest.freqanalysis.availableTapers
.. autodata:: syncopy.specest.freqanalysis.availableWavelets
See also
--------
syncopy.specest.mtmfft.mtmfft : (multi-)tapered Fourier transform of multi-channel time series data
syncopy.specest.mtmconvol.mtmconvol : time-frequency analysis of multi-channel time series data with a sliding window FFT
syncopy.specest.wavelet.wavelet : time-frequency analysis of multi-channel time series data using a wavelet transform
numpy.fft.fft : NumPy's reference FFT implementation
scipy.signal.stft : SciPy's Short Time Fourier Transform
"""
# Make sure our one mandatory input object can be processed
try:
data_parser(data,
varname="data",
dataclass="AnalogData",
writable=None,
empty=False)
except Exception as exc:
raise exc
timeAxis = data.dimord.index("time")
# Get everything of interest in local namespace
defaults = get_defaults(freqanalysis)
lcls = locals()
# Ensure a valid computational method was selected
if method not in availableMethods:
lgl = "'" + "or '".join(opt + "' " for opt in availableMethods)
raise SPYValueError(legal=lgl, varname="method", actual=method)
# Ensure a valid output format was selected
if output not in spectralConversions.keys():
lgl = "'" + "or '".join(opt + "' "
for opt in spectralConversions.keys())
raise SPYValueError(legal=lgl, varname="output", actual=output)
# Parse all Boolean keyword arguments
for vname in ["keeptrials", "keeptapers"]:
if not isinstance(lcls[vname], bool):
raise SPYTypeError(lcls[vname], varname=vname, expected="Bool")
# If only a subset of `data` is to be processed, make some necessary adjustments
# and compute minimal sample-count across (selected) trials
if data._selection is not None:
trialList = data._selection.trials
sinfo = np.zeros((len(trialList), 2))
for tk, trlno in enumerate(trialList):
trl = data._preview_trial(trlno)
tsel = trl.idx[timeAxis]
if isinstance(tsel, list):
sinfo[tk, :] = [0, len(tsel)]
else:
sinfo[tk, :] = [
trl.idx[timeAxis].start, trl.idx[timeAxis].stop
]
else:
trialList = list(range(len(data.trials)))
sinfo = data.sampleinfo
lenTrials = np.diff(sinfo).squeeze()
numTrials = len(trialList)
# Set default padding options: after this, `pad` is either `None`, `False` or `str`
defaultPadding = {"mtmfft": "nextpow2", "mtmconvol": None, "wavelet": None}
if pad is None or pad is True:
pad = defaultPadding[method]
# Sliding window FFT does not support "fancy" padding
if method == "mtmconvol" and isinstance(pad, str):
msg = "method 'mtmconvol' only supports in-place padding for windows " +\
"exceeding trial boundaries. Your choice of `pad = '{}'` will be ignored. "
SPYWarning(msg.format(pad))
pad = None
# Ensure padding selection makes sense: do not pad on a by-trial basis but
# use the longest trial as reference and compute `padlength` from there
# (only relevant for "global" padding options such as `maxlen` or `nextpow2`)
if pad:
if not isinstance(pad, str):
raise SPYTypeError(pad, varname="pad", expected="str or None")
if pad == "maxlen":
padlength = lenTrials.max()
prepadlength = True
postpadlength = False
elif pad == "nextpow2":
padlength = 0
for ltrl in lenTrials:
padlength = max(padlength, _nextpow2(ltrl))
pad = "absolute"
prepadlength = True
postpadlength = False
padding(data._preview_trial(trialList[0]),
padtype,
pad=pad,
padlength=padlength,
prepadlength=prepadlength,
postpadlength=postpadlength)
# Compute `minSampleNum` accounting for padding
minSamplePos = lenTrials.argmin()
minSampleNum = padding(data._preview_trial(trialList[minSamplePos]),
padtype,
pad=pad,
padlength=padlength,
prepadlength=True).shape[timeAxis]
else:
if method == "mtmfft" and np.unique(
(np.floor(lenTrials / 2))).size > 1:
lgl = "trials of approximately equal length for method 'mtmfft'"
act = "trials of unequal length"
raise SPYValueError(legal=lgl, varname="data", actual=act)
minSampleNum = lenTrials.min()
# Compute length (in samples) of shortest trial
minTrialLength = minSampleNum / data.samplerate
# Basic sanitization of frequency specifications
if foi is not None:
if isinstance(foi, str):
if foi == "all":
foi = None
else:
raise SPYValueError(legal="'all' or `None` or list/array",
varname="foi",
actual=foi)
else:
try:
array_parser(foi,
varname="foi",
hasinf=False,
hasnan=False,
lims=[0, data.samplerate / 2],
dims=(None, ))
except Exception as exc:
raise exc
foi = np.array(foi, dtype="float")
if foilim is not None:
if isinstance(foilim, str):
if foilim == "all":
foilim = None
else:
raise SPYValueError(legal="'all' or `None` or `[fmin, fmax]`",
varname="foilim",
actual=foilim)
else:
try:
array_parser(foilim,
varname="foilim",
hasinf=False,
hasnan=False,
lims=[0, data.samplerate / 2],
dims=(2, ))
except Exception as exc:
raise exc
if foi is not None and foilim is not None:
lgl = "either `foi` or `foilim` specification"
act = "both"
raise SPYValueError(legal=lgl, varname="foi/foilim", actual=act)
# FIXME: implement detrending
# see also https://docs.obspy.org/_modules/obspy/signal/detrend.html#polynomial
if polyremoval is not None:
raise NotImplementedError("Detrending has not been implemented yet.")
