"""Process data from one form into another.
For example, taking spike times and binning them into non-overlapping bins and convolving spike
times with a gaussian kernel.
"""
import numpy as np
import pandas as pd
from scipy import interpolate, sparse
from brainbox import core
from iblutil.numerical import bincount2D
from iblutil.util import Bunch
import logging
_logger = logging.getLogger(__name__)
[docs]
def sync(dt, times=None, values=None, timeseries=None, offsets=None, interp='zero',
fillval=np.nan):
"""
Function for resampling a single or multiple time series to a single, evenly-spaced, delta t
between observations. Uses interpolation to find values.
Can be used on raw numpy arrays of timestamps and values using the 'times' and 'values' kwargs
and/or on brainbox.core.TimeSeries objects passed to the 'timeseries' kwarg. If passing both
TimeSeries objects and numpy arrays, the offsets passed should be for the TS objects first and
then the numpy arrays.
Uses scipy's interpolation library to perform interpolation.
See scipy.interp1d for more information regarding interp and fillval parameters.
:param dt: Separation of points which the output timeseries will be sampled at
:type dt: float
:param timeseries: A group of time series to perform alignment or a single time series.
Must have time stamps.
:type timeseries: tuple of TimeSeries objects, or a single TimeSeries object.
:param times: time stamps for the observations in 'values']
:type times: np.ndarray or list of np.ndarrays
:param values: observations corresponding to the timestamps in 'times'
:type values: np.ndarray or list of np.ndarrays
:param offsets: tuple of offsets for time stamps of each time series. Offsets for passed
TimeSeries objects first, then offsets for passed numpy arrays. defaults to None
:type offsets: tuple of floats, optional
:param interp: Type of interpolation to use. Refer to scipy.interpolate.interp1d for possible
values, defaults to np.nan
:type interp: str
:param fillval: Fill values to use when interpolating outside of range of data. See interp1d
for possible values, defaults to np.nan
:return: TimeSeries object with each row representing synchronized values of all
input TimeSeries. Will carry column names from input time series if all of them have column
names.
"""
#########################################
# Checks on inputs and input processing #
#########################################
# Initialize a list to contain times/values pairs if no TS objs are passed
if timeseries is None:
timeseries = []
# If a single time series is passed for resampling, wrap it in an iterable
elif isinstance(timeseries, core.TimeSeries):
timeseries = [timeseries]
# Yell at the user if they try to pass stuff to timeseries that isn't a TimeSeries object
elif not all([isinstance(ts, core.TimeSeries) for ts in timeseries]):
raise TypeError('All elements of \'timeseries\' argument must be brainbox.core.TimeSeries '
'objects. Please uses \'times\' and \'values\' for np.ndarray args.')
# Check that if something is passed to times or values, there is a corresponding equal-length
# argument for the other element.
if (times is not None) or (values is not None):
if len(times) != len(values):
raise ValueError('\'times\' and \'values\' must have the same number of elements.')
if type(times[0]) is np.ndarray:
if not all([t.shape == v.shape for t, v in zip(times, values)]):
raise ValueError('All arrays in \'times\' must match the shape of the'
' corresponding entry in \'values\'.')
# If all checks are passed, convert all times and values args into TimeSeries objects
timeseries.extend([core.TimeSeries(t, v) for t, v in zip(times, values)])
else:
# If times and values are only numpy arrays and lists of arrays, pair them and add
timeseries.append(core.TimeSeries(times, values))
# Adjust each timeseries by the associated offset if necessary then load into a list
if offsets is not None:
tstamps = [ts.times + os for ts, os in zip(timeseries, offsets)]
else:
tstamps = [ts.times for ts in timeseries]
# If all input timeseries have column names, put them together for the output TS
if all([ts.columns is not None for ts in timeseries]):
colnames = []
for ts in timeseries:
colnames.extend(ts.columns)
else:
colnames = None
#################
# Main function #
#################
# Get the min and max values for all timeseries combined after offsetting
tbounds = np.array([(np.amin(ts), np.amax(ts)) for ts in tstamps])
if not np.all(np.isfinite(tbounds)):
# If there is a np.inf or np.nan in the time stamps for any of the timeseries this will
# break any further code so we check for all finite values and throw an informative error.
raise ValueError('NaN or inf encountered in passed timeseries.\
Please either drop or fill these values.')
