def apply_filter(x, filter=None):
if x.shape[0] == 0:
return x
b, a = filter
try:
out_arr = signal.filtfilt(b, a, x, axis=0)
except TypeError:
out_arr = np.zeros_like(x)
for i_ch in range(x.shape[1]):
out_arr[:, i_ch] = signal.filtfilt(b, a, x[:, i_ch])
return out_arr
def apply_filter(x, filter=None):
if x.shape[0] == 0:
return x
b, a = filter
try:
out_arr = signal.filtfilt(b, a, x, axis=0)
except TypeError:
out_arr = np.zeros_like(x)
for i_ch in range(x.shape[1]):
out_arr[:, i_ch] = signal.filtfilt(b, a, x[:, i_ch])
return out_arr
def compute_pcs(x, npcs=None, masks=None):
"""Compute the PCs of an array x, where each row is an observation.
x can be a 2D or 3D array. In the latter case, the PCs are computed
and concatenated iteratively along the last axis."""
# If x is a 3D array, compute the PCs by iterating over the last axis.
if x.ndim == 3:
if masks is None:
return np.dstack([
compute_pcs(x[..., i], npcs=npcs, masks=None)
for i in range(x.shape[-1])
])
else:
# We pass the masks to compute_pcs for each row of the 3D array.
assert isinstance(masks, np.ndarray)
assert masks.ndim == 2
assert masks.shape[0] == x.shape[0] # number of spikes
assert masks.shape[1] == x.shape[-1] # number of channels
return np.dstack([
compute_pcs(x[..., i], npcs=npcs, masks=masks[..., i])
for i in range(x.shape[-1])
])
# Now, we assume x is a 2D array.
assert x.ndim == 2
# Check the masks.
if masks is not None:
assert masks.ndim == 1
assert masks.shape[0] == x.shape[0] # number of spikes
# Only select those rows in x that are *unmasked* (mask>0).
x = np.compress(masks > 0, x, axis=0)
if len(x) == 0:
return np.zeros((npcs, x.shape[-1]), dtype=np.float32)
# Take the covariance matrix.
cov_ss = np.cov(x.astype(np.float64), rowvar=0)
# Compute the eigenelements.
vals, vecs = np.linalg.eigh(cov_ss)
pcs = vecs.T.astype(np.float32)[np.argsort(vals)[::-1]]
# Take the first npcs components.
if npcs is not None:
return pcs[:npcs, ...]
else:
return pcs
def compute_pcs(x, npcs=None, masks=None):
"""Compute the PCs of an array x, where each row is an observation.
x can be a 2D or 3D array. In the latter case, the PCs are computed
and concatenated iteratively along the last axis."""
# If x is a 3D array, compute the PCs by iterating over the last axis.
if x.ndim == 3:
if masks is None:
return np.dstack([compute_pcs(x[..., i], npcs=npcs, masks=None)
for i in range(x.shape[-1])])
else:
# We pass the masks to compute_pcs for each row of the 3D array.
assert isinstance(masks, np.ndarray)
assert masks.ndim == 2
assert masks.shape[0] == x.shape[0] # number of spikes
assert masks.shape[1] == x.shape[-1] # number of channels
return np.dstack([compute_pcs(x[..., i], npcs=npcs,
masks=masks[..., i])
for i in range(x.shape[-1])])
# Now, we assume x is a 2D array.
assert x.ndim == 2
# Check the masks.
if masks is not None:
assert masks.ndim == 1
assert masks.shape[0] == x.shape[0] # number of spikes
# Only select those rows in x that are *unmasked* (mask>0).
x = np.compress(masks>0, x, axis=0)
if len(x) == 0:
return np.zeros((npcs, x.shape[-1]), dtype=np.float32)
# Take the covariance matrix.
cov_ss = np.cov(x.astype(np.float64), rowvar=0)
# Compute the eigenelements.
vals, vecs = np.linalg.eigh(cov_ss)
pcs = vecs.T.astype(np.float32)[np.argsort(vals)[::-1]]
# Take the first npcs components.
if npcs is not None:
return pcs[:npcs,...]
else:
return pcs
def compute_pcs(x, npcs=None, masks=None):
"""Compute the PCs of an array x, where each row is an observation.
x can be a 2D or 3D array. In the latter case, the PCs are computed
and concatenated iteratively along the last axis."""