try:
scalar_parser(polyremoval,
varname="polyremoval",
lims=[0, 8],
ntype="int_like")
except Exception as exc:
raise exc
# Prepare keyword dict for logging (use `lcls` to get actually provided
# keyword values, not defaults set above)
log_dct = {
"method": method,
"output": output,
"keeptapers": keeptapers,
"keeptrials": keeptrials,
"polyremoval": polyremoval,
"pad": lcls["pad"],
"padtype": lcls["padtype"],
"padlength": lcls["padlength"],
"foi": lcls["foi"]
}
# 1st: Check time-frequency inputs to prepare/sanitize `toi`
if method in ["mtmconvol", "wavelet"]:
# Get start/end timing info respecting potential in-place selection
if toi is None:
raise SPYTypeError(toi,
varname="toi",
expected="scalar or array-like or 'all'")
if data._selection is not None:
tStart = data._selection.trialdefinition[:, 2] / data.samplerate
else:
tStart = data._t0 / data.samplerate
tEnd = tStart + lenTrials / data.samplerate
# Process `toi`: we have to account for three scenarios: (1) center sliding
# windows on all samples in (selected) trials (2) `toi` was provided as
# percentage indicating the degree of overlap b/w time-windows and (3) a set
# of discrete time points was provided. These three cases are encoded in
# `overlap, i.e., ``overlap > 1` => all, `0 < overlap < 1` => percentage,
# `overlap < 0` => discrete `toi`
if isinstance(toi, str):
if toi != "all":
lgl = "`toi = 'all'` to center analysis windows on all time-points"
raise SPYValueError(legal=lgl, varname="toi", actual=toi)
overlap = 1.1
toi = None
equidistant = True
elif isinstance(toi, Number):
if method == "wavelet":
lgl = "array of time-points wavelets are to be centered on"
act = "scalar value"
raise SPYValueError(legal=lgl, varname="toi", actual=act)
try:
scalar_parser(toi, varname="toi", lims=[0, 1])
except Exception as exc:
raise exc
overlap = toi
equidistant = True
else:
overlap = -1
try:
array_parser(toi,
varname="toi",
hasinf=False,
hasnan=False,
lims=[tStart.min(), tEnd.max()],
dims=(None, ))
except Exception as exc:
raise exc
toi = np.array(toi)
tSteps = np.diff(toi)
if (tSteps < 0).any():
lgl = "ordered list/array of time-points"
act = "unsorted list/array"
raise SPYValueError(legal=lgl, varname="toi", actual=act)
# This is imho a bug in NumPy - even `arange` and `linspace` may produce
# arrays that are numerically not exactly equidistant - `unique` will
# show several entries here - use `allclose` to identify "even" spacings
equidistant = np.allclose(tSteps, [tSteps[0]] * tSteps.size)
# If `toi` was 'all' or a percentage, use entire time interval of (selected)
# trials and check if those trials have *approximately* equal length
if toi is None:
if not np.allclose(lenTrials, [minSampleNum] * lenTrials.size):
msg = "processing trials of different lengths (min = {}; max = {} samples)" +\
" with `toi = 'all'`"
SPYWarning(msg.format(int(minSampleNum), int(lenTrials.max())))
if pad is False:
lgl = "`pad` to be `None` or `True` to permit zero-padding " +\
"at trial boundaries to accommodate windows if `0 < toi < 1` " +\
"or if `toi` is 'all'"
act = "False"
raise SPYValueError(legal=lgl, actual=act, varname="pad")
# Code recycling: `overlap`, `equidistant` etc. are really only relevant
# for `mtmconvol`, but we use padding calc below for `wavelet` as well
if method == "mtmconvol":
try:
scalar_parser(t_ftimwin,
varname="t_ftimwin",
lims=[1 / data.samplerate, minTrialLength])
except Exception as exc:
raise exc
else:
t_ftimwin = 0
nperseg = int(t_ftimwin * data.samplerate)
minSampleNum = nperseg
halfWin = int(nperseg / 2)
# `mtmconvol`: compute no. of samples overlapping across adjacent windows
if overlap < 0: # `toi` is equidistant range or disjoint points
noverlap = nperseg - max(1, int(tSteps[0] * data.samplerate))
elif 0 <= overlap <= 1: # `toi` is percentage
noverlap = min(nperseg - 1, int(overlap * nperseg))
else: # `toi` is "all"
noverlap = nperseg - 1
# `toi` is array
if overlap < 0:
# Compute necessary padding at begin/end of trials to fit sliding windows
offStart = ((toi[0] - tStart) * data.samplerate).astype(np.intp)
padBegin = halfWin - offStart
padBegin = ((padBegin > 0) * padBegin).astype(np.intp)
offEnd = ((tEnd - toi[-1]) * data.samplerate).astype(np.intp)
padEnd = halfWin - offEnd
padEnd = ((padEnd > 0) * padEnd).astype(np.intp)
# Abort if padding was explicitly forbidden
if pad is False and (np.any(padBegin) or np.any(padBegin)):
lgl = "windows within trial bounds"
act = "windows exceeding trials no. " +\
"".join(str(trlno) + ", "\
for trlno in np.array(trialList)[(padBegin + padEnd) > 0])[:-2]
raise SPYValueError(legal=lgl, varname="pad", actual=act)
# Compute sample-indices (one slice/list per trial) from time-selections
soi = []
if not equidistant:
for tk in range(numTrials):
starts = (data.samplerate * (toi - tStart[tk]) -
halfWin).astype(np.intp)
starts += padBegin[tk]
stops = (data.samplerate * (toi - tStart[tk]) + halfWin +
1).astype(np.intp)
stops += padBegin[tk]
stops = np.maximum(stops, stops - starts, dtype=np.intp)
soi.append([
slice(start, stop)
for start, stop in zip(starts, stops)
])
else:
for tk in range(numTrials):
start = int(data.samplerate * (toi[0] - tStart[tk]) -
halfWin)
stop = int(data.samplerate * (toi[-1] - tStart[tk]) +
halfWin + 1)
soi.append(slice(max(0, start), max(stop, stop - start)))
# `toi` is percentage or "all"
else:
padBegin = np.zeros((numTrials, ))
padEnd = np.zeros((numTrials, ))
soi = [slice(None)] * numTrials
# For wavelets, we need to first trim the data (via `preSelect`), then
# extract the wanted time-points (`postSelect`)
if method == "wavelet":
# Simply recycle the indexing work done for `mtmconvol` (i.e., `soi`)
preSelect = []
if not equidistant:
for tk in range(numTrials):
preSelect.append(slice(soi[tk][0].start, soi[tk][-1].stop))
else:
preSelect = soi
# If `toi` is an array, convert "global" indices to "local" ones
# (select within `preSelect`'s selection), otherwise just take all
if overlap < 0:
postSelect = []
for tk in range(numTrials):
smpIdx = np.minimum(
lenTrials[tk] - 1,
data.samplerate * (toi - tStart[tk]) - offStart[tk] +
padBegin[tk])
postSelect.append(smpIdx.astype(np.intp))
else:
postSelect = [slice(None)] * numTrials
# Update `log_dct` w/method-specific options (use `lcls` to get actually
# provided keyword values, not defaults set in here)
if toi is None:
toi = "all"
log_dct["toi"] = lcls["toi"]
# Check options specific to mtm*-methods (particularly tapers and foi/freqs alignment)
if "mtm" in method:
# See if taper choice is supported
if taper not in availableTapers:
lgl = "'" + "or '".join(opt + "' " for opt in availableTapers)
raise SPYValueError(legal=lgl, varname="taper", actual=taper)
taper = getattr(spwin, taper)
# Advanced usage: see if `taperopt` was provided - if not, leave it empty
taperopt = kwargs.get("taperopt", {})
if not isinstance(taperopt, dict):
raise SPYTypeError(taperopt,
varname="taperopt",
expected="dictionary")
# Construct array of maximally attainable frequencies
nFreq = int(np.floor(minSampleNum / 2) + 1)
freqs = np.linspace(0, data.samplerate / 2, nFreq)
# Match desired frequencies as close as possible to actually attainable freqs
if foi is not None:
foi, _ = best_match(freqs, foi, squash_duplicates=True)
elif foilim is not None:
foi, _ = best_match(freqs,
foilim,
span=True,
squash_duplicates=True)
else:
foi = freqs
# Abort if desired frequency selection is empty
if foi.size == 0:
lgl = "non-empty frequency specification"
act = "empty frequency selection"
raise SPYValueError(legal=lgl, varname="foi/foilim", actual=act)
# Set/get `tapsmofrq` if we're working w/Slepian tapers
if taper.__name__ == "dpss":
# Try to derive "sane" settings by using 3/4 octave smoothing of highest `foi`
# following Hipp et al. "Oscillatory Synchronization in Large-Scale
# Cortical Networks Predicts Perception", Neuron, 2011
if tapsmofrq is None:
foimax = foi.max()
tapsmofrq = (foimax * 2**(3 / 4 / 2) -
foimax * 2**(-3 / 4 / 2)) / 2
else:
try:
scalar_parser(tapsmofrq,
varname="tapsmofrq",
lims=[1, np.inf])
except Exception as exc:
raise exc
# Get/compute number of tapers to use (at least 1 and max. 50)
nTaper = taperopt.get("Kmax", 1)
if not taperopt:
nTaper = int(
max(
2,
min(
50,
np.floor(tapsmofrq * minSampleNum * 1 /
data.samplerate))))
taperopt = {"NW": tapsmofrq, "Kmax": nTaper}
else:
nTaper = 1
# Warn the user in case `tapsmofrq` has no effect
if tapsmofrq is not None and taper.__name__ != "dpss":
msg = "`tapsmofrq` is only used if `taper` is `dpss`!"