tmin, tmax = np.amin(tbounds[:, 0]), np.amax(tbounds[:, 1])
if fillval == 'extrapolate':
# If extrapolation is enabled we can ensure we have a full coverage of the data by
# extending the t max to be an whole integer multiple of dt above tmin.
# The 0.01% fudge factor is to account for floating point arithmetic errors.
newt = np.arange(tmin, tmax + 1.0001 * (dt - (tmax - tmin) % dt), dt)
else:
newt = np.arange(tmin, tmax, dt)
tsinterps = [interpolate.interp1d(ts.times, ts.values, kind=interp, fill_value=fillval, axis=0)
for ts in timeseries]
syncd = core.TimeSeries(newt, np.hstack([tsi(newt) for tsi in tsinterps]), columns=colnames)
return syncd
[docs]
def compute_cluster_average(spike_clusters, spike_var):
"""
Quickish way to compute the average of some quantity across spikes in each cluster given
quantity for each spike
:param spike_clusters: cluster idx of each spike
:param spike_var: variable of each spike (e.g spike amps or spike depths)
:return: cluster id, average of quantity for each cluster, no. of spikes per cluster
"""
clust, inverse, counts = np.unique(spike_clusters, return_inverse=True, return_counts=True)
_spike_var = sparse.csr_matrix((spike_var, (inverse, np.zeros(inverse.size, dtype=int))))
spike_var_avg = np.ravel(_spike_var.toarray()) / counts
return clust, spike_var_avg, counts
[docs]
def bin_spikes(spikes, binsize, interval_indices=False):
"""
Wrapper for bincount2D which is intended to take in a TimeSeries object of spike times
and cluster identities and spit out spike counts in bins of a specified width binsize, also in
another TimeSeries object. Can either return a TS object with each row labeled with the
corresponding interval or the value of the left edge of the bin.
:param spikes: Spike times and cluster identities of sorted spikes
:type spikes: TimeSeries object with \'clusters\' column and timestamps
:param binsize: Width of the non-overlapping bins in which to bin spikes
:type binsize: float
:param interval_indices: Whether to use intervals as the time stamps for binned spikes, rather
than the left edge value of the bins, defaults to False
:type interval_indices: bool, optional
:return: Object with 2D array of shape T x N, for T timesteps and N clusters, and the
associated time stamps.
:rtype: TimeSeries object
"""
if type(spikes) is not core.TimeSeries:
raise TypeError('Input spikes need to be in TimeSeries object format')
if not hasattr(spikes, 'clusters'):
raise AttributeError('Input spikes need to have a clusters attribute. Make sure you set '
'columns=(\'clusters\',)) when constructing spikes.')
rates, tbins, clusters = bincount2D(spikes.times, spikes.clusters, binsize)
if interval_indices:
intervals = pd.interval_range(tbins[0], tbins[-1], freq=binsize, closed='left')
return core.TimeSeries(times=intervals, values=rates.T[:-1], columns=clusters)
else:
return core.TimeSeries(times=tbins, values=rates.T, columns=clusters)
[docs]
def get_units_bunch(spks_b, *args):
"""
Returns a bunch, where the bunch keys are keys from `spks` with labels of spike information
(e.g. unit IDs, times, features, etc.), and the values for each key are arrays with values for
each unit: these arrays are ordered and can be indexed by unit id.
Parameters
----------
spks_b : bunch
A spikes bunch containing fields with spike information (e.g. unit IDs, times, features,
etc.) for all spikes.
features : list of strings (optional positional arg)
A list of names of labels of spike information (which must be keys in `spks`) that specify
which labels to return as keys in `units`. If not provided, all keys in `spks` are returned
as keys in `units`.
Returns
-------
units_b : bunch
A bunch with keys of labels of spike information (e.g. cluster IDs, times, features, etc.)
whose values are arrays that hold values for each unit. The arrays for each key are ordered
by unit ID.
Examples
--------
1) Create a units bunch given a spikes bunch, and get the amps for unit #4 from the units
bunch.
>>> from brainbox import processing
>>> import one.alf.io as alfio
>>> import ibllib.ephys.spikes as e_spks
(*Note, if there is no 'alf' directory, make 'alf' directory from 'ks2' output directory):
>>> e_spks.ks2_to_alf(path_to_ks_out, path_to_alf_out)
>>> spks_b = alfio.load_object(path_to_alf_out, 'spikes')
>>> units_b = processing.get_units_bunch(spks_b)
# Get amplitudes for unit 4.
>>> amps = units_b['amps']['4']
TODO add computation time estimate?
