def learnModel(self, X):
"""
Learn X using a matrix factorisation method. If self.rank is an integer
then we factorise with that rank. If it is an array then we compute the
complete regularisation path and return a list of matrices.
"""
if isinstance(self.rank, int):
model = nimfa.mf(X, method=self.method, max_iter=self.maxIter, rank=self.rank)
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
predX = W.dot(H)
return predX
else:
predXList = []
model = nimfa.mf(X, method=self.method, max_iter=self.maxIter, rank=self.rank[0])
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
predXList.append(W.dot(H))
for i in range(1, self.rank.shape[0]):
model = nimfa.mf(X, method=self.method, max_iter=self.maxIter, rank=self.rank[i], W=W, H=H)
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
predXList.append(W.dot(H))
return predXList
def run_nmf(V):
"""
Run standard nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
# Euclidean
rank = 10
model = nimfa.mf(V,
seed="random_vcol",
rank=rank,
method="nmf",
max_iter=12,
initialize_only=True,
update='euclidean',
objective='fro')
fit = nimfa.mf_run(model)
print_info(fit)
# divergence
model = nimfa.mf(V,
seed="random_vcol",
rank=rank,
method="nmf",
max_iter=12,
initialize_only=True,
update='divergence',
objective='div')
fit = nimfa.mf_run(model)
print_info(fit)
def nmf(Xtrn, Xtst):
# Init matricies
Xtrn_n = np.shape(Xtrn)[0]
Xtst_n = np.shape(Xtst)[0]
Xtrn_nmf = np.zeros((Xtrn_n, my_rank))
Xtst_nmf = np.zeros((Xtst_n, my_rank))
print(file_name + ': Running non-negative matrix facorization w/ rank = ' + str(my_rank))
#Xtrn_fctr = nimfa.mf(Xtrn, method = 'nmf', seed = "fixed", max_iter = iters,
# rank = my_rank, update = 'euclidean', objective = 'fro')
print(file_name + ': \t on traning...')
for i in xrange(Xtrn_n):
Xtrn_fctr = nimfa.mf(Xtrn[i,:], method = 'lsnmf', max_iter = iters, rank = my_rank)
Xtrn_res = nimfa.mf_run(Xtrn_fctr)
Xtrn_nmf[i,:] = Xtrn_res.basis()
if (i%10000 == 0): print(file_name + ': \t iter ' + str(i))
print(file_name + ' \t on testing...')
for i in xrange(Xtst_n):
Xtst_fctr = nimfa.mf(Xtst[i,:], method = 'lsnmf', max_iter = iters, rank = my_rank)
Xtst_res = nimfa.mf_run(Xtrn_fctr)
Xtst_nmf[i,:] = Xtst_res.basis()
if (i%10000 == 0): print(file_name + ': \t iter ' + str(i))
"""
Xtrn_sm = Xtrn_res.summary()
Xtst_sm = Xtst_res.summary()
print(file_name + ': \t\t RSS \t Explained Var \t Iters')
print(file_name + ': Xtrn: \t' + str(Xtrn_sm['rss']) + '\t' +
str(Xtrn_sm['evar']) + '\t' + str(Xtrn_sm['n_iter']))
print(file_name + ': Xtst: ' + str(Xtst_sm['rss']) + '\t' +
str(Xtst_sm['evar']) + '\t' + str(Xtst_sm['n_iter']))
"""
return (Xtrn_nmf, Xtst_nmf)
def run_nmf(V):
"""
Run standard nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
# Euclidean
rank = 10
model = nimfa.mf(V,
seed = "random_vcol",
rank = rank,
method = "nmf",
max_iter = 12,
initialize_only = True,
update = 'euclidean',
objective = 'fro')
fit = nimfa.mf_run(model)
print_info(fit)
# divergence
model = nimfa.mf(V,
seed = "random_vcol",
rank = rank,
method = "nmf",
max_iter = 12,
initialize_only = True,
update = 'divergence',
objective = 'div')
fit = nimfa.mf_run(model)
print_info(fit)
def run_bd(V):
"""
Run Bayesian decomposition.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed = "random_c",
rank = rank,
method = "bd",
max_iter = 12,
initialize_only = True,
alpha = np.mat(np.zeros((V.shape[0], rank))),
beta = np.mat(np.zeros((rank, V.shape[1]))),
theta = .0,
k = .0,
sigma = 1.,
skip = 100,
stride = 1,
n_w = np.mat(np.zeros((rank, 1))),
n_h = np.mat(np.zeros((rank, 1))),
n_sigma = False)
fit = nimfa.mf_run(model)
print_info(fit)
def factorize(V):
"""
Perform SNMF/R factorization on the sparse MovieLens data matrix.
Return basis and mixture matrices of the fitted factorization model.
:param V: The MovieLens data matrix.
:type V: `scipy.sparse.csr_matrix`
"""
model = nimfa.mf(V,
seed="random_vcol",
rank=12,
method="snmf",
max_iter=15,
initialize_only=True,
version='r',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
print("Performing %s %s %d factorization ..." % (model, model.seed, model.rank))
fit = nimfa.mf_run(model)
print("... Finished")
sparse_w, sparse_h = fit.fit.sparseness()
print("""Stats:
- iterations: %d
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" % (fit.fit.n_iter, fit.distance(metric='euclidean'), sparse_w, sparse_h))
return fit.basis(), fit.coef()
def run_snmnmf(V, V1):
"""
Run sparse network-regularized multiple NMF.
:param V: First target matrix to estimate.
:type V: :class:`numpy.matrix`
:param V1: Second target matrix to estimate.