SPYWarning(msg)
# Update `log_dct` w/method-specific options (use `lcls` to get actually
# provided keyword values, not defaults set in here)
log_dct["taper"] = lcls["taper"]
log_dct["tapsmofrq"] = lcls["tapsmofrq"]
log_dct["nTaper"] = nTaper
# Check for non-default values of options not supported by chosen method
kwdict = {"wav": wav, "width": width}
for name, kwarg in kwdict.items():
if kwarg is not lcls[name]:
msg = "option `{}` has no effect in methods `mtmfft` and `mtmconvol`!"
SPYWarning(msg.format(name))
# Now, prepare explicit compute-classes for chosen method
if method == "mtmfft":
# Check for non-default values of options not supported by chosen method
kwdict = {"t_ftimwin": t_ftimwin, "toi": toi}
for name, kwarg in kwdict.items():
if kwarg is not lcls[name]:
msg = "option `{}` has no effect in method `mtmfft`!"
SPYWarning(msg.format(name))
# Set up compute-class
specestMethod = MultiTaperFFT(samplerate=data.samplerate,
foi=foi,
nTaper=nTaper,
timeAxis=timeAxis,
taper=taper,
taperopt=taperopt,
tapsmofrq=tapsmofrq,
pad=pad,
padtype=padtype,
padlength=padlength,
keeptapers=keeptapers,
polyremoval=polyremoval,
output_fmt=output)
elif method == "mtmconvol":
# Set up compute-class
specestMethod = MultiTaperFFTConvol(soi,
list(padBegin),
list(padEnd),
samplerate=data.samplerate,
noverlap=noverlap,
nperseg=nperseg,
equidistant=equidistant,
toi=toi,
foi=foi,
nTaper=nTaper,
timeAxis=timeAxis,
taper=taper,
taperopt=taperopt,
pad=pad,
padtype=padtype,
padlength=padlength,
prepadlength=prepadlength,
postpadlength=postpadlength,
keeptapers=keeptapers,
polyremoval=polyremoval,
output_fmt=output)
elif method == "wavelet":
# Check for non-default values of `taper`, `tapsmofrq`, `keeptapers` and
# `t_ftimwin` (set to 0 above)
kwdict = {
"taper": taper,
"tapsmofrq": tapsmofrq,
"keeptapers": keeptapers
}
for name, kwarg in kwdict.items():
if kwarg is not lcls[name]:
msg = "option `{}` has no effect in method `wavelet`!"
SPYWarning(msg.format(name))
if t_ftimwin != 0:
msg = "option `t_ftimwin` has no effect in method `wavelet`!"
SPYWarning(msg)
# Check wavelet selection
if wav not in availableWavelets:
lgl = "'" + "or '".join(opt + "' " for opt in availableWavelets)
raise SPYValueError(legal=lgl, varname="wav", actual=wav)
if wav not in ["Morlet", "Paul"]:
msg = "the chosen wavelet '{}' is real-valued and does not provide " +\
"any information about amplitude or phase of the data. This wavelet function " +\
"may be used to isolate peaks or discontinuities in the signal. "
SPYWarning(msg.format(wav))
# Check for consistency of `width`, `order` and `wav`
if wav == "Morlet":
try:
scalar_parser(width, varname="width", lims=[1, np.inf])
except Exception as exc:
raise exc
wfun = getattr(spywave, wav)(w0=width)
else:
if width != lcls["width"]:
msg = "option `width` has no effect for wavelet '{}'"
SPYWarning(msg.format(wav))
if wav == "Paul":
try:
scalar_parser(order,
varname="order",
lims=[4, np.inf],
ntype="int_like")
except Exception as exc:
raise exc
wfun = getattr(spywave, wav)(m=order)
elif wav == "DOG":
try:
scalar_parser(order,
varname="order",
lims=[1, np.inf],
ntype="int_like")
except Exception as exc:
raise exc
wfun = getattr(spywave, wav)(m=order)
else:
if order is not None:
msg = "option `order` has no effect for wavelet '{}'"
SPYWarning(msg.format(wav))
wfun = getattr(spywave, wav)()
# Process frequency selection (`toi` was taken care of above): `foilim`
# selections are wrapped into `foi` thus the seemingly weird if construct
# Note: SLURM workers don't like monkey-patching, so let's pretend
# `get_optimal_wavelet_scales` is a class method by passing `wfun` as its
# first argument
if foi is None:
scales = _get_optimal_wavelet_scales(
wfun, int(minTrialLength * data.samplerate),
1 / data.samplerate)
if foilim is not None:
foi = np.arange(foilim[0], foilim[1] + 1)
if foi is not None:
foi[foi < 0.01] = 0.01
scales = wfun.scale_from_period(1 / foi)
scales = scales[::
-1] # FIXME: this only makes sense if `foi` was sorted -> cf Issue #94
# Update `log_dct` w/method-specific options (use `lcls` to get actually
# provided keyword values, not defaults set in here)
log_dct["wav"] = lcls["wav"]
log_dct["width"] = lcls["width"]
log_dct["order"] = lcls["order"]
# Set up compute-class
specestMethod = WaveletTransform(preSelect,
postSelect,
list(padBegin),
list(padEnd),
samplerate=data.samplerate,
toi=toi,
scales=scales,
timeAxis=timeAxis,
wav=wfun,
polyremoval=polyremoval,
output_fmt=output)
# If provided, make sure output object is appropriate
if out is not None:
try:
data_parser(out,
varname="out",
writable=True,
empty=True,
dataclass="SpectralData",
dimord=SpectralData().dimord)
except Exception as exc:
raise exc
new_out = False
else:
out = SpectralData(dimord=SpectralData._defaultDimord)
new_out = True
# Perform actual computation
specestMethod.initialize(data,
chan_per_worker=kwargs.get("chan_per_worker"),
keeptrials=keeptrials)
specestMethod.compute(data,
out,
parallel=kwargs.get("parallel"),
log_dict=log_dct)
# Either return newly created output object or simply quit
return out if new_out else None
def save(out,
container=None,
tag=None,
filename=None,
overwrite=False,
memuse=100):
r"""Save Syncopy data object to disk
The underlying array data object is stored in a HDF5 file, the metadata in
a JSON file. Both can be placed inside a Syncopy container, which is a
regular directory with the extension '.spy'.
