def supervised_separation(self, input_data, f, iter=50, adaptive_alpha=0, additional_basis=0):
self.f=f
self.time_vec=input_data[:,0:1]
self.adaptive_alpha=adaptive_alpha
self.additional_basis=additional_basis
input_data=input_data[:,1:].T
baseline=input_data.min()
input_data=input_data-baseline
# Modify the input data based on the feature length
data=pcnmf(feature_length=self.feature_length).matrix_conv(input_data)
# Check the learning rate (adaptive_alpha) for each source (adaptive NMF)
if isinstance(adaptive_alpha, int) or isinstance(adaptive_alpha, float):
adaptive_alpha = np.ones(self.source_num)*adaptive_alpha
else:
if len(adaptive_alpha) != self.source_num:
print("Error: The model has " +str(self.source_num) +" sources. Please specify adaptive_alpha for every source")
return
# Add additional basis for feature learning (semi-supervised NMF)
if additional_basis>0:
self.W_cluster=np.append(self.W_cluster, np.ones(additional_basis)*self.source_num)
self.source_num=self.source_num+1
self.basis_num=self.basis_num+additional_basis
adaptive_alpha = np.append(adaptive_alpha, 1)
# supervised NMF
W, H = NMF_init(data, self.basis_num, init='random')
W[:, 0:self.basis_num-additional_basis] = self.W
violation=0
Ht=H.T
Ht = check_array(H.T, order='C')
X = check_array(data, accept_sparse='csr')
W = np.array(W)
for run in range(iter):
prev_W = np.array(W)
# update H
violation += basis_update(X.T, W=Ht, Ht=W, l1_reg=0, l2_reg=0, shuffle=False, random_state=None)
# update W
violation += basis_update(X, W=W, Ht=Ht, l1_reg=0, l2_reg=0, shuffle=False, random_state=None)
for i in range(self.source_num):
index = np.where(self.W_cluster == i)
if adaptive_alpha[i] == 0:
W[:,index] = prev_W[:,index]
else:
W[:,index] = prev_W[:,index]*(1-adaptive_alpha[i]) + W[:,index]*adaptive_alpha[i]
self.W=W
self.H=Ht.T
self.nmf_output(input_data, self.time_vec, baseline)
def supervised_separation(self, input_data, f, iter=50):
self.f = f
self.time_vec = input_data[:, 0:1]
input_data = input_data[:, 1:].T
baseline = input_data.min()
input_data = input_data - baseline
# Modify the input data based the feature width
print('Supervised NMF')
data = self.matrix_conv(input_data)
# supervised NMF
W, H = NMF_init(data, self.basis_num, init='random')
violation = 0
Ht = H.T
Ht = check_array(H.T, order='C')
X = check_array(data, accept_sparse='csr')
print('Learning temporal activitations...')
for run in range(iter):
W = self.W
violation += basis_update(X.T,
W=Ht,
Ht=W,
l1_reg=0,
l2_reg=0,
shuffle=False,
random_state=None)
self.H = Ht.T
self.pcnmf_output(input_data, self.time_vec, baseline)
self.time_vec = self.time_vec[:, 0]