:type V1: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(target = (V, V1),
seed = "random_c",
rank = rank,
method = "snmnmf",
max_iter = 12,
initialize_only = True,
A = abs(sp.rand(V1.shape[1], V1.shape[1], density = 0.7, format = 'csr')),
B = abs(sp.rand(V.shape[1], V1.shape[1], density = 0.7, format = 'csr')),
gamma = 0.01,
gamma_1 = 0.01,
lamb = 0.01,
lamb_1 = 0.01)
fit = nimfa.mf_run(model)
# print all quality measures concerning first target and mixture matrix in multiple NMF
print_info(fit, idx = 0)
# print all quality measures concerning second target and mixture matrix in multiple NMF
print_info(fit, idx = 1)
def run(self,
seed='random_vcol',
method='nmf',
rank=3,
max_iter=65,
display_N_tokens=5,
display_N_documents=3):
#Re-initialise clusters
if self.clusters != []:
self.clusters = []
self.construct_term_doc_matrix(
pca=False
) #We cannot perform PCA with NMF because we only want non-negative vectors
V = self.td_matrix
model = nimfa.mf(V,
seed=seed,
method=method,
rank=rank,
max_iter=max_iter)
fitted = nimfa.mf_run(model)
w = fitted.basis()
h = fitted.coef()
self.split_documents(w,
h,
self.document_dict,
self.attributes,
display_N_tokens=display_N_tokens,
display_N_documents=display_N_documents)
#Just testing remove it
self.showfeatures(w, h, [
self.document_dict.values()[i]["raw"]
for i in range(numpy.shape(w)[0])
], self.attributes)
def decompose_nmf(spectrum_array,n_spectra): print '\nDecomposing spectra using NMF...' fctr = nimfa.mf(spectrum_array, method="nmf", max_iter=10000, rank=n_spectra, update='divergence', objective='div') fctr_res = nimfa.mf_run(fctr) a=n.transpose(n.array(fctr_res.basis())) coeffs=n.array(fctr_res.coef()) return a,coeffs
def factorization(V, rank=4):
"""
use nmf to factorize V
:rtype : (1) the projection matrix (2) the feature vector of V
"""
fctr = nimfa.mf(
V,
method="nmf",
max_iter=30,
rank=rank,
update="divergence",
objective="div",
callback_init=init_info,
callback=init_info,
)
fctr_res = nimfa.mf_run(fctr)
print "calculate generized inverse"
projection = pinv(fctr_res.basis().todense())
print "inverse finished"
return {
"projection": projection,
"feature": (projection * V),
"basis": fctr_res.basis(),
"coef": fctr_res.coef().todense(),
}
def factorize(V):
"""
Perform LSNMF factorization on the ORL faces data matrix.
Return basis and mixture matrices of the fitted factorization model.
:param V: The ORL faces data matrix.
:type V: `numpy.matrix`
"""
model = nimfa.mf(V,
seed = "random_vcol",
rank = 25,
method = "lsnmf",
max_iter = 50,
initialize_only = True,
sub_iter = 10,
inner_sub_iter = 10,
beta = 0.1,
min_residuals = 1e-8)
print "Performing %s %s %d factorization ..." % (model, model.seed, model.rank)
fit = nimfa.mf_run(model)
print "... Finished"
print """Stats:
- iterations: %d
- final projected gradients norm: %5.3f
- Euclidean distance: %5.3f""" % (fit.fit.n_iter, fit.distance(), fit.distance(metric = 'euclidean'))
return fit.basis(), fit.coef()
def run_bd(V):
"""
Run Bayesian decomposition.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed="random_c",
rank=rank,
method="bd",
max_iter=12,
initialize_only=True,
alpha=np.mat(np.zeros((V.shape[0], rank))),
beta=np.mat(np.zeros((rank, V.shape[1]))),
theta=.0,
k=.0,
sigma=1.,
skip=100,
stride=1,
n_w=np.mat(np.zeros((rank, 1))),
n_h=np.mat(np.zeros((rank, 1))),
n_sigma=False)
fit = nimfa.mf_run(model)
print_info(fit)
def run(self, **params):
if not self.dataConsolided:
print "NIMFA_SNMNMF: preparing data"
self.consolideTheData()
self.dataConsolided = True
print "NIMFA_SNMNMF: starting"
#
V = self.miRNA.as_matrix()
V1 = self.mRNA.as_matrix()
A = csr_matrix(self.gene2gene)
B = csr_matrix(self.miRNA2gene)
fctr = nimfa.mf(target = (V, V1),
seed = params['seed'], # e.g., "random_c",
rank = params['rank'], # e.g., 50,
method = "snmnmf",
max_iter = params['max_iter'], # e.g., 500,
initialize_only = True,
A = A ,
B = B,
n_run = 3,
gamma = self.g1,
gamma_1 = self.g2,
lamb = self.l1,
lamb_1 = self.l2)
fctr_res = nimfa.mf_run(fctr)
print "NIMFA_SNMNMF: done"
# extract the results
self.W = DataFrame(fctr_res.basis(), index = self.miRNA.index)
self.H1_miRNA = DataFrame(fctr_res.coef(0), columns = self.miRNA.columns)
self.H2_genes = DataFrame(fctr_res.coef(1), columns = self.mRNA.columns)
self.performance = NIMFA_SNMNMFPerformance(fctr_res)
def factorize(V):
"""
Perform NMF - Divergence factorization on the sparse Medlars data matrix.
Return basis and mixture matrices of the fitted factorization model.
:param V: The Medlars data matrix.
:type V: `scipy.sparse.csr_matrix`
"""
model = nimfa.mf(V,
seed="random_vcol",
rank=12,
method="nmf",
max_iter=15,
initialize_only=True,
update='divergence',
objective='div')
print("Performing %s %s %d factorization ..." %
(model, model.seed, model.rank))
fit = nimfa.mf_run(model)
print("... Finished")
sparse_w, sparse_h = fit.fit.sparseness()
print("""Stats:
- iterations: %d
- KL Divergence: %5.3f
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" %
(fit.fit.n_iter, fit.distance(), fit.distance(metric='euclidean'),
sparse_w, sparse_h))
return fit.basis(), fit.coef()
def factorize(V):
"""
Perform NMF - Divergence factorization on the sparse Medlars data matrix.