Parameters
----------
out : Syncopy data object
Object to be stored on disk.
container : str
Path to Syncopy container folder (\*.spy) to be used for saving. If
omitted, the extension '.spy' will be added to the folder name.
tag : str
Tag to be appended to container basename
filename : str
Explicit path to data file. This is only necessary if the data should
not be part of a container folder. An extension (\*.<dataclass>) is
added if omitted. The `tag` argument is ignored.
overwrite : bool
If `True` an existing HDF5 file and its accompanying JSON file is
overwritten (without prompt).
memuse : scalar
Approximate in-memory cache size (in MB) for writing data to disk
(only relevant for :class:`syncopy.VirtualData` or memory map data sources)
Returns
-------
Nothing : None
Notes
------
Syncopy objects may also be saved using the class method ``.save`` that
acts as a wrapper for :func:`syncopy.save`, e.g.,
>>> save(obj, container="new_spy_container")
is equivalent to
>>> obj.save(container="new_spy_container")
However, once a Syncopy object has been saved, the class method ``.save``
can be used as a shortcut to quick-save recent changes, e.g.,
>>> obj.save()
writes the current state of `obj` to the data/meta-data files on-disk
associated with `obj` (overwriting both in the process). Similarly,
>>> obj.save(tag='newtag')
saves `obj` in the current container 'new_spy_container' under a different
tag.
Examples
--------
Save the Syncopy data object `obj` on disk in the current working directory
without creating a spy-container
>>> spy.save(obj, filename="session1")
>>> # --> os.getcwd()/session1.<dataclass>
>>> # --> os.getcwd()/session1.<dataclass>.info
Save `obj` without creating a spy-container using an absolute path
>>> spy.save(obj, filename="/tmp/session1")
>>> # --> /tmp/session1.<dataclass>
>>> # --> /tmp/session1.<dataclass>.info
Save `obj` in a new spy-container created in the current working directory
>>> spy.save(obj, container="container.spy")
>>> # --> os.getcwd()/container.spy/container.<dataclass>
>>> # --> os.getcwd()/container.spy/container.<dataclass>.info
Save `obj` in a new spy-container created by providing an absolute path
>>> spy.save(obj, container="/tmp/container.spy")
>>> # --> /tmp/container.spy/container.<dataclass>
>>> # --> /tmp/container.spy/container.<dataclass>.info
Save `obj` in a new (or existing) spy-container under a different tag
>>> spy.save(obj, container="session1.spy", tag="someTag")
>>> # --> os.getcwd()/session1.spy/session1_someTag.<dataclass>
>>> # --> os.getcwd()/session1.spy/session1_someTag.<dataclass>.info
See also
--------
syncopy.load : load data created with :func:`syncopy.save`
"""
# Make sure `out` is a valid Syncopy data object
data_parser(out, varname="out", writable=None, empty=False)
if filename is None and container is None:
raise SPYError('filename and container cannot both be `None`')
if container is not None and filename is None:
# construct filename from container name
if not isinstance(container, str):
raise SPYTypeError(container, varname="container", expected="str")
if not os.path.splitext(container)[1] == ".spy":
container += ".spy"
fileInfo = filename_parser(container)
filename = os.path.join(fileInfo["folder"], fileInfo["container"],
fileInfo["basename"])
# handle tag
if tag is not None:
if not isinstance(tag, str):
raise SPYTypeError(tag, varname="tag", expected="str")
filename += '_' + tag
elif container is not None and filename is not None:
raise SPYError(
"container and filename cannot be used at the same time")
if not isinstance(filename, str):
raise SPYTypeError(filename, varname="filename", expected="str")
# add extension if not part of the filename
if "." not in os.path.splitext(filename)[1]:
filename += out._classname_to_extension()
try:
scalar_parser(memuse, varname="memuse", lims=[0, np.inf])
except Exception as exc:
raise exc
if not isinstance(overwrite, bool):
raise SPYTypeError(overwrite, varname="overwrite", expected="bool")
# Parse filename for validity and construct full path to HDF5 file
fileInfo = filename_parser(filename)
if fileInfo["extension"] != out._classname_to_extension():
raise SPYError("""Extension in filename ({ext}) does not match data
class ({dclass})""".format(ext=fileInfo["extension"],
dclass=out.__class__.__name__))
dataFile = os.path.join(fileInfo["folder"], fileInfo["filename"])
# If `out` is to replace its own on-disk representation, be more careful
if overwrite and dataFile == out.filename:
replace = True
else:
replace = False
# Prevent `out` from trying to re-create its own data file
if replace:
out.data.flush()
h5f = out.data.file
dat = out.data
trl = h5f["trialdefinition"]
else:
if not os.path.exists(fileInfo["folder"]):
try:
os.makedirs(fileInfo["folder"])
except IOError:
raise SPYIOError(fileInfo["folder"])
except Exception as exc:
raise exc
else:
if os.path.exists(dataFile):
if not os.path.isfile(dataFile):
raise SPYIOError(dataFile)
if overwrite:
try:
h5f = h5py.File(dataFile, mode="w")
h5f.close()
except Exception as exc:
msg = "Cannot overwrite {} - file may still be open. "
msg += "Original error message below\n{}"
raise SPYError(msg.format(dataFile, str(exc)))
else:
raise SPYIOError(dataFile, exists=True)
h5f = h5py.File(dataFile, mode="w")
# Save each member of `_hdfFileDatasetProperties` in target HDF file
for datasetName in out._hdfFileDatasetProperties:
dataset = getattr(out, datasetName)
# Member is a memory map
if isinstance(dataset, np.memmap):
# Given memory cap, compute how many data blocks can be grabbed
# per swipe (divide by 2 since we're working with an add'l tmp array)
memuse *= 1024**2 / 2
nrow = int(
memuse /
(np.prod(dataset.shape[1:]) * dataset.dtype.itemsize))
rem = int(dataset.shape[0] % nrow)
n_blocks = [nrow] * int(
dataset.shape[0] // nrow) + [rem] * int(rem > 0)
# Write data block-wise to dataset (use `clear` to wipe blocks of
# mem-maps from memory)
dat = h5f.create_dataset(datasetName,
dtype=dataset.dtype,
shape=dataset.shape)
for m, M in enumerate(n_blocks):
dat[m * nrow:m * nrow +
M, :] = out.data[m * nrow:m * nrow + M, :]
out.clear()
# Member is a HDF5 dataset
else:
dat = h5f.create_dataset(datasetName, data=dataset)
# Now write trial-related information
trl_arr = np.array(out.trialdefinition)
if replace:
trl[()] = trl_arr
trl.flush()
else:
trl = h5f.create_dataset("trialdefinition",
data=trl_arr,
maxshape=(None, trl_arr.shape[1]))
# Write to log already here so that the entry can be exported to json
infoFile = dataFile + FILE_EXT["info"]
out.log = "Wrote files " + dataFile + "\n\t\t\t" + 2 * " " + infoFile
# While we're at it, write cfg entries
out.cfg = {
"method": sys._getframe().f_code.co_name,
"files": [dataFile, infoFile]
}
# Assemble dict for JSON output: order things by their "readability"
outDict = OrderedDict(startInfoDict)
outDict["filename"] = fileInfo["filename"]
outDict["dataclass"] = out.__class__.__name__
outDict["data_dtype"] = dat.dtype.name
outDict["data_shape"] = dat.shape
outDict["data_offset"] = dat.id.get_offset()
outDict["trl_dtype"] = trl.dtype.name
outDict["trl_shape"] = trl.shape
outDict["trl_offset"] = trl.id.get_offset()
if isinstance(out.data, np.ndarray):
if np.isfortran(out.data):
outDict["order"] = "F"
else:
outDict["order"] = "C"
for key in out._infoFileProperties:
value = getattr(out, key)
if isinstance(value, np.ndarray):
value = value.tolist()
# potentially nested dicts
elif isinstance(value, dict):
value = dict(value)
_dict_converter(value)
outDict[key] = value
# Save relevant stuff as HDF5 attributes
for key in out._hdfFileAttributeProperties:
if outDict[key] is None:
h5f.attrs[key] = "None"
else:
try:
h5f.attrs[key] = outDict[key]
except RuntimeError:
msg = "Too many entries in `{}` - truncating HDF5 attribute. " +\
"Please refer to {} for complete listing."
