def get_clusters_data(self, load_all=None):
"""Return a list of Bunch instances, with attributes pos and spike_ids."""
if not len(self.cluster_ids):
return
cluster_ids = list(self.cluster_ids)
# Don't need the background when splitting.
if not load_all:
# Add None cluster which means background spikes.
cluster_ids = [None] + cluster_ids
bunchs = self.amplitudes[self.amplitude_name](cluster_ids, load_all=load_all) or ()
# Add a pos attribute in bunchs in addition to x and y.
for i, (cluster_id, bunch) in enumerate(zip(cluster_ids, bunchs)):
spike_ids = _as_array(bunch.spike_ids)
spike_times = _as_array(bunch.spike_times)
amplitudes = _as_array(bunch.amplitudes)
assert spike_ids.shape == spike_times.shape == amplitudes.shape
# Ensure that bunch.pos exists, as it used by the LassoMixin.
bunch.pos = np.c_[spike_times, amplitudes]
assert bunch.pos.ndim == 2
bunch.cluster_id = cluster_id
bunch.color = (
selected_cluster_color(i - 1, self.marker_alpha)
# Background amplitude color.
if cluster_id is not None else (.5, .5, .5, .5))
return bunchs
def _extend_assignment(spike_ids, old_spike_clusters, spike_clusters_rel,
new_cluster_id):
# 1. Add spikes that belong to modified clusters.
# 2. Find new cluster ids for all changed clusters.
old_spike_clusters = _as_array(old_spike_clusters)
spike_ids = _as_array(spike_ids)
assert isinstance(spike_clusters_rel, (list, np.ndarray))
spike_clusters_rel = _as_array(spike_clusters_rel)
assert spike_clusters_rel.min() >= 0
# We renumber the new cluster indices.
new_spike_clusters = (spike_clusters_rel +
(new_cluster_id - spike_clusters_rel.min()))
# We find the spikes belonging to modified clusters.
extended_spike_ids = _extend_spikes(spike_ids, old_spike_clusters)
if len(extended_spike_ids) == 0:
return spike_ids, new_spike_clusters
# We take their clusters.
extended_spike_clusters = old_spike_clusters[extended_spike_ids]
# Use relative numbers in extended_spike_clusters.
_, extended_spike_clusters = np.unique(extended_spike_clusters,
return_inverse=True)
# Generate new cluster numbers.
k = new_spike_clusters.max() + 1
extended_spike_clusters += (k - extended_spike_clusters.min())
# Finally, we concatenate spike_ids and extended_spike_ids.
return _concatenate_spike_clusters(
(spike_ids, new_spike_clusters),
(extended_spike_ids, extended_spike_clusters))
def _concatenate_spike_clusters(*pairs):
"""Concatenate a list of pairs (spike_ids, spike_clusters)."""
pairs = [(_as_array(x), _as_array(y)) for (x, y) in pairs]
concat = np.vstack([np.hstack((x[:, None], y[:, None])) for x, y in pairs])
reorder = np.argsort(concat[:, 0])
concat = concat[reorder, :]
return concat[:, 0].astype(np.int64), concat[:, 1].astype(np.int64)
def select_spikes_from_chunked(spike_times,
chunk_bounds,
max_n_spikes,
skip_chunks=0):
"""Select a maximum number of spikes among the specified ones so as to minimize the
number of chunks that contain those spikes."""
if len(spike_times) <= max_n_spikes:
return spike_times
spike_times = _as_array(spike_times)
chunk_bounds = _as_array(chunk_bounds)
spike_chunks = np.searchsorted(chunk_bounds, spike_times, side='right') - 1
chunk_sizes = np.bincount(spike_chunks)
best_chunks = np.argsort(chunk_sizes)[::-1][skip_chunks:]
keep = np.zeros(len(spike_times), dtype=np.bool)
total = 0
for chunk_idx in best_chunks:
in_chunks = np.isin(spike_chunks, chunk_idx)
n_spikes_chunk = np.sum(in_chunks)
if total + n_spikes_chunk > max_n_spikes:
# Truncate to get the exact number of requested spikes.
last = np.nonzero(
np.cumsum(in_chunks) <= max_n_spikes - total)[0][-1]
in_chunks[last + 1:] = False
keep[in_chunks] = True
total += n_spikes_chunk
if total >= max_n_spikes:
break
return spike_times[keep]
def _increment(arr, indices):
"""Increment some indices in a 1D vector of non-negative integers.