# Ensure x is a 3D array.
if x.ndim == 2:
x = x[..., None]
assert x.ndim == 3
# Ensure double precision
x = x.astype(np.float64)
nspikes, nsamples, nchannels = x.shape
if masks is not None:
assert isinstance(masks, np.ndarray)
assert masks.ndim == 2
assert masks.shape[0] == x.shape[0] # number of spikes
assert masks.shape[1] == x.shape[2] # number of channels
# Compute regularization cov matrix.
if masks is not None:
unmasked = masks > 0
# The last dimension is now time. The second dimension is channel.
x_swapped = np.swapaxes(x, 1, 2)
# This is the list of all unmasked spikes on all channels.
# shape: (n_unmasked_spikes, nsamples)
unmasked_all = x_swapped[unmasked, :]
# Let's compute the regularization cov matrix of this beast.
# shape: (nsamples, nsamples)
cov_reg = np.cov(unmasked_all, rowvar=0)
else:
cov_reg = np.eye(nsamples)
assert cov_reg.shape == (nsamples, nsamples)
pcs_list = []
# Loop over channels
for channel in range(nchannels):
x_channel = x[:, :, channel]
# Compute cov matrix for the channel
if masks is not None:
# Unmasked waveforms on that channel
# shape: (n_unmasked, nsamples)
x_channel = np.compress(masks[:, channel] > 0, x_channel, axis=0)
assert x_channel.ndim == 2
# Don't compute the cov matrix if there are no unmasked spikes
# on that channel.
alpha = 1. / nspikes
if x_channel.shape[0] <= 1:
cov = alpha * cov_reg
else:
cov_channel = np.cov(x_channel, rowvar=0)
assert cov_channel.shape == (nsamples, nsamples)
cov = alpha * cov_reg + cov_channel
# Compute the eigenelements
vals, vecs = np.linalg.eigh(cov)
pcs = vecs.T.astype(np.float32)[np.argsort(vals)[::-1]]
# Take the first npcs components.
if npcs is not None:
pcs_list.append(pcs[:npcs, ...])
else:
pcs_list.append(pcs)
# Return the concatenation of the PCs on all channels, along the 3d axis,
# except if there is only one element in the 3d axis. In this case
# we convert to a 2D array.
pcs = np.dstack(pcs_list)
assert pcs.ndim == 3
if pcs.shape[2] == 1:
pcs = pcs[:, :, 0]
assert pcs.ndim == 2
return pcs
def compute_pcs(x, npcs=None, masks=None):
"""Compute the PCs of an array x, where each row is an observation.
x can be a 2D or 3D array. In the latter case, the PCs are computed
and concatenated iteratively along the last axis."""
# Ensure x is a 3D array.
if x.ndim == 2:
x = x[..., None]
assert x.ndim == 3
# Ensure double precision
x = x.astype(np.float64)
nspikes, nsamples, nchannels = x.shape
if masks is not None:
assert isinstance(masks, np.ndarray)
assert masks.ndim == 2
assert masks.shape[0] == x.shape[0] # number of spikes
assert masks.shape[1] == x.shape[2] # number of channels
# Compute regularization cov matrix.
if masks is not None:
unmasked = masks > 0
# The last dimension is now time. The second dimension is channel.
x_swapped = np.swapaxes(x, 1, 2)
# This is the list of all unmasked spikes on all channels.
# shape: (n_unmasked_spikes, nsamples)
unmasked_all = x_swapped[unmasked, :]
# Let's compute the regularization cov matrix of this beast.
# shape: (nsamples, nsamples)
cov_reg = np.cov(unmasked_all, rowvar=0)
else:
cov_reg = np.eye(nsamples)
assert cov_reg.shape == (nsamples, nsamples)
pcs_list = []
# Loop over channels
for channel in range(nchannels):
x_channel = x[:, :, channel]
# Compute cov matrix for the channel
if masks is not None:
# Unmasked waveforms on that channel
# shape: (n_unmasked, nsamples)
x_channel = np.compress(masks[:, channel] > 0,
x_channel, axis=0)
assert x_channel.ndim == 2
# Don't compute the cov matrix if there are no unmasked spikes
# on that channel.
alpha = 1. / nspikes
if x_channel.shape[0] <= 1:
cov = alpha * cov_reg
else:
cov_channel = np.cov(x_channel, rowvar=0)
assert cov_channel.shape == (nsamples, nsamples)
cov = alpha * cov_reg + cov_channel
# Compute the eigenelements
vals, vecs = np.linalg.eigh(cov)
pcs = vecs.T.astype(np.float32)[np.argsort(vals)[::-1]]
# Take the first npcs components.
if npcs is not None:
pcs_list.append(pcs[:npcs,...])
else:
pcs_list.append(pcs)
# Return the concatenation of the PCs on all channels, along the 3d axis,
# except if there is only one element in the 3d axis. In this case
# we convert to a 2D array.
pcs = np.dstack(pcs_list)
assert pcs.ndim == 3
if pcs.shape[2] == 1:
pcs = pcs[:, :, 0]
assert pcs.ndim == 2
return pcs