info_fle = os.path.split(
os.path.split(filename.format(ext=FILE_EXT["info"]))[0])[1]
info_fle = os.path.join(
info_fle,
os.path.basename(filename.format(ext=FILE_EXT["info"])))
SPYWarning(msg.format(key, info_fle))
h5f.attrs[key] = [outDict[key][0], "...", outDict[key][-1]]
# Re-assign filename after saving (and remove source in case it came from `__storage__`)
if not replace:
h5f.close()
if __storage__ in out.filename:
out.data.file.close()
os.unlink(out.filename)
out.data = dataFile
# Compute checksum and finally write JSON (automatically overwrites existing)
outDict["file_checksum"] = hash_file(dataFile)
with open(infoFile, 'w') as out_json:
json.dump(outDict, out_json, indent=4)
return
def _prep_spectral_plots(self, name, **inputArgs):
"""
Local helper that performs sanity checks and sets up data selection
Parameters
----------
self : :class:`~syncopy.SpectralData` object
Syncopy :class:`~syncopy.SpectralData` object that is being processed by
the respective :meth:`.singlepanelplot` or :meth:`.multipanelplot` class methods
defined in this module.
name : str
Name of caller (i.e., "singlepanelplot" or "multipanelplot")
inputArgs : dict
Input arguments of caller (i.e., :meth:`.singlepanelplot` or :meth:`.multipanelplot`)
collected in dictionary
Returns
-------
dimArrs : tuple
Four-element tuple containing (in this order): `trList`, list of (selected)
trials to visualize, `chArr`, 1D :class:`numpy.ndarray` of channel specifiers
based on provided user selection, `freqArr`, 1D :class:`numpy.ndarray` of
frequency specifiers based on provided user selection, `tpArr`,
1D :class:`numpy.ndarray` of taper specifiers based on provided user selection.
Note that `"all"` and `None` selections are converted to arrays ready for
indexing.
dimCounts : tuple
Four-element tuple holding sizes of corresponding selection arrays comprised
in `dimArrs`. Elements are (in this order): number of (selected) trials
`nTrials`, number of (selected) channels `nChan`, number of (selected)
frequencies `nFreq`, number of (selected) tapers `nTap`.
isTimeFrequency : bool
If `True`, input object contains time-frequency data, `False` otherwise
complexConversion : callable
Lambda function that performs complex-to-float conversion of Fourier
coefficients (if necessary).
pltDtype : str or :class:`numpy.dtype`
Numeric type of (potentially converted) complex Fourier coefficients.
dataLbl : str
Caption for y-axis or colorbar (depending on value of `isTimeFrequency`).
Notes
-----
This is an auxiliary method that is intended purely for internal use. Please
refer to the user-exposed methods :func:`~syncopy.singlepanelplot` and/or
:func:`~syncopy.multipanelplot` to actually generate plots of Syncopy data objects.
See also
--------
:meth:`syncopy.plotting.spy_plotting._prep_plots` : General basic input parsing for all Syncopy plotting routines
"""
# Basic sanity checks for all plotting routines w/any Syncopy object
_prep_plots(self, name, **inputArgs)
# Ensure our binary flags are actually binary
if not isinstance(inputArgs["avg_channels"], bool):
raise SPYTypeError(inputArgs["avg_channels"],
varname="avg_channels",
expected="bool")
if not isinstance(inputArgs["avg_tapers"], bool):
raise SPYTypeError(inputArgs["avg_tapers"],
varname="avg_tapers",
expected="bool")
if not isinstance(inputArgs.get("avg_trials", True), bool):
raise SPYTypeError(inputArgs["avg_trials"],
varname="avg_trials",
expected="bool")
# Pass provided selections on to `Selector` class which performs error
# checking and generates required indexing arrays
self._selection = {
"trials": inputArgs["trials"],
"channels": inputArgs["channels"],
"tapers": inputArgs["tapers"],
"toilim": inputArgs["toilim"],
"foilim": inputArgs["foilim"]
}
# Ensure any optional keywords controlling plotting appearance make sense
if inputArgs["title"] is not None:
if not isinstance(inputArgs["title"], str):
raise SPYTypeError(inputArgs["title"],
varname="title",
expected="str")
if inputArgs["grid"] is not None:
if not isinstance(inputArgs["grid"], bool):
raise SPYTypeError(inputArgs["grid"],
varname="grid",
expected="bool")
# Get trial/channel/taper count and collect quantities in tuple
trList = self._selection.trials
nTrials = len(trList)
chArr = self.channel[self._selection.channel]
nChan = chArr.size
freqArr = self.freq[self._selection.freq]
nFreq = freqArr.size
tpArr = np.arange(self.taper.size)[self._selection.taper]
nTap = tpArr.size
dimCounts = (nTrials, nChan, nFreq, nTap)
dimArrs = (trList, chArr, freqArr, tpArr)
# Determine whether we're dealing w/tf data
isTimeFrequency = False
if any([t.size > 1 for t in self.time]):
isTimeFrequency = True
# Ensure provided min/max range for plotting TF data makes sense
vminmax = False
if inputArgs.get("vmin", None) is not None:
try:
scalar_parser(inputArgs["vmin"], varname="vmin")
except Exception as exc:
raise exc
vminmax = True
if inputArgs.get("vmax", None) is not None:
try:
scalar_parser(inputArgs["vmax"], varname="vmax")
except Exception as exc:
raise exc
vminmax = True
if inputArgs.get("vmin", None) and inputArgs.get("vmax", None):
if inputArgs["vmin"] >= inputArgs["vmax"]:
lgl = "minimal data range bound to be less than provided maximum "
act = "vmax < vmin"
raise SPYValueError(legal=lgl, varname="vmin/vamx", actual=act)
if vminmax and not isTimeFrequency:
msg = "`vmin` and `vmax` is only used for time-frequency visualizations"
SPYWarning(msg)
# Check for complex entries in data and set datatype for plotting arrays
# constructed below (always use floats w/same precision as data)
if "complex" in self.data.dtype.name:
msg = "Found complex Fourier coefficients - visualization will use absolute values."