Return basis and mixture matrices of the fitted factorization model.
:param V: The Medlars data matrix.
:type V: `scipy.sparse.csr_matrix`
"""
model = nimfa.mf(V,
seed = "random_vcol",
rank = 12,
method = "nmf",
max_iter = 15,
initialize_only = True,
update = 'divergence',
objective = 'div')
print "Performing %s %s %d factorization ..." % (model, model.seed, model.rank)
fit = nimfa.mf_run(model)
print "... Finished"
sparse_w, sparse_h = fit.fit.sparseness()
print """Stats:
- iterations: %d
- KL Divergence: %5.3f
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" % (fit.fit.n_iter, fit.distance(), fit.distance(metric = 'euclidean'), sparse_w, sparse_h)
return fit.basis(), fit.coef()
def run_snmnmf(V, V1):
"""
Run sparse network-regularized multiple NMF.
:param V: First target matrix to estimate.
:type V: :class:`numpy.matrix`
:param V1: Second target matrix to estimate.
:type V1: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(
target=(V, V1),
seed="random_c",
rank=rank,
method="snmnmf",
max_iter=12,
initialize_only=True,
A=abs(sp.rand(V1.shape[1], V1.shape[1], density=0.7, format='csr')),
B=abs(sp.rand(V.shape[1], V1.shape[1], density=0.7, format='csr')),
gamma=0.01,
gamma_1=0.01,
lamb=0.01,
lamb_1=0.01)
fit = nimfa.mf_run(model)
# print all quality measures concerning first target and mixture matrix in
# multiple NMF
print_info(fit, idx=0)
# print all quality measures concerning second target and mixture matrix
# in multiple NMF
print_info(fit, idx=1)
def factorize(V):
"""
Perform LSNMF factorization on the CBCL faces data matrix.
Return basis and mixture matrices of the fitted factorization model.
:param V: The CBCL faces data matrix.
:type V: `numpy.matrix`
"""
model = nimfa.mf(V,
seed="random_vcol",
rank=49,
method="lsnmf",
max_iter=50,
initialize_only=True,
sub_iter=10,
inner_sub_iter=10,
beta=0.1,
min_residuals=1e-8)
print("Performing %s %s %d factorization ..." %
(model, model.seed, model.rank))
fit = nimfa.mf_run(model)
print("... Finished")
sparse_w, sparse_h = fit.fit.sparseness()
print("""Stats:
- iterations: %d
- final projected gradients norm: %5.3f
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" %
(fit.fit.n_iter, fit.distance(), fit.distance(metric='euclidean'),
sparse_w, sparse_h))
return fit.basis(), fit.coef()
def max_guess_select(ratings, users, rank=9, user=None):
matrix = sp.dok_matrix((len(users), len(users)))
for k, v in ratings.items():
matrix[users[k[0]], users[k[1]]] = v
matrix[users[k[1]], users[k[0]]] = v
# Run sparse matrix factorisation
factor = nimfa.mf(matrix,
seed="random_c",
rank=rank,
method="snmf",
max_iter=12,
initialize_only=True,
version='r',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
result = nimfa.mf_run(factor)
if user is None:
# Pick a user to expand
user = min(users, key=lambda u: len([i for i in ratings if u in i]))
recommendations = result.fitted()
rval = max([
i
for i in users if (i, user) not in ratings and (user, i) not in ratings
],
key=lambda x: recommendations[users[user], users[x]])
return user, rval
def factorize(V):
"""
Perform SNMF/R factorization on the sparse MovieLens data matrix.
Return basis and mixture matrices of the fitted factorization model.
:param V: The MovieLens data matrix.
:type V: `scipy.sparse.csr_matrix`
"""
model = nimfa.mf(V,
seed="random_vcol",
rank=12,
method="snmf",
max_iter=15,
initialize_only=True,
version='r',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
print "Performing %s %s %d factorization ..." % (model, model.seed, model.rank)
fit = nimfa.mf_run(model)
print "... Finished"
sparse_w, sparse_h = fit.fit.sparseness()
print """Stats:
- iterations: %d
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" % (fit.fit.n_iter, fit.distance(metric='euclidean'), sparse_w, sparse_h)
return fit.basis(), fit.coef()
def run(self, **params):
if not self.dataConsolided:
print "NIMFA_SNMNMF: preparing data"
self.consolideTheData()
self.dataConsolided = True
print "NIMFA_SNMNMF: starting"
#
V = self.miRNA.as_matrix()
V1 = self.mRNA.as_matrix()
A = csr_matrix(self.gene2gene)
B = csr_matrix(self.miRNA2gene)
fctr = nimfa.mf(
target=(V, V1),
seed=params['seed'], # e.g., "random_c",
rank=params['rank'], # e.g., 50,
method="snmnmf",
max_iter=params['max_iter'], # e.g., 500,
initialize_only=True,
A=A,
B=B,
n_run=1,
gamma=self.g1,
gamma_1=self.g2,
lamb=self.l1,
lamb_1=self.l2)
fctr_res = nimfa.mf_run(fctr)
