def fit(self, data):
""" Fit the iONMF model.
Parameters
----------
data: dictionary of array describing samples with multiple data sources.
data = {
"data_source_1": X_1 array [n_samples, n_features_1],
"data_source_2": X_2 array [n_samples, n_features_2],
...
"data_source_N": X_N array [n_samples, n_features_N],
}
Data sources must match in the number of rows.
"""
self.keys_ = sorted(data.keys())
self.n_ = [data[ky].shape[1] for ky in self.keys_]
self.m = data[self.keys_[0]].shape[0]
if not all([data[ky].shape[0] == self.m for ky in self.keys_]):
raise ValueError("The number of rows must match for all matrices!")
# Fill training data matrix
X = zeros((self.m, sum(self.n_)))
t = 0
for ny, ky in zip(self.n_, self.keys_,):
X[:, t:t+ny] = data[ky]
t += ny
# Run factorization
W, H = onmf(X, rank=self.rank,
max_iter=self.max_iter,
alpha=self.alpha)
# Set model variables
self.coef_ = W
self.basis_ = dict()
t = 0
for ny, ky in zip(self.n_, self.keys_):
self.basis_[ky] = H[:, t:t+ny]
t += ny
self.instantiated = True
def predict(self, data_test):
""" Predict the values for test samples based on a (non-empty)
subset of avalible data sources .
Parameters
----------
data_test: dictionary of array describing samples with multiple data sources.
data_test = {
"data_source_1": X_1 array [n_samples, n_features_1],
"data_source_2": X_2 array [n_samples, n_features_2],
...
"data_source_N": X_N array [n_samples, n_features_N],
}
Test data sources must match in the number of rows.
Dictionary keys must match keys in the traning data.
"""
if not self.instantiated:
raise ValueError("Run the method iONMF.fit first!")
keys_test = sorted(data_test.keys())
if set(keys_test) - set(self.keys_):
raise ValueError("Test data dictionary contains unknown keys!")
m_test = len(data_test[keys_test[0]])
if not all([data_test[ky].shape[0] == m_test for ky in keys_test]):
raise ValueError("The number of rows must match for all matrices!")
X_test = hstack([data_test[ky] for ky in keys_test])
H_test = hstack([self.basis_[ky] for ky in keys_test])
# Infer values of W for test samples based on a non-empty subset of
# observed data sources (matrices)
W_test, _ = onmf(X_test, rank=self.rank, alpha=self.alpha,
H_init=H_test)
# Predict all remaining matrices
remaining_keys = set(self.keys_) - set(keys_test)
predictions = dict([(ky, W_test.dot(self.basis_[ky]))
for ky in remaining_keys])
return predictions