SPYWarning(msg)
complexConversion = lambda x: np.absolute(x).real
pltDtype = "f{}".format(self.data.dtype.itemsize)
dataLbl = "Absolute Frequency [dB]"
else:
complexConversion = lambda x: x
pltDtype = self.data.dtype
dataLbl = "Power [dB]"
return dimArrs, dimCounts, isTimeFrequency, complexConversion, pltDtype, dataLbl
def padding(data,
padtype,
pad="absolute",
padlength=None,
prepadlength=None,
postpadlength=None,
unit="samples",
create_new=True):
"""
Perform data padding on Syncopy object or :class:`numpy.ndarray`
**Usage Summary**
Depending on the value of `pad` the following padding length specifications
are supported:
+------------+----------------------+---------------+----------------------+----------------------+
| `pad` | `data` | `padlength` | `prepadlength` | `postpadlength` |
+============+======================+===============+======================+======================+
| 'absolute' | Syncopy object/array | number | `None`/`bool` | `None`/`bool` |
+------------+----------------------+---------------+----------------------+----------------------+
| 'relative' | Syncopy object/array | number/`None` | number/`None`/`bool` | number/`None`/`bool` |
+------------+----------------------+---------------+----------------------+----------------------+
| 'maxlen' | Syncopy object | `None`/`bool` | `None`/`bool` | `None`/`bool` |
+------------+----------------------+---------------+----------------------+----------------------+
| 'nextpow2' | Syncopy object/array | `None`/`bool` | `None`/`bool` | `None`/`bool` |
+------------+----------------------+---------------+----------------------+----------------------+
* `data` can be either a Syncopy object containing multiple trials or a
:class:`numpy.ndarray` representing a single trial
* (pre/post)padlength: can be either `None`, `True`/`False` or a positive
number: if `True` indicates where to pad, e.g., by using ``pad =
'maxlen'`` and ``prepadlength = True``, `data` is padded at the beginning
of each trial. **Only** if `pad` is 'relative' are scalar values supported
for `prepadlength` and `postpadlength`
* ``pad = 'absolute'``: pad to desired absolute length, e.g., by using ``pad
= 5`` and ``unit = 'time'`` all trials are (if necessary) padded to 5s
length. Here, `padlength` **has** to be provided, `prepadlength` and
`postpadlength` can be `None` or `True`/`False`
* ``pad = 'relative'``: pad by provided `padlength`, e.g., by using
``padlength = 20`` and ``unit = 'samples'``, 20 samples are padded
symmetrically around (before and after) each trial. Use ``padlength = 20``
and ``prepadlength = True`` **or** directly ``prepadlength = 20`` to pad
before each trial. Here, at least one of `padlength`, `prepadlength` or
`postpadlength` **has** to be provided.
* ``pad = 'maxlen'``: (only valid for **Syncopy objects**) pad up to maximal
trial length found in `data`. All lengths have to be either Boolean
indicating padding location or `None` (if all are `None`, symmetric
padding is performed)
* ``pad = 'nextpow2'``: pad each trial up to closest power of two. All
lengths have to be either Boolean indicating padding location or `None`
(if all are `None`, symmetric padding is performed)
Full documentation below.
Parameters
----------
data : Syncopy object or :class:`numpy.ndarray`
Non-empty Syncopy data object or array representing numeric data to be
padded. **NOTE**: if `data` is a :class:`numpy.ndarray`, it is assumed
that it represents recordings from only a single trial, where its first
axis corresponds to time. In other words, `data` is a
'time'-by-'channel' array such that its rows reflect samples and its
columns represent channels. If `data` is a Syncopy object, trial
information and dimensional order are fetched from `data.trials` and
`data.dimord`, respectively.
padtype : str
Padding value(s) to be used. Available options are:
* 'zero' : pad using zeros
* 'nan' : pad using `np.nan`'s
* 'mean' : pad with by-channel mean value across each trial
* 'localmean' : pad with by-channel mean value using only `padlength` or
`prepadlength`/`postpadlength` number of boundary-entries for averaging
* 'edge' : pad with trial-boundary values
* 'mirror' : pad with reflections of trial-boundary values
pad : str
Padding mode to be used. Available options are:
* 'absolute' : pad each trial to achieve a desired absolute length such
that all trials have identical length post padding. If `pad` is `absolute`
a `padlength` **has** to be provided, `prepadlength` and `postpadlength`
may be `True` or `False`, respectively (see Examples for details).
* 'relative' : pad each trial by provided `padlength` such that all trials
are extended by the same amount regardless of their original lengths.
If `pad` is `relative`, `prepadlength` and `postpadlength` can either
be specified directly (using numerical values) or implicitly by only
providing `padlength` and setting `prepadlength` and `postpadlength`
to `True` or `False`, respectively (see Examples for details). If `pad`
is `relative` at least one of `padlength`, `prepadlength` or `postpadlength`
**has** to be provided.
* 'maxlen' : only usable if `data` is a Syncopy object. If `pad` is
'maxlen' all trials are padded to achieve the length of the longest
trial in `data`, i.e., post padding, all trials have the same length,
which equals the size of the longest trial pre-padding. For
``pad = 'maxlen'``, `padlength`, `prepadlength` as well as `postpadlength`
have to be either Boolean or `None` indicating the preferred padding
location (pre-trial, post-trial or symmetrically pre- and post-trial).
If all are `None`, symmetric padding is performed (see Examples for
details).
* 'nextpow2' : pad each trial to achieve a length equals the closest power
of two of its original length. For ``pad = 'nextpow2'``, `padlength`,
`prepadlength` as well as `postpadlength` have to be either Boolean
or `None` indicating the preferred padding location (pre-trial, post-trial
or symmetrically pre- and post-trial). If all are `None`, symmetric
padding is performed (see Examples for details).
padlength : None, bool or positive scalar
Length to be padded to `data` (if `padlength` is scalar-valued) or
padding location (if `padlength` is Boolean). Depending on the value of
`pad`, `padlength` can be used to pre-pend (if `padlength` is a positive
number and `prepadlength` is `True`) or append trials (if `padlength` is
a positive number and `postpadlength` is `True`). If neither
`prepadlength` nor `postpadlength` are specified (i.e, both are `None`),
symmetric pre- and post-trial padding is performed (i.e., ``0.5 * padlength``
before and after each trial - note that odd sample counts are rounded downward
to the nearest even integer). If ``unit = 'time'``, `padlength` is assumed
to be given in seconds, otherwise (``unit = 'samples'``), `padlength` is
interpreted as sample-count. Note that only ``pad = 'relative'`` and
``pad = 'absolute'`` support numeric values of `padlength`.
prepadlength : None, bool or positive scalar
Length to be pre-pended before each trial (if `prepadlength` is
scalar-valued) or pre-padding flag (if `prepadlength` is `True`). If
`prepadlength` is `True`, pre-padding length is either directly inferred
from `padlength` or implicitly derived from chosen padding mode defined
by `pad`. If ``unit = 'time'``, `prepadlength` is assumed to be given in
seconds, otherwise (``unit = 'samples'``), `prepadlength` is interpreted
as sample-count. Note that only ``pad = 'relative'`` supports numeric
values of `prepadlength`.
postpadlength : None, bool or positive scalar
Length to be appended after each trial (if `postpadlength` is
scalar-valued) or post-padding flag (if `postpadlength` is `True`). If
`postpadlength` is `True`, post-padding length is either directly inferred
from `padlength` or implicitly derived from chosen padding mode defined
by `pad`. If ``unit = 'time'``, `postpadlength` is assumed to be given in
seconds, otherwise (``unit = 'samples'``), `postpadlength` is interpreted
as sample-count. Note that only ``pad = 'relative'`` supports numeric
values of `postpadlength`.