print "NIMFA_SNMNMF: done"
# extract the results
self.W = DataFrame(fctr_res.basis(), index=self.miRNA.index)
self.H1_miRNA = DataFrame(fctr_res.coef(0), columns=self.miRNA.columns)
self.H2_genes = DataFrame(fctr_res.coef(1), columns=self.mRNA.columns)
self.performance = NIMFA_SNMNMFPerformance(fctr_res)
def nmfMatrix(self, V):
print "---"
print "NMF"
print "---"
V = np.array(V)
print "Target matrix"
print V
fctr = nimfa.mf(V, seed = 'random_vcol', method = 'lsnmf', rank = 40, max_iter = 10)
fctr_res = nimfa.mf_run(fctr)
W = fctr_res.basis()
print "Basis matrix"
print W
H = fctr_res.coef()
print "Coef"
print H
print "Estimate"
print np.dot(W, H)
print 'Rss: %5.4f' % fctr_res.fit.rss()
print 'Evar: %5.4f' % fctr_res.fit.evar()
print 'K-L divergence: %5.4f' % fctr_res.distance(metric = 'kl')
print 'Sparseness, W: %5.4f, H: %5.4f' % fctr_res.fit.sparseness()
return W, H
def nmf(matrix, k=c_K):
fctr = nimfa.mf(matrix,
seed='random_vcol',
method='lsnmf',
rank=k,
max_iter=c_NMF_MAXITR)
fctr_result = nimfa.mf_run(fctr)
return fctr_result.basis(), fctr_result.coef()
def nmfMatrix(self, V, method, rank, maxIter):
print "---"
print "NMF"
print "---"
V = np.array(V)
print "Target matrix"
print V.shape[0]
print V.shape[1]
print V
# X = sp.rand(V.shape[0], V.shape[1], density=1).tocsr()
# NMFの際の、基底数やイテレーションの設定
# rank = 8
# maxIter = 2000
# method = "snmf"
# init2arizer = nimfa.methods.seeding.random_vcol.Random_vcol()
initiarizer = nimfa.methods.seeding.random.Random()
initW, initH = initiarizer.initialize(V, rank, {})
fctr = nimfa.mf(V, seed = 'random_vcol', method = method, rank = rank, max_iter = maxIter)
# fctr = nimfa.mf(V, method = "lsnmf", rank = rank, max_iter = maxIter, W = initW, H = initH)
fctr_res = nimfa.mf_run(fctr)
W = fctr_res.basis()
print "Basis matrix"
print W.shape[0]
print W.shape[1]
print W
H = fctr_res.coef()
print "Coef"
print H.shape[0]
print H.shape[1]
print H
print "Estimate"
print np.dot(W, H)
print 'Rss: %5.4f' % fctr_res.fit.rss()
print 'Evar: %5.4f' % fctr_res.fit.evar()
print 'K-L divergence: %5.4f' % fctr_res.distance(metric = 'kl')
print 'Sparseness, W: %5.4f, H: %5.4f' % fctr_res.fit.sparseness()
sm = fctr_res.summary()
print type(sm)
# print "Rss: %8.3f" % sm['rss']
# # Print explained variance.
# print "Evar: %8.3f" % sm['evar']
# # Print actual number of iterations performed
# print "Iterations: %d" % sm['n_iter']
# プロットの際に不具合が生じるため,numpy.ndarray型に変換
NW = np.asarray(W)
NH = np.asarray(H)
return NW, NH, sm
def learnModel(self, X):
"""
Learn X using a matrix factorisation method. If self.rank is an integer
then we factorise with that rank. If it is an array then we compute the
complete regularisation path and return a list of matrices.
"""
if isinstance(self.rank, int):
model = nimfa.mf(X,
method=self.method,
max_iter=self.maxIter,
rank=self.rank)
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
predX = W.dot(H)
return predX
else:
predXList = []
model = nimfa.mf(X,
method=self.method,
max_iter=self.maxIter,
rank=self.rank[0])
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
predXList.append(W.dot(H))
for i in range(1, self.rank.shape[0]):
model = nimfa.mf(X,
method=self.method,
max_iter=self.maxIter,
rank=self.rank[i],
W=W,
H=H)
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
predXList.append(W.dot(H))
return predXList
def cluster_nmf(vectors, num_clusters):
"""
Takes in vectors and clusters them using Non Negative Matrix Factorization.
Inputs:
vectors -- matrix containing rows of vectors
num_clusters -- number of clusters to create
"""
print "Starting NMF clustering"
start_time = time.time()
# Run NMF
vectors_matrix = numpy.matrix(vectors)
vectors_matrix = vectors_matrix.transpose()
print "Created vectors_matrix"
# Generate random matrix factors which we will pass as fixed factors to nimfa.nmf
init_W = numpy.random.rand(vectors_matrix.shape[0], num_clusters)
init_H = numpy.random.rand(num_clusters, vectors_matrix.shape[1])
print "Generated random matrix factors"
fctr = nimfa.mf(vectors_matrix,
method="nmf",
seed="fixed",
W=init_W,
H=init_H,
rank=num_clusters)
fctr_res = nimfa.mf_run(fctr)
print "NIMFA"
# Basis matrix
W = fctr_res.basis()
# Mixture matrix
H = fctr_res.coef()
print "Extracted Basis and Mixture matrices"
# get assignments
assignment = []
for index in range(H.shape[1]):
column = list(H[:, index])
assignment.append(column.index(max(column)))
print "Assignments extracted"
