def test_project_pcs():
x1 = np.random.randn(20000, 2) * np.array([[10., 1.]])
x2 = np.random.randn(20000, 2) * np.array([[1., 10.]])
x = np.dstack((x1, x2))
# Compute the PCs.
pcs = compute_pcs(x)
# Project the PCs.
x_proj = project_pcs(x[0, ...], pcs)
assert x_proj.shape == (2, 2)
def save_features(experiment, **prm):
"""Compute the features from the waveforms and save them in the experiment
dataset."""
nwaveforms_max = prm['pca_nwaveforms_max']
npcs = prm['nfeatures_per_channel']
kwik = experiment._files['kwik']
for chgrp in iterkeys(experiment.channel_groups):
spikes = experiment.channel_groups[chgrp].spikes
# Extract a subset of the saveforms.
nspikes = len(spikes)
# We convert the extendable features_masks array to a
# contiguous array.
if prm.get('features_contiguous', True):
# Make sure to update the PyTables node after the recreation,
# to avoid ClosedNodeError.
spikes.features_masks = to_contiguous(spikes.features_masks,
nspikes=nspikes)
else:
warn(("The features array has not been converted to a contiguous "
"array."))
# Skip the channel group if there are no spikes.
if nspikes == 0:
continue
nwaveforms = min(nspikes, nwaveforms_max)
step = excerpt_step(nspikes, nexcerpts=nwaveforms, excerpt_size=1)
waveforms_subset = spikes.waveforms_filtered[::step]
# With this option, PCs are directly provided in the PRM file as
# a NumPy array
if prm.get('canonical_pcs', None) is not None:
pcs = prm['canonical_pcs']
assert isinstance(pcs, np.ndarray)
else:
# We take the masks in order to compute the PCs only on
# the unmasked spikes, for each channel.
masks = spikes.features_masks[::step, ::npcs,
1] # (nspikes, nchannels)
# Compute the PCs.
pcs = compute_pcs(waveforms_subset, npcs=npcs, masks=masks)
# Add PCs to the KWIK file
kwik.createArray(experiment.channel_groups[chgrp]._node,
'pca_waveforms', pcs)
# Project the waveforms on the PCs and compute the features.
# WARNING: optimization: we could load and project waveforms by chunks.
for i, waveform in enumerate(spikes.waveforms_filtered):
# Convert waveforms from int16 to float32 with scaling
# before computing PCA so as to avoid getting huge numbers.
waveform = convert_dtype(waveform, np.float32)
features = project_pcs(waveform, pcs)
spikes.features_masks[i, :, 0] = features.ravel()
def test_project_pcs():
x1 = np.random.randn(20000, 2) * np.array([[10., 1.]])
x2 = np.random.randn(20000, 2) * np.array([[1., 10.]])
x = np.dstack((x1, x2))
# Compute the PCs.
pcs = compute_pcs(x)
# Project the PCs.
x_proj = project_pcs(x[0,...], pcs)
assert x_proj.shape == (2, 2)
def test_compute_pcs_3d():
"""Test PCA on a 3D array."""
x1 = np.random.randn(20000, 2) * np.array([[10., 1.]])
x2 = np.random.randn(20000, 2) * np.array([[1., 10.]])
x = np.dstack((x1, x2))
# Compute the PCs.
pcs = compute_pcs(x)
assert pcs.ndim == 3
assert np.linalg.norm(pcs[0, :, 0] - np.array([-1., 0.])) < 1e-2
assert np.linalg.norm(pcs[1, :, 0] - np.array([0., -1.])) < 1e-2
assert np.linalg.norm(pcs[0, :, 1] - np.array([0, 1.])) < 1e-2
assert np.linalg.norm(pcs[1, :, 1] - np.array([-1., 0.])) < 1e-2
def test_compute_pcs():
"""Test PCA on a 2D array."""
# Horizontal ellipsoid.
x = np.random.randn(20000, 2) * np.array([[10., 1.]])
# Rotate the points by pi/4.
a = 1. / np.sqrt(2.)
rot = np.array([[a, -a], [a, a]])
x = np.dot(x, rot)
# Compute the PCs.
pcs = compute_pcs(x)
assert pcs.ndim == 2
assert (np.abs(pcs) - a).max() < 1e-2
def test_compute_pcs_3d():
"""Test PCA on a 3D array."""
x1 = np.random.randn(20000, 2) * np.array([[10., 1.]])
x2 = np.random.randn(20000, 2) * np.array([[1., 10.]])
x = np.dstack((x1, x2))
# Compute the PCs.
pcs = compute_pcs(x)
assert pcs.ndim == 3
assert np.linalg.norm(pcs[0,:,0] - np.array([-1., 0.])) < 1e-2
assert np.linalg.norm(pcs[1,:,0] - np.array([0., -1.])) < 1e-2
assert np.linalg.norm(pcs[0,:,1] - np.array([0, 1.])) < 1e-2
assert np.linalg.norm(pcs[1,:,1] - np.array([-1., 0.])) < 1e-2
def test_compute_pcs():
"""Test PCA on a 2D array."""
# Horizontal ellipsoid.
x = np.random.randn(20000, 2) * np.array([[10., 1.]])
# Rotate the points by pi/4.
a = 1./np.sqrt(2.)
rot = np.array([[a, -a], [a, a]])
x = np.dot(x, rot)
# Compute the PCs.
pcs = compute_pcs(x)
assert pcs.ndim == 2
assert (np.abs(pcs) - a).max() < 1e-2
def save_features(experiment, **prm):
"""Compute the features from the waveforms and save them in the experiment
dataset."""
nwaveforms_max = prm['pca_nwaveforms_max']
npcs = prm['nfeatures_per_channel']
kwik = experiment._files['kwik']
for chgrp in iterkeys(experiment.channel_groups):
spikes = experiment.channel_groups[chgrp].spikes
# Extract a subset of the saveforms.
nspikes = len(spikes)
# We convert the extendable features_masks array to a
# contiguous array.
if prm.get('features_contiguous', True):
# Make sure to update the PyTables node after the recreation,
# to avoid ClosedNodeError.
spikes.features_masks = to_contiguous(spikes.features_masks, nspikes=nspikes)
else:
warn(("The features array has not been converted to a contiguous "
"array."))
# Skip the channel group if there are no spikes.
if nspikes == 0:
continue
nwaveforms = min(nspikes, nwaveforms_max)
step = excerpt_step(nspikes, nexcerpts=nwaveforms, excerpt_size=1)
waveforms_subset = spikes.waveforms_filtered[::step]
# With this option, PCs are directly provided in the PRM file as
# a NumPy array
if prm.get('canonical_pcs', None) is not None:
pcs = prm['canonical_pcs']
assert isinstance(pcs, np.ndarray)
else:
# We take the masks in order to compute the PCs only on
# the unmasked spikes, for each channel.
masks = spikes.features_masks[::step,::npcs,1] # (nspikes, nchannels)
# Compute the PCs.
pcs = compute_pcs(waveforms_subset, npcs=npcs, masks=masks)
# Add PCs to the KWIK file
kwik.createArray(experiment.channel_groups[chgrp]._node, 'pca_waveforms',
pcs)
# Project the waveforms on the PCs and compute the features.
# WARNING: optimization: we could load and project waveforms by chunks.
for i, waveform in enumerate(spikes.waveforms_filtered):
# Convert waveforms from int16 to float32 with scaling
# before computing PCA so as to avoid getting huge numbers.
waveform = convert_dtype(waveform, np.float32)
features = project_pcs(waveform, pcs)
spikes.features_masks[i,:,0] = features.ravel()