unit : str
Unit of numerical values given by `padlength` and/or `prepadlength`
and/or `postpadlength`. If ``unit = 'time'``, `padlength`,
`prepadlength`, and `postpadlength` are assumed to be given in seconds,
otherwise (``unit = 'samples'``), `padlength`, `prepadlength`, and
`postpadlength` are interpreted as sample-counts. **Note** Providing
padding lengths in seconds (i.e., ``unit = 'time'``) is only supported
if `data` is a Syncopy object.
create_new : bool
If `True`, a padded copy of the same type as `data` is returned (a
:class:`numpy.ndarray` or Syncopy object). If `create_new` is `False`,
either a single dictionary (if `data` is a :class:`numpy.ndarray`) or a
``len(data.trials)``-long list of dictionaries (if `data` is a Syncopy
object) with all necessary options for performing the actual padding
operation with :func:`numpy.pad` is returned.
Returns
-------
pad_dict : dict, if `data` is a :class:`numpy.ndarray` and ``create_new = False``
Dictionary whose items contain all necessary parameters for calling
:func:`numpy.pad` to perform the desired padding operation on `data`.
pad_dicts : list, if `data` is a Syncopy object and ``create_new = False``
List of dictionaries for calling :func:`numpy.pad` to perform the
desired padding operation on all trials found in `data`.
out : :class:`numpy.ndarray`, if `data` is a :class:`numpy.ndarray` and ``create_new = True``
Padded version (deep copy) of `data`
out : Syncopy object, if `data` is a Syncopy object and ``create_new = True``
Padded version (deep copy) of `data`
Notes
-----
This method emulates (and extends) FieldTrip's `ft_preproc_padding` by
providing a convenience wrapper for NumPy's :func:`numpy.pad` that performs
the actual heavy lifting.
Examples
--------
Consider the following small array representing a toy-problem-trial of `ns`
samples across `nc` channels:
>>> nc = 7; ns = 30
>>> trl = np.random.randn(ns, nc)
We start by padding a total of 10 zeros symmetrically to `trl`
>>> padded = spy.padding(trl, 'zero', pad='relative', padlength=10)
>>> padded[:6, :]
array([[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[-1.0866, 2.3358, 0.8758, 0.5196, 0.8049, -0.659 , -0.9173]])
>>> padded[-6:, :]
array([[ 0.027 , 1.8069, 1.5249, -0.7953, -0.8933, 1.0202, -0.6862],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ],
[ 0. , 0. , 0. , 0. , 0. , 0. , 0. ]])
>>> padded.shape
(40, 7)
Note that the above call is equivalent to
>>> padded_ident = spy.padding(trl, 'zero', pad='relative', padlength=10, prepadlength=True, postpadlength=True)
>>> np.array_equal(padded_ident, padded)
True
>>> padded_ident = spy.padding(trl, 'zero', pad='relative', prepadlength=5, postpadlength=5)
>>> np.array_equal(padded_ident, padded)
True
Similarly,
>>> prepad = spy.padding(trl, 'nan', pad='relative', prepadlength=10)
is the same as
>>> prepad_ident = spy.padding(trl, 'nan', pad='relative', padlength=10, prepadlength=True)
>>> np.allclose(prepad, prepad_ident, equal_nan=True)
True
Define bogus trials on `trl` and create a dummy object with unit samplerate
>>> tdf = np.vstack([np.arange(0, ns, 5),
np.arange(5, ns + 5, 5),
np.ones((int(ns / 5), )),
np.ones((int(ns / 5), )) * np.pi]).T
>>> adata = spy.AnalogData(trl, trialdefinition=tdf, samplerate=1)
Pad each trial to the closest power of two by appending by-trial channel
averages. However, do not perform actual padding, but only prepare dictionaries
of parameters to be passed on to :func:`numpy.pad`
>>> pad_dicts = spy.padding(adata, 'mean', pad='nextpow2', postpadlength=True, create_new=False)
>>> len(pad_dicts) == len(adata.trials)
True
>>> pad_dicts[0]
{'pad_width': array([[0, 3],
[0, 0]]), 'mode': 'mean'}
Similarly, the following call generates a list of dictionaries preparing
absolute padding by prepending zeros with :func:`numpy.pad`
>>> pad_dicts = spy.padding(adata, 'zero', pad='absolute', padlength=10, prepadlength=True, create_new=False)
>>> pad_dicts[0]
{'pad_width': array([[5, 0],
[0, 0]]), 'mode': 'constant', 'constant_values': 0}
See also
--------
numpy.pad : fast array padding in NumPy
"""
# Detect whether input is data object or array-like
if any(["BaseData" in str(base) for base in data.__class__.__mro__]):
try:
data_parser(data,
varname="data",
dataclass="AnalogData",
empty=False)
except Exception as exc:
raise exc
timeAxis = data.dimord.index("time")
spydata = True
elif data.__class__.__name__ == "FauxTrial":
if len(data.shape) != 2:
lgl = "two-dimensional AnalogData trial segment"
act = "{}-dimensional trial segment"
raise SPYValueError(legal=lgl,
varname="data",
actual=act.format(len(data.shape)))
timeAxis = data.dimord.index("time")
spydata = False
else:
try:
array_parser(data, varname="data", dims=2)
except Exception as exc:
raise exc
timeAxis = 0
spydata = False
# FIXME: Creation of new spy-object currently not supported
if not isinstance(create_new, bool):
raise SPYTypeError(create_new, varname="create_new", expected="bool")
if spydata and create_new:
raise NotImplementedError(
"Creation of padded spy objects currently not supported. ")
# Use FT-compatible options (sans FT option 'remove')
if not isinstance(padtype, str):
raise SPYTypeError(padtype, varname="padtype", expected="string")
options = ["zero", "nan", "mean", "localmean", "edge", "mirror"]
if padtype not in options:
lgl = "'" + "or '".join(opt + "' " for opt in options)
raise SPYValueError(legal=lgl, varname="padtype", actual=padtype)
# Check `pad` and ensure we can actually perform the requested operation
if not isinstance(pad, str):
raise SPYTypeError(pad, varname="pad", expected="string")
options = ["absolute", "relative", "maxlen", "nextpow2"]
if pad not in options:
lgl = "'" + "or '".join(opt + "' " for opt in options)
raise SPYValueError(legal=lgl, varname="pad", actual=pad)
if pad == "maxlen" and not spydata:
lgl = "syncopy data object when using option 'maxlen'"
raise SPYValueError(legal=lgl, varname="pad", actual="maxlen")
# Make sure a data object was provided if we're working with time values
if not isinstance(unit, str):
raise SPYTypeError(unit, varname="unit", expected="string")
options = ["samples", "time"]
if unit not in options:
lgl = "'" + "or '".join(opt + "' " for opt in options)
raise SPYValueError(legal=lgl, varname="unit", actual=unit)
if unit == "time" and not spydata:
raise SPYValueError(
legal="syncopy data object when using option 'time'",
varname="unit",
actual="time")
# Set up dictionary for type-checking of provided padding lengths
nt_dict = {"samples": "int_like", "time": None}
# If we're padding up to an absolute bound or the max. length across
# trials, compute lower bound for padding (in samples or seconds)
if pad in ["absolute", "maxlen"]:
if spydata:
maxTrialLen = np.diff(data.sampleinfo).max()
else:
maxTrialLen = data.shape[
timeAxis] # if `pad="absolute" and data is array
else:
maxTrialLen = np.inf
if unit == "time":
padlim = maxTrialLen / data.samplerate
else:
padlim = maxTrialLen
# To ease option processing, collect padding length keywords in dict
plengths = {
"padlength": padlength,
"prepadlength": prepadlength,
"postpadlength": postpadlength
}
# In case of relative padding, we need at least one scalar value to proceed
if pad == "relative":