# Print the loss function (Euclidean distance between target matrix and its estimate).
print "Euclidean distance: %5.3e" % fctr_res.distance(metric="euclidean")
end_time = time.time()
print "Clustering required", (end_time - start_time), "seconds"
return assignment
def _NIMFA_NMF(self, X, nBases):
model = nimfa.mf(X, seed="nndsvd", rank=nBases, method="nmf", initialize_only=True)
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
self.W = W.todense()
self.H = H.todense()
return (self.W, self.H)
def _NIMFA_NMF(self, X, nBases):
model = nimfa.mf(X, seed = 'nndsvd', rank = nBases,
method = "nmf", initialize_only = True)
fit = nimfa.mf_run(model)
W = fit.basis()
H = fit.coef()
self.W = W.todense()
self.H = H.todense()
return (self.W, self.H)
def max_guess_select(ratings, users, rank=9, user=None):
matrix = sp.dok_matrix((len(users), len(users)))
for k, v in ratings.items():
matrix[users[k[0]], users[k[1]]] = v
matrix[users[k[1]], users[k[0]]] = v
# Run sparse matrix factorisation
factor = nimfa.mf(matrix, seed="random_c", rank=rank, method="snmf", max_iter=12, initialize_only=True, version='r', eta=1., beta=1e-4, i_conv=10, w_min_change=0)
result = nimfa.mf_run(factor)
if user is None:
# Pick a user to expand
user = min(users, key=lambda u: len([i for i in ratings if u in i]))
recommendations = result.fitted()
rval = max([i for i in users if (i, user) not in ratings and (user, i) not in ratings], key=lambda x: recommendations[users[user], users[x]])
return user, rval
def fit(self, k=100, max_iter=15, method='lsnmf'):
if self.recommender_data.preference_matrix.shape[1] < k:
k = self.recommender_data.preference_matrix.shape[1]
model = nimfa.mf(self.recommender_data.preference_matrix,
seed="random_vcol",
rank=k,
method=method,
max_iter=max_iter)
fit = nimfa.mf_run(model)
self.user_matrix = fit.basis().todense()
self.item_matrix = fit.coef().todense()
def _factorize(matrix):
"Factorize the matrix to get pc"
# Build the model
model = mf(matrix,
seed="random_vcol",
rank=15,
method="nmf",
max_iter=15,
initialize_only=True,
update='divergence',
objective='div')
# Then fit it
fit = mf_run(model)
return fit.basis(), fit.coef()
def run(self, seed = 'random_vcol', method='nmf', rank=3, max_iter=65, display_N_tokens = 5, display_N_documents = 3):
#Re-initialise clusters
if self.clusters != []:
self.clusters = []
self.construct_term_doc_matrix(pca=False) #We cannot perform PCA with NMF because we only want non-negative vectors
V = self.td_matrix
model = nimfa.mf(V, seed = seed, method = method, rank = rank, max_iter = max_iter)
fitted = nimfa.mf_run(model)
w = fitted.basis()
h = fitted.coef()
self.split_documents(w,h, self.document_dict, self.attributes, display_N_tokens = display_N_tokens, display_N_documents = display_N_documents)
#Just testing remove it
self.showfeatures(w,h, [self.document_dict.values()[i]["raw"] for i in range(numpy.shape(w)[0])], self.attributes)
def run_snmf(V):
"""
Run sparse nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
# SNMF/R
rank = 10
model = nimfa.mf(V,
seed = "random_c",
rank = rank,
method = "snmf",
max_iter = 12,
initialize_only = True,
version = 'r',
eta = 1.,
beta = 1e-4,
i_conv = 10,
w_min_change = 0)
fit = nimfa.mf_run(model)
print_info(fit)
# SNMF/L
model = nimfa.mf(V,
seed = "random_vcol",
rank = rank,
method = "snmf",
max_iter = 12,
initialize_only = True,
version = 'l',
eta = 1.,
beta = 1e-4,
i_conv = 10,
w_min_change = 0)
fit = nimfa.mf_run(model)
print_info(fit)
def factor_eval(data, ranks, nrun=40, method="nmf", max_iter=2000):
coefs = []
for rank in ranks:
fctr = nimfa.mf(data,
method=method,
max_iter=max_iter,
rank=rank,
n_run=nrun,
track_factor=True)
fctr_res = nimfa.mf_run(fctr)
sm = fctr_res.summary()
coef = sm['cophenetic']
print coef
coefs.append(coef)
return coefs
def run_snmf(V):
"""
Run sparse nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
# SNMF/R
rank = 10
model = nimfa.mf(V,
seed="random_c",
rank=rank,
method="snmf",
max_iter=12,
initialize_only=True,
version='r',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
fit = nimfa.mf_run(model)
print_info(fit)
# SNMF/L
model = nimfa.mf(V,
seed="random_vcol",
rank=rank,
method="snmf",
max_iter=12,
initialize_only=True,
version='l',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
fit = nimfa.mf_run(model)
print_info(fit)
def run_nsnmf(V):
"""
Run nonsmooth nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed = "random",
rank = rank,
method = "nsnmf",
max_iter = 12,
initialize_only = True,
theta = 0.5)
fit = nimfa.mf_run(model)
print_info(fit)
def run_pmf(V):
"""
Run probabilistic matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed = "random_vcol",
rank = rank,
method = "pmf",
max_iter = 12,
initialize_only = True,
rel_error = 1e-5)
fit = nimfa.mf_run(model)
print_info(fit)
def run_one(V, rank):
"""
Run standard NMF on leukemia data set. 50 runs of Standard NMF are performed and obtained consensus matrix
averages all 50 connectivity matrices.
:param V: Target matrix with gene expression data.
:type V: `numpy.matrix` (of course it could be any format of scipy.sparse, but we will use numpy here)
:param rank: Factorization rank.