Repeated indices are taken into account."""
arr = _as_array(arr)
indices = _as_array(indices)
bbins = np.bincount(indices)
arr[:len(bbins)] += bbins
return arr
def _in_polygon(points, polygon):
"""Return the points that are inside a polygon."""
from matplotlib.path import Path
points = _as_array(points)
polygon = _as_array(polygon)
assert points.ndim == 2
assert polygon.ndim == 2
if len(polygon):
polygon = np.vstack((polygon, polygon[0]))
path = Path(polygon, closed=True)
return path.contains_points(points)
def apply_filter(x, filter=None, axis=0):
"""Apply a filter to an array."""
x = _as_array(x)
if x.shape[axis] == 0:
return x
b, a = filter
with warnings.catch_warnings():
warnings.simplefilter("ignore")
return signal.filtfilt(b, a, x, axis=axis)
def _check_positions(self, positions):
if positions is None:
return
positions = _as_array(positions)
if positions.shape[0] != self.n_channels:
raise ValueError("'positions' "
"(shape {0:s})".format(str(positions.shape)) +
" and 'n_channels' "
"({0:d})".format(self.n_channels) +
" do not match.")
def test_as_array():
ae(_as_array(3), [3])
ae(_as_array([3]), [3])
ae(_as_array(3.), [3.])
ae(_as_array([3.]), [3.])
with raises(ValueError):
_as_array(map)
def _unique(x):
"""Faster version of np.unique().
This version is restricted to 1D arrays of non-negative integers.
It is only faster if len(x) >> len(unique(x)).
"""
if x is None or len(x) == 0:
return np.array([], dtype=np.int64)
# WARNING: only keep positive values.
# cluster=-1 means "unclustered".
x = _as_array(x)
x = x[x >= 0]
bc = np.bincount(x)
return np.nonzero(bc)[0]
def firing_rate(spike_clusters,
cluster_ids=None,
bin_size=None,
duration=None):
"""Compute the average number of spikes per cluster per bin."""
# Take the cluster order into account.
if cluster_ids is None:
clusters = _unique(spike_clusters)
else:
clusters = _as_array(cluster_ids)
# Like spike_clusters, but with 0..n_clusters-1 indices.
spike_clusters_i = _index_of(spike_clusters, clusters)
assert duration > 0
assert bin_size > 0
bc = np.bincount(spike_clusters_i)
return bc * np.c_[bc] * (bin_size / duration)
def __init__(self,
spike_clusters,
new_cluster_id=None,
spikes_per_cluster=None):
super(Clustering, self).__init__()
self._undo_stack = History(base_item=(None, None, None))
# Spike -> cluster mapping.
self._spike_clusters = _as_array(spike_clusters)
self._spikes_per_cluster = {}
self._n_spikes = len(self._spike_clusters)
self._spike_ids = np.arange(self._n_spikes).astype(np.int64)
# We can pass the precomputed spikes_per_cluster dictionary for
# performance reasons.
self._update_cluster_ids(to_add=spikes_per_cluster)
self._new_cluster_id_0 = int(new_cluster_id
or self._spike_clusters.max() + 1)
self._new_cluster_id = self._new_cluster_id_0
assert self._new_cluster_id >= 0
assert np.all(self._spike_clusters < self._new_cluster_id)
# Keep a copy of the original spike clusters assignment.
self._spike_clusters_base = self._spike_clusters.copy()
def _do_assign(self, spike_ids, new_spike_clusters):
"""Make spike-cluster assignments after the spike selection has
been extended to full clusters."""
# Ensure spike_clusters has the right shape.
spike_ids = _as_array(spike_ids)
if len(new_spike_clusters) == 1 and len(spike_ids) > 1:
new_spike_clusters = np.ones(
len(spike_ids), dtype=np.int64) * new_spike_clusters[0]
old_spike_clusters = self._spike_clusters[spike_ids]
assert len(spike_ids) == len(old_spike_clusters)
assert len(new_spike_clusters) == len(spike_ids)
# Update the spikes per cluster structure.
old_clusters = _unique(old_spike_clusters)
# NOTE: shortcut to a merge if this assignment is effectively a merge
# i.e. if all spikes are assigned to a single cluster.
# The fact that spike selection has been previously extended to
# whole clusters is critical here.
new_clusters = _unique(new_spike_clusters)
if len(new_clusters) == 1:
return self._do_merge(spike_ids, old_clusters, new_clusters[0])
# We return the UpdateInfo structure.
up = _assign_update_info(spike_ids, old_spike_clusters,
new_spike_clusters)
# We update the new cluster id (strictly increasing during a session).
self._new_cluster_id = max(self._new_cluster_id, max(up.added) + 1)
# We make the assignments.
self._spike_clusters[spike_ids] = new_spike_clusters
# OPTIM: we update spikes_per_cluster manually.
new_spc = _spikes_per_cluster(new_spike_clusters, spike_ids)
self._update_cluster_ids(to_remove=old_clusters, to_add=new_spc)
up.all_cluster_ids = list(self.cluster_ids)
return up
def firing_rate(spike_clusters,
cluster_ids=None,
bin_size=None,
duration=None):
"""Compute the average number of spikes per cluster per bin."""