# If `padlength = None`, pre- or post- need to be set; if `padlength`
# is set, both pre- and post- need to be `None` or `True`/`False`.
# After this code block, pre- and post- are guaranteed to be numeric.
if padlength is None:
for key in ["prepadlength", "postpadlength"]:
if plengths[key] is not None:
try:
scalar_parser(plengths[key],
varname=key,
ntype=nt_dict[unit],
lims=[0, np.inf])
except Exception as exc:
raise exc
else:
plengths[key] = 0
else:
try:
scalar_parser(padlength,
varname="padlength",
ntype=nt_dict[unit],
lims=[0, np.inf])
except Exception as exc:
raise exc
for key in ["prepadlength", "postpadlength"]:
if not isinstance(plengths[key], (bool, type(None))):
raise SPYTypeError(plengths[key],
varname=key,
expected="bool or None")
if prepadlength is None and postpadlength is None:
prepadlength = True
postpadlength = True
else:
prepadlength = prepadlength is not None
postpadlength = postpadlength is not None
if prepadlength and postpadlength:
plengths["prepadlength"] = padlength / 2
plengths["postpadlength"] = padlength / 2
else:
plengths["prepadlength"] = prepadlength * padlength
plengths["postpadlength"] = postpadlength * padlength
# Under-determined: abort if requested padding length is 0
if all(value == 0 for value in plengths.values() if value is not None):
lgl = "either non-zero value of `padlength` or `prepadlength` " + \
"and/or `postpadlength` to be set"
raise SPYValueError(legal=lgl,
varname="padlength",
actual="0|None")
else:
# For absolute padding, the desired length has to be >= max. trial length
if pad == "absolute":
try:
scalar_parser(padlength,
varname="padlength",
ntype=nt_dict[unit],
lims=[padlim, np.inf])
except Exception as exc:
raise exc
for key in ["prepadlength", "postpadlength"]:
if not isinstance(plengths[key], (bool, type(None))):
raise SPYTypeError(plengths[key],
varname=key,
expected="bool or None")
# For `maxlen` or `nextpow2` we don't want any numeric entries at all
else:
for key, value in plengths.items():
if not isinstance(value, (bool, type(None))):
raise SPYTypeError(value,
varname=key,
expected="bool or None")
# Warn of potential conflicts
if padlength and (prepadlength or postpadlength):
msg = "Found `padlength` and `prepadlength` and/or " +\
"`postpadlength`. Symmetric padding is performed. "
SPYWarning(msg)
# If both pre-/post- are `None`, set them to `True` to use symmetric
# padding, otherwise convert `None` entries to `False`
if prepadlength is None and postpadlength is None:
plengths["prepadlength"] = True
plengths["postpadlength"] = True
else:
plengths["prepadlength"] = plengths["prepadlength"] is not None
plengths["postpadlength"] = plengths["postpadlength"] is not None
# Update pre-/post-padding and (if required) convert time to samples
prepadlength = plengths["prepadlength"]
postpadlength = plengths["postpadlength"]
if unit == "time":
if pad == "relative":
prepadlength = int(prepadlength * data.samplerate)
postpadlength = int(postpadlength * data.samplerate)
elif pad == "absolute":
padlength = int(padlength * data.samplerate)
# Construct dict of keywords for ``np.pad`` depending on chosen `padtype`
kws = {
"zero": {
"mode": "constant",
"constant_values": 0
},
"nan": {
"mode": "constant",
"constant_values": np.nan
},
"localmean": {
"mode": "mean",
"stat_length": -1
},
"mean": {
"mode": "mean"
},
"edge": {
"mode": "edge"
},
"mirror": {
"mode": "reflect"
}
}
# If in put was syncopy data object, padding is done on a per-trial basis
if spydata:
# A list of input keywords for ``np.pad`` is constructed, no matter if
# we actually want to build a new object or not
pad_opts = []
for trl in data.trials:
nSamples = trl.shape[timeAxis]
if pad == "absolute":
padding = (padlength - nSamples) / (prepadlength +
postpadlength)
elif pad == "relative":
padding = True
elif pad == "maxlen":
padding = (maxTrialLen - nSamples) / (prepadlength +
postpadlength)
elif pad == "nextpow2":
padding = (_nextpow2(nSamples) - nSamples) / (prepadlength +
postpadlength)
pw = np.zeros((2, 2), dtype=int)
pw[timeAxis, :] = [prepadlength * padding, postpadlength * padding]
pad_opts.append(dict({"pad_width": pw}, **kws[padtype]))
if padtype == "localmean":
pad_opts[-1]["stat_length"] = pw[timeAxis, :]
if create_new:
pass
else:
return pad_opts
# Input was a array/FauxTrial (i.e., single trial) - we have to do the padding just once
else:
nSamples = data.shape[timeAxis]
if pad == "absolute":
padding = (padlength - nSamples) / (prepadlength + postpadlength)
elif pad == "relative":
padding = True
elif pad == "nextpow2":
padding = (_nextpow2(nSamples) - nSamples) / (prepadlength +
postpadlength)
pw = np.zeros((2, 2), dtype=int)
pw[timeAxis, :] = [prepadlength * padding, postpadlength * padding]
pad_opts = dict({"pad_width": pw}, **kws[padtype])
if padtype == "localmean":
pad_opts["stat_length"] = pw[timeAxis, :]
if create_new:
if isinstance(data, np.ndarray):
return np.pad(data, **pad_opts)
else: # FIXME: currently only supports FauxTrial
shp = list(data.shape)
shp[timeAxis] += pw[timeAxis, :].sum()
idx = list(data.idx)
if isinstance(idx[timeAxis], slice):
idx[timeAxis] = slice(idx[timeAxis].start,
idx[timeAxis].start + shp[timeAxis])
else:
idx[timeAxis] = pw[timeAxis, 0] * [idx[timeAxis][0]] + idx[timeAxis] \
+ pw[timeAxis, 1] * [idx[timeAxis][-1]]
return data.__class__(shp, idx, data.dtype, data.dimord)
else:
return pad_opts
def test_none(self):
with pytest.raises(SPYTypeError):
scalar_parser(None,
varname="value",
ntype="int_like",
lims=[10, 1000])
def test_complex_invalid(self):
value = complex(2, -1)
with pytest.raises(SPYValueError):
scalar_parser(value, lims=[-3, 1])
def test_complex_valid(self):
value = complex(2, -1)
scalar_parser(value, lims=[-3, 5]) # valid