:type rank: `int`
"""
print "================= Rank = %d =================" % rank
consensus = np.mat(np.zeros((V.shape[1], V.shape[1])))
for i in xrange(50):
# Standard NMF with Euclidean update equations is used. For initialization random Vcol method is used.
# Objective function is the number of consecutive iterations in which the connectivity matrix has not changed.
# We demand that factorization does not terminate before 30 consecutive iterations in which connectivity matrix
# does not change. For a backup we also specify the maximum number of iterations. Note that the satisfiability
# of one stopping criteria terminates the run (there is no chance for divergence).
model = nimfa.mf(
V,
method="nmf",
rank=rank,
seed="random_vcol",
max_iter=200,
update="euclidean",
objective="conn",
conn_change=40,
initialize_only=True,
)
fit = nimfa.mf_run(model)
print "%2d / 50 :: %s - init: %s ran with ... %3d / 200 iters ..." % (
i + 1,
fit.fit,
fit.fit.seed,
fit.fit.n_iter,
)
# Compute connectivity matrix of factorization.
# Again, we could use multiple runs support of the nimfa library, track factorization model across 50 runs and then
# just call fit.consensus()
consensus += fit.fit.connectivity()
# averaging connectivity matrices
consensus /= 50.0
# reorder consensus matrix
p_consensus = reorder(consensus)
# plot reordered consensus matrix
plot(p_consensus, rank)
def run_nsnmf(V):
"""
Run nonsmooth nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed="random",
rank=rank,
method="nsnmf",
max_iter=12,
initialize_only=True,
theta=0.5)
fit = nimfa.mf_run(model)
print_info(fit)
def run_pmf(V):
"""
Run probabilistic matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed="random_vcol",
rank=rank,
method="pmf",
max_iter=12,
initialize_only=True,
rel_error=1e-5)
fit = nimfa.mf_run(model)
print_info(fit)
def run_psmf(V):
"""
Run probabilistic sparse matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
prng = np.random.RandomState()
model = nimfa.mf(V,
seed=None,
rank=rank,
method="psmf",
max_iter=12,
initialize_only=True,
prior=prng.uniform(low=0., high=1., size=10))
fit = nimfa.mf_run(model)
print_info(fit)
def run_psmf(V):
"""
Run probabilistic sparse matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
prng = np.random.RandomState()
model = nimfa.mf(V,
seed = None,
rank = rank,
method = "psmf",
max_iter = 12,
initialize_only = True,
prior = prng.uniform(low = 0., high = 1., size = 10))
fit = nimfa.mf_run(model)
print_info(fit)
def run_bmf(V):
"""
Run binary matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed = "random_vcol",
rank = rank,
method = "bmf",
max_iter = 12,
initialize_only = True,
lambda_w = 1.1,
lambda_h = 1.1)
fit = nimfa.mf_run(model)
print_info(fit)
def nmf(X, method='sklearn', **nmfparams):
"""
Calculates the non-negative matrix factorization of an input matrix
"""
#TODO: Documentation
if method == 'sklearn':
model = NMF(**nmfparams)
H = model.fit_transform(X)
W = model.components_
elif method == 'nimfa':
model_tmp = nimfa.mf(X, **nmfparams)
model = nimfa.mf_run(model_tmp)
H = model.coef()
W = model.basis()
return (H, W, model)
def nmf(X, method='sklearn', **nmfparams):
"""
Calculates the non-negative matrix factorization of an input matrix
"""
#TODO: Documentation
if method == 'sklearn':
model = NMF(**nmfparams)
H = model.fit_transform(X)
W = model.components_
elif method == 'nimfa':
model_tmp = nimfa.mf(X, **nmfparams)
model = nimfa.mf_run(model_tmp)
H = model.coef()
W = model.basis()
return (H, W, model)
def run_bmf(V):
"""
Run binary matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed="random_vcol",
rank=rank,
method="bmf",
max_iter=12,
initialize_only=True,
lambda_w=1.1,
lambda_h=1.1)
fit = nimfa.mf_run(model)
print_info(fit)
def run_lfnmf(V):
"""
Run local fisher nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
pnrg = np.random.RandomState()
model = nimfa.mf(V,
seed=None,
W=abs(pnrg.randn(V.shape[0], rank)),
H=abs(pnrg.randn(rank, V.shape[1])),
rank=rank,
method="lfnmf",
max_iter=12,
initialize_only=True,
alpha=0.01)
fit = nimfa.mf_run(model)
print_info(fit)
def run_lsnmf(V):
"""
Run least squares nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed="random_vcol",
rank=rank,
method="lsnmf",
max_iter=12,
initialize_only=True,
sub_iter=10,
inner_sub_iter=10,
beta=0.1,
min_residuals=1e-5)
fit = nimfa.mf_run(model)
print_info(fit)
def run_lsnmf(V):
"""
Run least squares nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
model = nimfa.mf(V,
seed = "random_vcol",
rank = rank,
method = "lsnmf",
max_iter = 12,
initialize_only = True,
sub_iter = 10,
inner_sub_iter = 10,
beta = 0.1,
min_residuals = 1e-5)
fit = nimfa.mf_run(model)
print_info(fit)
def run_lfnmf(V):
"""
Run local fisher nonnegative matrix factorization.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
pnrg = np.random.RandomState()
model = nimfa.mf(V,
seed = None,
W = abs(pnrg.randn(V.shape[0], rank)),
H = abs(pnrg.randn(rank, V.shape[1])),
rank = rank,
method = "lfnmf",
max_iter = 12,
initialize_only = True,
alpha = 0.01)
fit = nimfa.mf_run(model)
print_info(fit)
def factorize(data):
"""
Perform factorization on S. cerevisiae FunCat annotated sequence data set (D1 FC seq).
Return factorized data, this is matrix factors as result of factorization (basis and mixture matrix).
:param data: Transformed data set containing attributes' values, class information and possibly additional meta information.
:type data: `tuple`
"""
V = data['attr']
"""model = nimfa.mf(V,
seed = "random_vcol",
rank = 40,
method = "nmf",
max_iter = 75,
initialize_only = True,
update = 'euclidean',
objective = 'fro')"""
model = nimfa.mf(V,
seed="random_vcol",
rank=40,
method="snmf",
max_iter=5,
initialize_only=True,
version='l',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
print("Performing %s %s %d factorization ..." % (model, model.seed, model.rank))
fit = nimfa.mf_run(model)
print("... Finished")
sparse_w, sparse_h = fit.fit.sparseness()
print("""Stats:
- iterations: %d
- KL Divergence: %5.3f
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" % (fit.fit.n_iter, fit.distance(), fit.distance(metric='euclidean'), sparse_w, sparse_h))
data['W'] = fit.basis()
data['H'] = fit.coef()
return data
def run_one(V, rank):
"""
Run standard NMF on medulloblastoma data set. 50 runs of Standard NMF are performed and obtained consensus matrix
averages all 50 connectivity matrices.