"""
# Initialize `units`
units_b = Bunch()
# Get the keys to return for `units`:
if not args:
feat_keys = list(spks_b.keys())
else:
feat_keys = args[0]
# Get unit id for each spike and number of units. *Note: `n_units` might not equal `len(units)`
# because some clusters may be empty (due to a "wontfix" bug in ks2).
spks_unit_id = spks_b['clusters']
n_units = np.max(spks_unit_id)
units = np.unique(spks_b['clusters'])
# For each key in `units`, iteratively get each unit's values and add as a key to a bunch,
# `feat_bunch`. After iterating through all units, add `feat_bunch` as a key to `units`:
for feat in feat_keys:
# Initialize `feat_bunch` with a key for each unit.
feat_bunch = Bunch((str(unit), np.array([])) for unit in np.arange(n_units))
for unit in units:
unit_idxs = np.where(spks_unit_id == unit)[0]
feat_bunch[str(unit)] = spks_b[feat][unit_idxs]
units_b[feat] = feat_bunch
return units_b
[docs]
def filter_units(units_b, t, **kwargs):
"""
Filters units according to some parameters. **kwargs are the keyword parameters used to filter
the units.
Parameters
----------
units_b : bunch
A bunch with keys of labels of spike information (e.g. cluster IDs, times, features, etc.)
whose values are arrays that hold values for each unit. The arrays for each key are ordered
by unit ID.
t : float
Duration of time over which to calculate the firing rate and false positive rate.
Keyword Parameters
------------------
min_amp : float
The minimum mean amplitude (in V) of the spikes in the unit. Default value is 50e-6.
min_fr : float
The minimum firing rate (in Hz) of the unit. Default value is 0.5.
max_fpr : float
The maximum false positive rate of the unit (using the fp formula in Hill et al. (2011)
J Neurosci 31: 8699-8705). Default value is 0.2.
rp : float
The refractory period (in s) of the unit. Used to calculate `max_fp`. Default value is
0.002.
Returns
-------
filt_units : ndarray
The ids of the filtered units.
See Also
--------
get_units_bunch
Examples
--------
1) Filter units according to the default parameters.
>>> from brainbox import processing
>>> import one.alf.io as alfio
>>> import ibllib.ephys.spikes as e_spks
(*Note, if there is no 'alf' directory, make 'alf' directory from 'ks2' output directory):
>>> e_spks.ks2_to_alf(path_to_ks_out, path_to_alf_out)
# Get a spikes bunch, units bunch, and filter the units.
>>> spks_b = alfio.load_object(path_to_alf_out, 'spikes')
>>> units_b = processing.get_units_bunch(spks_b, ['times', 'amps', 'clusters'])
>>> T = spks_b['times'][-1] - spks_b['times'][0]
>>> filtered_units = processing.filter_units(units_b, T)
2) Filter units with no minimum amplitude, a minimum firing rate of 1 Hz, and a max false
positive rate of 0.2, given a refractory period of 2 ms.
>>> filtered_units = processing.filter_units(units_b, T, min_amp=0, min_fr=1)
TODO: `units_b` input arg could eventually be replaced by `clstrs_b` if the required metrics
are in `clstrs_b['metrics']`
"""
# Set params
params = {'min_amp': 50e-6, 'min_fr': 0.5, 'max_fpr': 0.2, 'rp': 0.002} # defaults
params.update(kwargs) # update from **kwargs
# Iteratively filter the units for each filter param #
# -------------------------------------------------- #
units = np.asarray(list(units_b.amps.keys()))
# Remove empty clusters
empty_cl = np.where([len(units_b.amps[unit]) == 0 for unit in units])[0]
filt_units = np.delete(units, empty_cl)
for param in params.keys():
if param == 'min_amp': # return units above with amp > `'min_amp'`
mean_amps = np.asarray([np.mean(units_b.amps[unit]) for unit in filt_units])
filt_idxs = np.where(mean_amps > params['min_amp'])[0]
filt_units = filt_units[filt_idxs]
elif param == 'min_fr': # return units with fr > `'min_fr'`
fr = np.asarray([len(units_b.amps[unit]) /
(units_b.times[unit][-1] - units_b.times[unit][0])
for unit in filt_units])
filt_idxs = np.where(fr > params['min_fr'])[0]
filt_units = filt_units[filt_idxs]
elif param == 'max_fpr': # return units with fpr < `'max_fpr'`
fpr = np.zeros_like(filt_units, dtype='float')
for i, unit in enumerate(filt_units):
n_spks = len(units_b.amps[unit])
n_isi_viol = len(np.where(np.diff(units_b.times[unit]) < params['rp'])[0])
# fpr is min of roots of solved quadratic equation (Hill, et al. 2011).
c = (t * n_isi_viol) / (2 * params['rp'] * n_spks**2) # 3rd term in quadratic
fpr[i] = np.min(np.abs(np.roots([-1, 1, c]))) # solve quadratic
filt_idxs = np.where(fpr < params['max_fpr'])[0]
filt_units = filt_units[filt_idxs]
return filt_units.astype(int)