# Take the cluster order into account.
if cluster_ids is None:
cluster_ids = _unique(spike_clusters)
else:
cluster_ids = _as_array(cluster_ids)
# Like spike_clusters, but with 0..n_clusters-1 indices.
spike_clusters_i = _index_of(spike_clusters, cluster_ids)
assert bin_size > 0
bc = np.bincount(spike_clusters_i)
# Handle the case where the last cluster(s) are empty.
if len(bc) < len(cluster_ids):
n = len(cluster_ids) - len(bc)
bc = np.concatenate((bc, np.zeros(n, dtype=bc.dtype)))
assert bc.shape == (len(cluster_ids), )
return bc * np.c_[bc] * (bin_size / (duration or 1.))
def _normalize(self, x):
x = _as_array(x)
tw = self._thresholds['weak']
ts = self._thresholds['strong']
return np.clip((x - tw) / (ts - tw), 0, 1)
def get(self, spike_ids, channels=None):
"""Load the waveforms of the specified spikes."""
if isinstance(spike_ids, slice):
spike_ids = _range_from_slice(spike_ids,
start=0,
stop=self.n_spikes)
if not hasattr(spike_ids, '__len__'):
spike_ids = [spike_ids]
if channels is None:
channels = slice(None, None, None)
nc = self.n_channels
else:
channels = np.asarray(channels, dtype=np.int32)
assert np.all(channels < self.n_channels)
nc = len(channels)
# Ensure a list of time samples are being requested.
spike_ids = _as_array(spike_ids)
n_spikes = len(spike_ids)
# Initialize the array.
# NOTE: last dimension is time to simplify things.
shape = (n_spikes, nc, self._n_samples_extract)
waveforms = np.zeros(shape, dtype=np.float32)
# No traces: return null arrays.
if self.n_samples_trace == 0:
return np.transpose(waveforms, (0, 2, 1))
# Load all spikes.
for i, spike_id in enumerate(spike_ids):
assert 0 <= spike_id < self.n_spikes
time = self._spike_samples[spike_id]
# Extract the waveforms on the unmasked channels.
try:
w = self._load_at(time, channels)
except ValueError as e: # pragma: no cover
logger.warning("Error while loading waveform: %s", str(e))
continue
assert w.shape == (self._n_samples_extract, nc)
waveforms[i, :, :] = w.T
# Filter the waveforms.
waveforms_f = waveforms.reshape((-1, self._n_samples_extract))
# Only filter the non-zero waveforms.
unmasked = waveforms_f.max(axis=1) != 0
waveforms_f[unmasked] = self._filter(waveforms_f[unmasked], axis=1)
waveforms_f = waveforms_f.reshape(
(n_spikes, nc, self._n_samples_extract))
# Remove the margin.
margin_before, margin_after = self._filter_margin
if margin_after > 0:
assert margin_before >= 0
waveforms_f = waveforms_f[:, :, margin_before:-margin_after]
assert waveforms_f.shape == (n_spikes, nc, self.n_samples_waveforms)
# NOTE: we transpose before returning the array.
return np.transpose(waveforms_f, (0, 2, 1))
def _diff_shifted(arr, steps=1):
arr = _as_array(arr)
return arr[steps:] - arr[:len(arr) - steps]
def correlograms(
spike_times,
spike_clusters,
cluster_ids=None,
sample_rate=1.,
bin_size=None,
window_size=None,
symmetrize=True,
):
"""Compute all pairwise cross-correlograms among the clusters appearing
in `spike_clusters`.
Parameters
----------
spike_times : array-like
Spike times in seconds.
spike_clusters : array-like
Spike-cluster mapping.
cluster_ids : array-like
The list of *all* unique clusters, in any order. That order will be used
in the output array.
bin_size : float
Size of the bin, in seconds.
window_size : float
Size of the window, in seconds.
Returns
-------
correlograms : array
A `(n_clusters, n_clusters, winsize_samples)` array with all pairwise
CCGs.
"""
assert sample_rate > 0.
assert np.all(np.diff(spike_times) >= 0), ("The spike times must be "
"increasing.")