:param V: Target matrix with gene expression data.
:type V: `numpy.matrix` (of course it could be any format of scipy.sparse, but we will use numpy here)
:param rank: Factorization rank.
:type rank: `int`
"""
print("================= Rank = %d =================" % rank)
consensus = np.mat(np.zeros((V.shape[1], V.shape[1])))
for i in range(50):
# Standard NMF with Euclidean update equations is used. For initialization random Vcol method is used.
# Objective function is the number of consecutive iterations in which the connectivity matrix has not changed.
# We demand that factorization does not terminate before 30 consecutive iterations in which connectivity matrix
# does not change. For a backup we also specify the maximum number of iterations. Note that the satisfiability
# of one stopping criteria terminates the run (there is no chance for
# divergence).
model = nimfa.mf(V,
method="nmf",
rank=rank,
seed="random_vcol",
max_iter=200,
update='euclidean',
objective='conn',
conn_change=40,
initialize_only=True)
fit = nimfa.mf_run(model)
print("%2d / 50 :: %s - init: %s ran with ... %3d / 200 iters ..." %
(i + 1, fit.fit, fit.fit.seed, fit.fit.n_iter))
# Compute connectivity matrix of factorization.
# Again, we could use multiple runs support of the nimfa library, track factorization model across 50 runs and then
# just call fit.consensus()
consensus += fit.fit.connectivity()
# averaging connectivity matrices
consensus /= 50.
# reorder consensus matrix
p_consensus = reorder(consensus)
# plot reordered consensus matrix
plot(p_consensus, rank)
def run_factorization(data):
fctr = nimfa.mf(data, seed = "random_c", rank = 15, method = "snmf", max_iter = 50, initialize_only = True, version = 'r', eta = 1., beta = 1e-4, i_conv = 10, w_min_change = 0)
fctr_res = nimfa.mf_run(fctr)
np.set_printoptions(precision=3)
np.set_printoptions(suppress=True)
# Basis matrix. It is sparse, as input data was sparse as well.
W = fctr_res.basis()
# print "Basis matrix"
# print W.todense()
# print W.shape
# Mixture matrix. We print this tiny matrix in dense format.
H = fctr_res.coef()
# print "Coef"
# print H.todense()
# print H.shape
# Return the loss function according to Kullback-Leibler divergence. By default Euclidean metric is used.
print "Distance Kullback-Leibler: %5.3e" % fctr_res.distance(metric = "kl")
# Compute generic set of measures to evaluate the quality of the factorization
sm = fctr_res.summary()
# Print sparseness (Hoyer, 2004) of basis and mixture matrix
print "Sparseness Basis: %5.3f Mixture: %5.3f" % (sm['sparseness'][0], sm['sparseness'][1])
# Print actual number of iterations performed
print "Iterations: %d" % sm['n_iter']
# Print estimate of target matrix data
data_fact = np.dot(W.todense(), H.todense() )
rmse = 0.0
for i in range(data.shape[0]):
for j in range(data.shape[1]):
if data[i][j] > 0:
print data[i,j], data_fact[i, j]
rmse += (data[i, j] - data_fact[i, j])**2
print "RMSE:", rmse
print data, data_fact
def factorize(data):
"""
Perform factorization on S. cerevisiae FunCat annotated sequence data set (D1 FC seq).
Return factorized data, this is matrix factors as result of factorization (basis and mixture matrix).
:param data: Transformed data set containing attributes' values, class information and possibly additional meta information.
:type data: `tuple`
"""
V = data['attr']
"""model = nimfa.mf(V,
seed = "random_vcol",
rank = 40,
method = "nmf",
max_iter = 75,
initialize_only = True,
update = 'euclidean',
objective = 'fro')"""
model = nimfa.mf(V,
seed = "random_vcol",
rank = 40,
method = "snmf",
max_iter = 5,
initialize_only = True,
version = 'l',
eta = 1.,
beta = 1e-4,
i_conv = 10,
w_min_change = 0)
print "Performing %s %s %d factorization ..." % (model, model.seed, model.rank)
fit = nimfa.mf_run(model)
print "... Finished"
sparse_w, sparse_h = fit.fit.sparseness()
print """Stats:
- iterations: %d
- KL Divergence: %5.3f
- Euclidean distance: %5.3f
- Sparseness basis: %5.3f, mixture: %5.3f""" % (fit.fit.n_iter, fit.distance(), fit.distance(metric = 'euclidean'), sparse_w, sparse_h)
data['W'] = fit.basis()
data['H'] = fit.coef()
return data
def run(self):
# TODO: estimate rank
self.mask_bed()
NMF_Run.run(self)
import nimfa
print repr(self.masked_matrix)
print repr(self.masked_matrix.shape)
self.fctr = nimfa.mf(numpy.matrix(self.masked_matrix),
seed = "nndsvd",
rank = self.nmf_rank,
method = "bmf",
max_iter = self.max_iter,
initialize_only = True,
lambda_w = 1.1,
lambda_h = 1.1)
self.fctr_res = nimfa.mf_run(self.fctr)
print 'Rss: %5.4f' % self.fctr_res.fit.rss()
print 'Evar: %5.4f' % self.fctr_res.fit.evar()
print 'K-L divergence: %5.4f' % self.fctr_res.distance(metric = 'kl')
print 'Sparseness, W: %5.4f, H: %5.4f' % self.fctr_res.fit.sparseness()
print 'Iteration: %d' % self.fctr_res.n_iter
def probability_select(ratings, users, rank=9, user=None):
matrix = sp.dok_matrix((len(users), len(users)))
