def _assign_update_info(spike_ids, old_spike_clusters, new_spike_clusters):
old_clusters = _unique(old_spike_clusters)
new_clusters = _unique(new_spike_clusters)
largest_old_cluster = np.bincount(old_spike_clusters).argmax()
descendants = list(set(zip(old_spike_clusters, new_spike_clusters)))
update_info = UpdateInfo(
description='assign',
spike_ids=list(spike_ids),
spike_clusters=list(new_spike_clusters),
added=list(new_clusters),
deleted=list(old_clusters),
descendants=descendants,
largest_old_cluster=int(largest_old_cluster),
)
return update_info
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 _extend_spikes(spike_ids, spike_clusters):
"""Return all spikes belonging to the clusters containing the specified
spikes."""
# We find the spikes belonging to modified clusters.
# What are the old clusters that are modified by the assignment?
old_spike_clusters = spike_clusters[spike_ids]
unique_clusters = _unique(old_spike_clusters)
# Now we take all spikes from these clusters.
changed_spike_ids = _spikes_in_clusters(spike_clusters, unique_clusters)
# These are the new spikes that need to be reassigned.
extended_spike_ids = np.setdiff1d(changed_spike_ids,
spike_ids,
assume_unique=True)
return extended_spike_ids
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 _update_cluster_ids(self, to_remove=None, to_add=None):
# Update the list of non-empty cluster ids.
self._cluster_ids = _unique(self._spike_clusters)
# Clusters to remove.
if to_remove is not None:
for clu in to_remove:
self._spikes_per_cluster.pop(clu, None)
# Clusters to add.
if to_add:
for clu, spk in to_add.items():
self._spikes_per_cluster[clu] = spk
# If spikes_per_cluster is invalid, recompute the entire
# spikes_per_cluster array.
coherent = np.all(
np.in1d(self._cluster_ids, sorted(self._spikes_per_cluster)))
if not coherent:
logger.debug(
"Recompute spikes_per_cluster manually: this might take a while."
)
sc = self._spike_clusters
self._spikes_per_cluster = _spikes_per_cluster(sc)
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 make_depths(self):
"""Make spikes.depths.npy and clusters.depths.npy."""
channel_positions = self.model.channel_positions
assert channel_positions.ndim == 2
spike_clusters = self.model.spike_clusters
assert spike_clusters.ndim == 1
n_spikes = spike_clusters.shape[0]
self.cluster_ids = _unique(self.model.spike_clusters)
cluster_channels = np.load(self.out_path / 'clusters.peakChannel.npy')
assert cluster_channels.ndim == 1
n_clusters = cluster_channels.shape[0]
clusters_depths = channel_positions[cluster_channels, 1]
assert clusters_depths.shape == (n_clusters, )
spike_clusters_rel = _index_of(spike_clusters, self.cluster_ids)
assert spike_clusters_rel.max() < clusters_depths.shape[0]
spikes_depths = clusters_depths[spike_clusters_rel]
assert spikes_depths.shape == (n_spikes, )
self._save_npy('spikes.depths.npy', spikes_depths)
self._save_npy('clusters.depths.npy', clusters_depths)
def __init__(self, model):
self.model = model
self.dir_path = Path(model.dir_path)
self.spc = _spikes_per_cluster(model.spike_clusters)
self.cluster_ids = _unique(self.model.spike_clusters)
def merge_probes(ses_path):
"""
Merge spike sorting output from 2 probes and output in the session ALF folder the combined
output in IBL format
:param ses_path: session containing probes to be merged
:return: None
"""
def _sr(ap_file):
md = spikeglx.read_meta_data(ap_file.with_suffix('.meta'))
return spikeglx._get_fs_from_meta(md)
ses_path = Path(ses_path)
out_dir = ses_path.joinpath('alf').joinpath('tmp_merge')
ephys_files = glob_ephys_files(ses_path)
subdirs, labels, efiles_sorted, srates = zip(
*sorted([(ep.ap.parent, ep.label, ep, _sr(ep.ap)) for ep in ephys_files if ep.get('ap')]))
# if there is only one file, just convert the output to IBL format et basta
if len(subdirs) == 1:
ks2_to_alf(subdirs[0], ses_path / 'alf')
return
else:
_logger.info('converting individual spike-sorting outputs to ALF')
for subdir, label, ef, sr in zip(subdirs, labels, efiles_sorted, srates):
ks2alf_path = subdir / 'ks2_alf'
if ks2alf_path.exists():
shutil.rmtree(ks2alf_path, ignore_errors=True)
ks2_to_alf(subdir, ks2alf_path, label=label, sr=sr, force=True)
probe_info = [{'label': lab} for lab in labels]
mt = merge.Merger(subdirs=subdirs, out_dir=out_dir, probe_info=probe_info).merge()
# Create the cluster channels file, this should go in the model template as 2 methods
tmp = mt.sparse_templates.data
n_templates, n_samples, n_channels = tmp.shape
template_peak_channels = np.argmax(tmp.max(axis=1) - tmp.min(axis=1), axis=1)
cluster_probes = mt.channel_probes[template_peak_channels]
spike_clusters_rel = _index_of(mt.spike_clusters, _unique(mt.spike_clusters))
spike_probes = cluster_probes[spike_clusters_rel]
# sync spikes according to the probes
# how do you make sure they match the files:
for ind, probe in enumerate(efiles_sorted):
assert(labels[ind] == probe.label) # paranoid, make sure they are sorted
if not probe.get('ap'):
continue
sync_file = probe.ap.parent.joinpath(probe.ap.name.replace('.ap.', '.sync.')
).with_suffix('.npy')
if not sync_file.exists():
error_msg = f'No synchronisation file for {sync_file}'
_logger.error(error_msg)
raise FileNotFoundError(error_msg)
sync_points = np.load(sync_file)
fcn = interp1d(sync_points[:, 0] * srates[ind],
sync_points[:, 1], fill_value='extrapolate')
mt.spike_times[spike_probes == ind] = fcn(mt.spike_samples[spike_probes == ind])
# And convert to ALF
ac = alf.EphysAlfCreator(mt)
ac.convert(ses_path / 'alf', force=True)
# remove the temporary directory
shutil.rmtree(out_dir)
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