# Get the spike samples.
spike_times = np.asarray(spike_times, dtype=np.float64)
spike_samples = (spike_times * sample_rate).astype(np.int64)
spike_clusters = _as_array(spike_clusters)
assert spike_samples.ndim == 1
assert spike_samples.shape == spike_clusters.shape
# Find `binsize`.
bin_size = np.clip(bin_size, 1e-5, 1e5) # in seconds
binsize = int(sample_rate * bin_size) # in samples
assert binsize >= 1
# Find `winsize_bins`.
window_size = np.clip(window_size, 1e-5, 1e5) # in seconds
winsize_bins = 2 * int(.5 * window_size / bin_size) + 1
assert winsize_bins >= 1
assert winsize_bins % 2 == 1
# Take the cluster order into account.
if cluster_ids is None:
clusters = _unique(spike_clusters)
else:
clusters = _as_array(cluster_ids)
n_clusters = len(clusters)
# Like spike_clusters, but with 0..n_clusters-1 indices.
spike_clusters_i = _index_of(spike_clusters, clusters)
# Shift between the two copies of the spike trains.
shift = 1
# At a given shift, the mask precises which spikes have matching spikes
# within the correlogram time window.
mask = np.ones_like(spike_samples, dtype=np.bool)
correlograms = _create_correlograms_array(n_clusters, winsize_bins)
# The loop continues as long as there is at least one spike with
# a matching spike.
while mask[:-shift].any():
# Number of time samples between spike i and spike i+shift.
spike_diff = _diff_shifted(spike_samples, shift)
# Binarize the delays between spike i and spike i+shift.
spike_diff_b = spike_diff // binsize
# Spikes with no matching spikes are masked.
mask[:-shift][spike_diff_b > (winsize_bins // 2)] = False
# Cache the masked spike delays.
m = mask[:-shift].copy()
d = spike_diff_b[m]
# # Update the masks given the clusters to update.
# m0 = np.in1d(spike_clusters[:-shift], clusters)
# m = m & m0
# d = spike_diff_b[m]
d = spike_diff_b[m]
# Find the indices in the raveled correlograms array that need
# to be incremented, taking into account the spike clusters.
indices = np.ravel_multi_index(
(spike_clusters_i[:-shift][m], spike_clusters_i[+shift:][m], d),
correlograms.shape)
# Increment the matching spikes in the correlograms array.
_increment(correlograms.ravel(), indices)
shift += 1
# Remove ACG peaks.
correlograms[np.arange(n_clusters), np.arange(n_clusters), 0] = 0
if symmetrize:
return _symmetrize_correlograms(correlograms)
else:
return correlograms
def _zoom_aspect(self, zoom=None):
zoom = zoom if zoom is not None else self._zoom
zoom = _as_array(zoom)
aspect = (self._canvas_aspect * self._aspect if self._aspect is not None else 1.)
return zoom * aspect
def assign(self, spike_ids, spike_clusters_rel=0):
"""Make new spike cluster assignments.
Parameters
----------
spike_ids : array-like
List of spike ids.
spike_clusters_rel : array-like
Relative cluster ids of the spikes in `spike_ids`. This
must have the same size as `spike_ids`.
Returns
-------
up : UpdateInfo instance
Note
----
`spike_clusters_rel` contain *relative* cluster indices. Their values
don't matter: what matters is whether two give spikes
should end up in the same cluster or not. Adding a constant number
to all elements in `spike_clusters_rel` results in exactly the same
operation.
The final cluster ids are automatically generated by the `Clustering`
class. This is because we must ensure that all modified clusters
get brand new ids. The whole library is based on the assumption that
cluster ids are unique and "disposable". Changing a cluster always
results in a new cluster id being assigned.
If a spike is assigned to a new cluster, then all other spikes
belonging to the same cluster are assigned to a brand new cluster,
even if they were not changed explicitely by the `assign()` method.
In other words, the list of spikes affected by an `assign()` is almost
always a strict superset of the `spike_ids` parameter. The only case
where this is not true is when whole clusters change: this is called
a merge. It is implemented in a separate `merge()` method because it
is logically much simpler, and faster to execute.
"""
assert not isinstance(spike_ids, slice)
# Ensure `spike_clusters_rel` is an array-like.
if not hasattr(spike_clusters_rel, '__len__'):
spike_clusters_rel = spike_clusters_rel * np.ones(len(spike_ids),
dtype=np.int64)
spike_ids = _as_array(spike_ids)
if len(spike_ids) == 0:
return UpdateInfo()
assert len(spike_ids) == len(spike_clusters_rel)
assert spike_ids.min() >= 0
assert spike_ids.max() < self._n_spikes, "Some spikes don't exist."
# Normalize the spike-cluster assignment such that
# there are only new or dead clusters, not modified clusters.
# This implies that spikes not explicitly selected, but that
# belong to clusters affected by the operation, will be assigned
# to brand new clusters.
spike_ids, cluster_ids = _extend_assignment(spike_ids,
self._spike_clusters,
spike_clusters_rel,
self.new_cluster_id())
up = self._do_assign(spike_ids, cluster_ids)
undo_state = emit('request_undo_state', self, up)
# Add the assignment to the undo stack.
self._undo_stack.add((spike_ids, cluster_ids, undo_state))
emit('cluster', self, up)
return up