for k, v in ratings.items():
matrix[users[k[0]], users[k[1]]] = v
matrix[users[k[1]], users[k[0]]] = v
# Run sparse matrix factorisation
factor = nimfa.mf(matrix, seed="random_c", rank=rank, method="snmf", max_iter=12, initialize_only=True, version='r', eta=1., beta=1e-4, i_conv=10, w_min_change=0)
result = nimfa.mf_run(factor)
if len(ratings) >= len(users)**2:
return # all items expanded
if user is None:
# Pick a user to expand
user = min(users, key=lambda u: len([i for i in ratings if u in i]))
# Clusters (F)
clusters = result.basis()
# Matrix (M)
recommendations = result.fitted()
# All rated users (U)
user_rated = {i[0]: ratings[i] for i in ratings if user == i[1]}
user_rated.update({i[1]: ratings[i] for i in ratings if user == i[0]})
# Affiliations (A)
caff = [(sum(r * clusters[users[u], x] for u, r in user_rated.items())+1)/(len(user_rated)+1) for x in range(rank)]
# Confidence (d)
conf = sum(sum(clusters[users[u], x] for u in user_rated) for x in range(rank))/clusters.sum()
# Cluster confidences (C)
sums = clusters.sum(axis=0).tolist()[0]
cconf = [sum(clusters[users[u], x] for u in user_rated)/sums[x] for x in range(rank)]
cconf_norm = max(cconf) or 1
cconf = [i/cconf_norm for i in cconf]
# Find the user with the highest affinity to cluster
candidates = {i for i in users if i not in user_rated}
candidate = max(candidates, key=lambda x: conf * recommendations[users[user], users[x]] + (1-conf) * (sum((1-cconf[i])*caff[i]*clusters[users[x], i] for i in range(rank))/rank))
return user, candidate
def apply( self, X, k = 2 ): """ Apply NMF to the specified document-term matrix X. """ import nimfa self.W = None self.H = None initialize_only = self.max_iters < 1 if self.update == "euclidean": objective = "fro" else: objective = "div" alg = nimfa.mf(X, method = self.method, max_iter = self.max_iters, rank = k, seed = self.init_strategy, update = self.update, objective = objective, test_conv = self.test_conv ) res = nimfa.mf_run(alg) # TODO: fix try: self.W = res.basis().todense() self.H = res.coef().todense() except: self.W = res.basis() self.H = res.coef() # last number of iterations self.n_iter = res.n_iter
def run_icm(V):
"""
Run iterated conditional modes.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
pnrg = np.random.RandomState()
model = nimfa.mf(V,
seed = "nndsvd",
rank = rank,
method = "icm",
max_iter = 12,
initialize_only = True,
iiter = 20,
alpha = pnrg.randn(V.shape[0], rank),
beta = pnrg.randn(rank, V.shape[1]),
theta = 0.,
k = 0.,
sigma = 1.)
fit = nimfa.mf_run(model)
print_info(fit)
def run_icm(V):
"""
Run iterated conditional modes.
:param V: Target matrix to estimate.
:type V: :class:`numpy.matrix`
"""
rank = 10
pnrg = np.random.RandomState()
model = nimfa.mf(V,
seed="nndsvd",
rank=rank,
method="icm",
max_iter=12,
initialize_only=True,
iiter=20,
alpha=pnrg.randn(V.shape[0], rank),
beta=pnrg.randn(rank, V.shape[1]),
theta=0.,
k=0.,
sigma=1.)
fit = nimfa.mf_run(model)
print_info(fit)
def clustered_select(ratings, users, rank=9, user=None):
matrix = sp.dok_matrix((len(users), len(users)))
for k, v in ratings.items():
matrix[users[k[0]], users[k[1]]] = v
matrix[users[k[1]], users[k[0]]] = v
# Run sparse matrix factorisation
factor = nimfa.mf(matrix,
seed="random_c",
rank=rank,
method="snmf",
max_iter=12,
initialize_only=True,
version='r',
eta=1.,
beta=1e-4,
i_conv=10,
w_min_change=0)
result = nimfa.mf_run(factor)
if len(ratings) >= len(users)**2:
return # all items expanded
if user is None:
# Pick a user to expand
user = min(users, key=lambda u: len([i for i in ratings if u in i]))
# Pick a cluster
clusters = result.basis()
# Select all rated users
user_rated = {i[0]: ratings[i] for i in ratings if user == i[1]}
user_rated.update({i[1]: ratings[i] for i in ratings if user == i[0]})
cluster = max(range(rank),
key=lambda x: (sum(r * clusters[users[u], x]
for u, r in user_rated.items()) + 1) /
(len(user_rated) + 1)) # Maximise A_u(c)
# Find the user with the highest affinity to cluster
candidates = {i for i in users if i not in user_rated}
candidate = max(candidates, key=lambda x: clusters[users[x], cluster])
return user, candidate
# Initialization callback function
def init_info(model):
print "Initialized basis matrix\n", model.basis()
print "Initialized mixture matrix\n", model.coef()
# ICM rank 3 algorithm
# We specify callback_init parameter by passing a init_info function
# Callback is called after initialization and prior to factorization in each run.
fctr = nimfa.mf(V,
seed="random_c",
method="icm",
max_iter=10,
rank=3,
callback_init=init_info)
fctr_res = nimfa.mf_run(fctr)
# Basis matrix.
W = fctr_res.basis()
print "Resulting basis matrix"
print W
# Mixture matrix.
H = fctr_res.coef()
print "Resulting mixture matrix"
print H
sm = fctr_res.summary()
print "Rss: %8.3e" % sm['rss']
print "Evar: %8.3e" % sm['evar']
print "Iterations: %d" % sm['n_iter']