def _demixing_matrix(dataset):
"""Calculate the linear transformation to demix two channels.
Parameters
----------
dataset : ImagingDataset
The dataset which is to be demixed. This must have two channels.
Returns
-------
array
Matrix by which the data can be (left) multiplied to be demixed.
"""
from mdp.nodes import FastICANode
# Make matrix of the time averaged channels.
time_avgs = np.concatenate(
[im.reshape(-1, 1) for im in dataset.time_averages], axis=1)
# Perform ICA on the time averaged data.
node = FastICANode() # TODO: switch to ICA from scikit-learn ?
node(time_avgs)
W = np.dot(node.white.v, node.filters).T
# Reorder and normalize the rows so that the diagonal coefficients
# are 1 and the off diagonals have minimal magnitude.
if abs(old_div(W[0, 0], W[0, 1])) < abs(old_div(W[1, 0], W[1, 1])):
W = W[::-1]
W[0] /= W[0, 0]
W[1] /= W[1, 1]
assert np.allclose(np.diag(W), 1.)
return W
def _train(self, data, label=None):
""" Uses *data* to learn a decomposition into independent components."""
# We simply ignore the class label since we
# are doing unsupervised learning
if self.channel_names is None:
self.channel_names = data.channel_names
if self.wrapped_node is None:
self.wrapped_node = FastICANode()
self.wrapped_node.train(data)
def estimate_components_ica(d):
"""
Compute the ICA based on the input data d.
"""
U, s, Vt = pca_components_gf(d)
U = U[:, :NUM_COMPONENTS]
V = np.transpose(Vt)
V = V[:, :NUM_COMPONENTS]
f = FastICANode(whitened=True,
max_it=10000,
g='tanh',
fine_g='tanh',
max_it_fine=1000)
f.execute(V)
P = f.get_projmatrix()
Ur = np.dot(U, P)
Ur /= np.sum(Ur**2, axis=0)**0.5
return Ur
