def em_optimization(n_dw_matrix,
phi_matrix,
theta_matrix,
regularization_list,
iters_count=100,
loss_function=None,
iteration_callback=None,
const_phi=False,
params=None):
if loss_function is None:
loss_function = LogFunction()
if params is None:
params = {}
return_counters = params.get('return_counters', False)
phi_matrix = np.copy(phi_matrix)
theta_matrix = np.copy(theta_matrix)
docptr = get_docptr(n_dw_matrix)
wordptr = n_dw_matrix.indices
start_time = time.time()
for it in xrange(iters_count):
phi_matrix_tr = np.transpose(phi_matrix)
s_data = loss_function.calc_der(
memory_efficient_inner1d(theta_matrix, docptr, phi_matrix_tr,
wordptr))
A = scipy.sparse.csr_matrix((n_dw_matrix.data * s_data,
n_dw_matrix.indices, n_dw_matrix.indptr),
shape=n_dw_matrix.shape)
n_dt = A.dot(phi_matrix_tr) * theta_matrix
n_tw = np.transpose(
A.tocsc().transpose().dot(theta_matrix)) * phi_matrix
r_tw, r_dt = regularization_list[it](phi_matrix, theta_matrix, n_tw,
n_dt)
n_tw += r_tw
n_dt += r_dt
if not const_phi:
phi_matrix = get_prob_matrix_by_counters(n_tw)
theta_matrix = get_prob_matrix_by_counters(n_dt)
if iteration_callback is not None:
iteration_callback(it, phi_matrix, theta_matrix)
print 'Iters time', time.time() - start_time
if return_counters:
return phi_matrix, theta_matrix, n_tw, n_dt
else:
return phi_matrix, theta_matrix
def gradient_optimization(
n_dw_matrix,
phi_matrix,
theta_matrix,
regularization_gradient_list,
iters_count=100,
loss_function=None,
iteration_callback=None,
learning_rate=1.,
):
if loss_function is None:
loss_function = LogFunction()
phi_matrix = np.copy(phi_matrix)
theta_matrix = np.copy(theta_matrix)
docptr = get_docptr(n_dw_matrix)
wordptr = n_dw_matrix.indices
start_time = time.time()
for it in xrange(iters_count):
phi_matrix_tr = np.transpose(phi_matrix)
s_data = loss_function.calc_der(
memory_efficient_inner1d(theta_matrix, docptr, phi_matrix_tr,
wordptr))
A = scipy.sparse.csr_matrix((n_dw_matrix.data * s_data,
n_dw_matrix.indices, n_dw_matrix.indptr),
shape=n_dw_matrix.shape).tocsc()
g_tw = theta_matrix.T * A
g_dt = A.dot(phi_matrix_tr)
r_tw, r_dt = regularization_gradient_list[it](phi_matrix, theta_matrix,
phi_matrix, theta_matrix)
g_tw += r_tw
g_dt += r_dt
g_tw -= np.sum(g_tw * phi_matrix, axis=1)[:, np.newaxis]
g_dt -= np.sum(g_dt * theta_matrix, axis=1)[:, np.newaxis]
phi_matrix += g_tw * learning_rate
theta_matrix += g_dt * learning_rate
phi_matrix = get_prob_matrix_by_counters(phi_matrix)
theta_matrix = get_prob_matrix_by_counters(theta_matrix)
if iteration_callback is not None:
iteration_callback(it, phi_matrix, theta_matrix)
print 'Iters time', time.time() - start_time
return phi_matrix, theta_matrix
def naive_thetaless_em_optimization(n_dw_matrix,
phi_matrix,
regularization_list,
iters_count=100,
iteration_callback=None,
theta_matrix=None,
params=None):
if params is None:
params = {}
loss_function = LogFunction()
return_counters = params.get('return_counters', False)
phi_matrix = np.copy(phi_matrix)
docptr = get_docptr(n_dw_matrix)
wordptr = n_dw_matrix.indices
start_time = time.time()
for it in xrange(iters_count):
phi_rev_matrix = np.transpose(phi_matrix / np.sum(phi_matrix, axis=0))
phi_matrix_tr = np.transpose(phi_matrix)
theta_matrix = get_prob_matrix_by_counters(
n_dw_matrix.dot(phi_rev_matrix))
s_data = loss_function.calc_der(
memory_efficient_inner1d(theta_matrix, docptr, phi_matrix_tr,
wordptr))
A = scipy.sparse.csr_matrix((n_dw_matrix.data * s_data,
n_dw_matrix.indices, n_dw_matrix.indptr),
shape=n_dw_matrix.shape).tocsc()
n_tw = (A.T.dot(theta_matrix)).T * phi_matrix
r_tw, _ = regularization_list[it](phi_matrix, theta_matrix, n_tw,
theta_matrix)
n_tw += r_tw
phi_matrix = get_prob_matrix_by_counters(n_tw)
if iteration_callback is not None:
iteration_callback(it, phi_matrix, theta_matrix)
print 'Iters time', time.time() - start_time
if return_counters:
return phi_matrix, theta_matrix, n_tw, None
else:
return phi_matrix, theta_matrix
def artm_thetaless_em_optimization(n_dw_matrix,
phi_matrix,
regularization_list,
iters_count=100,
iteration_callback=None,
theta_matrix=None,
params=None):
if params is None:
params = {}
loss_function = LogFunction()
use_B_cheat = params.get('use_B_cheat', False)
return_counters = params.get('return_counters', False)
phi_matrix = np.copy(phi_matrix)
docptr = get_docptr(n_dw_matrix)
wordptr = n_dw_matrix.indices
D, _ = n_dw_matrix.shape
docsizes = []
indptr = n_dw_matrix.indptr
for doc_num in xrange(D):
size = indptr[doc_num + 1] - indptr[doc_num]
if use_B_cheat:
docsizes.extend([size] * size)
else:
docsizes.extend([
np.sum(n_dw_matrix.data[indptr[doc_num]:indptr[doc_num + 1]])
] * size)
docsizes = np.array(docsizes)
B = scipy.sparse.csr_matrix((1. * n_dw_matrix.data / docsizes,
n_dw_matrix.indices, n_dw_matrix.indptr),
shape=n_dw_matrix.shape).tocsc()
start_time = time.time()
for it in xrange(iters_count):
phi_matrix_tr = np.transpose(phi_matrix)
phi_rev_matrix = get_prob_matrix_by_counters(phi_matrix_tr)
theta_matrix = get_prob_matrix_by_counters(
n_dw_matrix.dot(phi_rev_matrix))
s_data = loss_function.calc_der(
memory_efficient_inner1d(theta_matrix, docptr, phi_matrix_tr,
wordptr))
A = scipy.sparse.csr_matrix((n_dw_matrix.data * s_data,
n_dw_matrix.indices, n_dw_matrix.indptr),
shape=n_dw_matrix.shape).tocsc()
n_tw = A.T.dot(theta_matrix).T * phi_matrix
r_tw, r_dt = regularization_list[it](phi_matrix, theta_matrix, n_tw,
theta_matrix)
theta_indices = theta_matrix > OPT_EPS
r_dt[theta_indices] /= theta_matrix[theta_indices]
r_dt[~theta_indices] = 0.
g_dt = A.dot(phi_matrix_tr) + r_dt
tmp = g_dt.T * B / (phi_matrix_tr.sum(axis=1) + OPT_EPS)
r_tw += (tmp - np.einsum('ij,ji->i', phi_rev_matrix, tmp)) * phi_matrix
n_tw += r_tw
phi_matrix = get_prob_matrix_by_counters(n_tw)
if iteration_callback is not None:
iteration_callback(it, phi_matrix, theta_matrix)
print 'Iters time', time.time() - start_time
if return_counters:
return phi_matrix, theta_matrix, n_tw, None
else:
return phi_matrix, theta_matrix